diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_0/index_tensor([4989, 4989])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_0/index_tensor([4989, 4989])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8e0f3b2d2b69d818a235c80d62491b1dfe33389e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_0/index_tensor([4989, 4989])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, medium shot, shot from above, focuses with powerful visual intensity on a young deity with flowing black hair, clad in an intricately crafted suit of crimson lotus armor, standing proudly amidst a devastating mountain torrent. The armor, seemingly forged from flaming lotus petals, each blade a deep crimson, gleaming with a faint golden hue at its edges, its surface shimmering with the ripples of water swept by the rapids. Beneath a sky thick with dark clouds and thunder, the dynamic, realistic play of light and shadow shifts constantly, each lightning flash illuminating his resolute and stern young face and fearless eyes that seem to hold a star. The turbulent torrent crashes upon him with unstoppable force, yet he remains unmoved, as steady as a rock. At this moment, the camera captures a stunning slow-motion particle effect: the massive splash of water from the impact is broken into countless individual droplets, which solidify in the air like suspended crystals and shimmering diamond dust. Each droplet clearly reflects the cold light of the lightning and the fiery red glow of the armor, creating a striking contrast between cold and warm. The entire scene, imbued with unparalleled tension in this extreme contrast between movement and stillness, showcases the absolute majesty and composure of the deity against the violent forces of nature.", + "seed": 16546, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_0/index_4989_seed_16546_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": -0.029353629751130938, + "vlm_logprob": -0.029353629751130938, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.029353629751130938, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013582786545157433 + ], + "opposite_logprob_tokens": [ + -4.513582706451416 + ], + "logprob_reward": -0.013582786545157433, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013582786545157433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5156044960021973, + 0.06475945562124252, + 0.07092764973640442, + 0.06323398649692535, + 0.04316526651382446, + 0.0400557778775692, + 0.030558764934539795, + 0.05348401889204979, + 0.05513788387179375, + 0.06307268142700195 + ], + "credit_map": [ + 0.06294456869363785, + 0.04991355910897255, + 0.09259771555662155, + 0.051501091569662094, + 0.05054273083806038, + 0.06382977962493896, + 0.07859516143798828, + 0.06567990034818649, + 0.008590678684413433, + 0.006437200121581554, + 0.009362335316836834, + 0.00616780761629343, + 0.00951011199504137, + 0.009254053235054016, + 0.006004345137625933, + 0.009432927705347538, + 0.015142381191253662, + 0.0062501560896635056, + 0.006164175923913717, + 0.0044803679920732975, + 0.006756259594112635, + 0.006510427221655846, + 0.0054891048930585384, + 0.020134778693318367, + 0.006429616827517748, + 0.005637823138386011, + 0.013776954263448715, + 0.006264095660299063, + 0.006879545282572508, + 0.005961291026324034, + 0.006095061544328928, + 0.0121895931661129, + 0.005376288667321205, + 0.006011683028191328, + 0.006823700852692127, + 0.005314067006111145, + 0.00483772624284029, + 0.006944113876670599, + 0.0035524622071534395, + 0.004305227193981409, + 0.006125899963080883, + 0.005671244114637375, + 0.003689014818519354, + 0.003970880061388016, + 0.007945588789880276, + 0.005031845066696405, + 0.00378781883046031, + 0.0038334850687533617, + 0.00388674670830369, + 0.0028243353590369225, + 0.005730287171900272, + 0.0035735515411943197, + 0.004318283870816231, + 0.0038789387326687574, + 0.0034884505439549685, + 0.0028581698425114155, + 0.006939276587218046, + 0.005347535014152527, + 0.006864945869892836, + 0.006398370489478111, + 0.00815918855369091, + 0.006659562233835459, + 0.006927070207893848, + 0.0061880676075816154, + 0.005642199423164129, + 0.004569504875689745, + 0.005766791291534901, + 0.0050052618607878685, + 0.005839830730110407, + 0.007084419950842857, + 0.006413158029317856, + 0.01481671817600727, + 0.007967923767864704, + 0.00817933864891529, + 0.0045461044646799564, + 0.005604693200439215, + 0.01524096168577671, + 0.0069073461927473545, + 0.010584050789475441, + 0.004042262211441994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.013582786545157433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002204132964834571 + ], + "opposite_logprob_tokens": [ + -6.127203941345215 + ], + "logprob_reward": -0.002204132964834571, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002204132964834571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387723684310913, + 0.11121860891580582, + 0.19810466468334198, + 0.16333071887493134, + 0.2803635597229004, + 0.02511926367878914, + 0.01946244388818741, + 0.03222421556711197, + 0.018588783219456673, + 0.01281537115573883 + ], + "credit_map": [ + 0.01875489391386509, + 0.011751910671591759, + 0.02703673206269741, + 0.012757409363985062, + 0.01303025707602501, + 0.01888115704059601, + 0.018564466387033463, + 0.017995528876781464, + 0.015150200575590134, + 0.006952300202101469, + 0.018889399245381355, + 0.0070104217156767845, + 0.011493140831589699, + 0.021841775625944138, + 0.007347282487899065, + 0.022534089162945747, + 0.04808678478002548, + 0.014046541415154934, + 0.020408809185028076, + 0.006561219692230225, + 0.008520269766449928, + 0.01226931530982256, + 0.014987916685640812, + 0.07322381436824799, + 0.021997805684804916, + 0.015625497326254845, + 0.030926614999771118, + 0.011489277705550194, + 0.017979063093662262, + 0.02547723986208439, + 0.01473477017134428, + 0.025100454688072205, + 0.029323037713766098, + 0.04029262810945511, + 0.041544463485479355, + 0.03424286097288132, + 0.04668663069605827, + 0.034918271005153656, + 0.02952645532786846, + 0.023829197511076927, + 0.0036540592554956675, + 0.002635819837450981, + 0.0024520931765437126, + 0.0029393876902759075, + 0.004213456995785236, + 0.0035250873770564795, + 0.0020913993939757347, + 0.0036079613491892815, + 0.0028605160769075155, + 0.0014106921153143048, + 0.004480334930121899, + 0.002447274047881365, + 0.001898733782581985, + 0.0022842332255095243, + 0.0022132075391709805, + 0.001867453334853053, + 0.004067218396812677, + 0.003035313682630658, + 0.004661265294998884, + 0.004711861722171307, + 0.0038699896540492773, + 0.003997152205556631, + 0.004214916378259659, + 0.0036665010266005993, + 0.001956822583451867, + 0.0018413376528769732, + 0.0021257123444229364, + 0.002043437212705612, + 0.002245507901534438, + 0.002367771929129958, + 0.002271219389513135, + 0.0037369730416685343, + 0.001838003401644528, + 0.001836809329688549, + 0.0007948459824547172, + 0.0014445557026192546, + 0.0021996942814439535, + 0.0013604273553937674, + 0.0025707348249852657, + 0.0007703002775087953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.002204132964834571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033692337572574615 + ], + "opposite_logprob_tokens": [ + -3.4086923599243164 + ], + "logprob_reward": -0.033692337572574615, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033692337572574615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21788392961025238, + 0.059822309762239456, + 0.22439926862716675, + 0.11146288365125656, + 0.14824654161930084, + 0.05286738649010658, + 0.0811963751912117, + 0.04633548855781555, + 0.03159956634044647, + 0.026186278089880943 + ], + "credit_map": [ + 0.03311954811215401, + 0.0275972131639719, + 0.03555288538336754, + 0.021618524566292763, + 0.027332311496138573, + 0.024949193000793457, + 0.025063978508114815, + 0.022650256752967834, + 0.00930128525942564, + 0.006082156207412481, + 0.006699679885059595, + 0.006021025124937296, + 0.007194649428129196, + 0.01023845188319683, + 0.006129813846200705, + 0.008155248127877712, + 0.05308847874403, + 0.016830332577228546, + 0.021574486047029495, + 0.007878953590989113, + 0.009798812679946423, + 0.01812327653169632, + 0.01668976992368698, + 0.08041515201330185, + 0.017304737120866776, + 0.013472305610775948, + 0.012726462446153164, + 0.011381802149116993, + 0.014470833353698254, + 0.015466137789189816, + 0.01533725205808878, + 0.011303355917334557, + 0.015013724565505981, + 0.015805425122380257, + 0.021641526371240616, + 0.015437011606991291, + 0.018804531544446945, + 0.028189275413751602, + 0.013261212967336178, + 0.020093826577067375, + 0.00807653646916151, + 0.005875531118363142, + 0.005140427965670824, + 0.005453130695968866, + 0.008651240728795528, + 0.008678670972585678, + 0.0049118162132799625, + 0.006080033257603645, + 0.012072374112904072, + 0.007676365319639444, + 0.015443551354110241, + 0.00892692431807518, + 0.008288199082016945, + 0.013529944233596325, + 0.007831545546650887, + 0.00742747075855732, + 0.006102594546973705, + 0.004143842030316591, + 0.005966650787740946, + 0.006734143011271954, + 0.006392365321516991, + 0.00488887308165431, + 0.006765981670469046, + 0.00534104136750102, + 0.005041788332164288, + 0.003994549624621868, + 0.004492293577641249, + 0.003143858863040805, + 0.004452997352927923, + 0.004100481979548931, + 0.0032250669319182634, + 0.0031485254876315594, + 0.0026063118129968643, + 0.0032635030802339315, + 0.0027677593752741814, + 0.002704374026507139, + 0.004122926853597164, + 0.0027586673386394978, + 0.004970499314367771, + 0.0029922369867563248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.033692337572574615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865341633558273 + ], + "opposite_logprob_tokens": [ + -3.0486533641815186 + ], + "logprob_reward": -0.04865341633558273, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865341633558273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13834957778453827, + 0.12791654467582703, + 0.35026785731315613, + 0.13933566212654114, + 0.07212486863136292, + 0.03149769455194473, + 0.04601902514696121, + 0.033606912940740585, + 0.029377378523349762, + 0.031504545360803604 + ], + "credit_map": [ + 0.021558502689003944, + 0.015239652246236801, + 0.026895303279161453, + 0.014557283371686935, + 0.01280487235635519, + 0.0163966566324234, + 0.014236434362828732, + 0.01666085608303547, + 0.012623315677046776, + 0.013667445629835129, + 0.02089092694222927, + 0.010284661315381527, + 0.018137404695153236, + 0.013702585361897945, + 0.014037473127245903, + 0.024572715163230896, + 0.08785980939865112, + 0.024689653888344765, + 0.03442287817597389, + 0.012442382983863354, + 0.014806223101913929, + 0.023546746000647545, + 0.025618473067879677, + 0.1268816441297531, + 0.014433912932872772, + 0.011192232370376587, + 0.027729826048016548, + 0.016397476196289062, + 0.013235248625278473, + 0.013968657702207565, + 0.01664135791361332, + 0.02573692798614502, + 0.007468050811439753, + 0.008142095059156418, + 0.00941741093993187, + 0.008044770918786526, + 0.00973332766443491, + 0.01060881745070219, + 0.008599509485065937, + 0.010110864415764809, + 0.003318258561193943, + 0.0037099362816661596, + 0.0026386110112071037, + 0.003909123130142689, + 0.0060278200544416904, + 0.003997809253633022, + 0.0035118740051984787, + 0.00438425550237298, + 0.007519167847931385, + 0.004549598786979914, + 0.00727440882474184, + 0.005173217039555311, + 0.005525005515664816, + 0.007518633734434843, + 0.004573951940983534, + 0.0038850330747663975, + 0.007782887667417526, + 0.0038573581259697676, + 0.0033428019378334284, + 0.003246848238632083, + 0.004641935229301453, + 0.004429719410836697, + 0.003552087116986513, + 0.002753267763182521, + 0.004607074894011021, + 0.0034522544592618942, + 0.0031520104967057705, + 0.002833219012245536, + 0.005662499461323023, + 0.0032595528755337, + 0.0029989543836563826, + 0.0034118087496608496, + 0.0035787089727818966, + 0.0035688059870153666, + 0.0038173815701156855, + 0.0028886189684271812, + 0.00638381065800786, + 0.003555281087756157, + 0.004977313801646233, + 0.002734618028625846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.04865341633558273 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863547533750534 + ], + "opposite_logprob_tokens": [ + -3.048635482788086 + ], + "logprob_reward": -0.04863547533750534, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863547533750534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1751810759305954, + 0.07406549155712128, + 0.1254100799560547, + 0.17144206166267395, + 0.2459225207567215, + 0.06165972352027893, + 0.03803575783967972, + 0.031678084284067154, + 0.029003601521253586, + 0.04760166257619858 + ], + "credit_map": [ + 0.022094428539276123, + 0.01988532952964306, + 0.023778464645147324, + 0.019521810114383698, + 0.019203606992959976, + 0.019226569682359695, + 0.028910094872117043, + 0.02256077155470848, + 0.008998088538646698, + 0.006190212443470955, + 0.013504806905984879, + 0.0077522993087768555, + 0.006220092531293631, + 0.006231081206351519, + 0.008069898933172226, + 0.017099017277359962, + 0.026869509369134903, + 0.012729921378195286, + 0.011800872161984444, + 0.010419766418635845, + 0.013431139290332794, + 0.011064407415688038, + 0.010263620875775814, + 0.028830843046307564, + 0.019398484379053116, + 0.016375232487916946, + 0.02759404107928276, + 0.017761608585715294, + 0.021483883261680603, + 0.017284145578742027, + 0.020717157050967216, + 0.030827516689896584, + 0.02649809420108795, + 0.028896305710077286, + 0.03396633639931679, + 0.024281509220600128, + 0.027842160314321518, + 0.043048277497291565, + 0.022608846426010132, + 0.03878096118569374, + 0.00801827758550644, + 0.0062885223887860775, + 0.00646078260615468, + 0.00871867686510086, + 0.009976820088922977, + 0.006645119749009609, + 0.006340049672871828, + 0.009211469441652298, + 0.006037652492523193, + 0.0035075617488473654, + 0.0047517623752355576, + 0.0031450677197426558, + 0.005688145756721497, + 0.006142186466604471, + 0.004370424430817366, + 0.004392956383526325, + 0.0041102017275989056, + 0.004216711036860943, + 0.003879666095599532, + 0.002957138931378722, + 0.006044335197657347, + 0.0037320503033697605, + 0.0029094151686877012, + 0.003828564425930381, + 0.004105950705707073, + 0.003043282311409712, + 0.004816300701349974, + 0.003239104524254799, + 0.003489522961899638, + 0.004103680606931448, + 0.0036523977760225534, + 0.0025533614680171013, + 0.00810153316706419, + 0.007006232626736164, + 0.004023911897093058, + 0.003155504120513797, + 0.010258181020617485, + 0.005495600402355194, + 0.006486596539616585, + 0.0030741016380488873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.04863547533750534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, medium shot, shot from above, focuses with powerful visual intensity on a young deity with flowing black hair, clad in an intricately crafted suit of crimson lotus armor, standing proudly amidst a devastating mountain torrent. The armor, seemingly forged from flaming lotus petals, each blade a deep crimson, gleaming with a faint golden hue at its edges, its surface shimmering with the ripples of water swept by the rapids. Beneath a sky thick with dark clouds and thunder, the dynamic, realistic play of light and shadow shifts constantly, each lightning flash illuminating his resolute and stern young face and fearless eyes that seem to hold a star. The turbulent torrent crashes upon him with unstoppable force, yet he remains unmoved, as steady as a rock. At this moment, the camera captures a stunning slow-motion particle effect: the massive splash of water from the impact is broken into countless individual droplets, which solidify in the air like suspended crystals and shimmering diamond dust. Each droplet clearly reflects the cold light of the lightning and the fiery red glow of the armor, creating a striking contrast between cold and warm. The entire scene, imbued with unparalleled tension in this extreme contrast between movement and stillness, showcases the absolute majesty and composure of the deity against the violent forces of nature.", + "seed": 16547, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_0/index_4989_seed_16547_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": -0.029276548698544503, + "vlm_logprob": -0.029276548698544503, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.029276548698544503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023429686203598976 + ], + "opposite_logprob_tokens": [ + -4.023429870605469 + ], + "logprob_reward": -0.023429686203598976, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.023429686203598976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16512440145015717, + 0.1380748599767685, + 0.03729372099041939, + 0.052846044301986694, + 0.05490974709391594, + 0.21038967370986938, + 0.12960852682590485, + 0.03426458314061165, + 0.0678546279668808, + 0.10963375121355057 + ], + "credit_map": [ + 0.03315969184041023, + 0.012401661835610867, + 0.039714884012937546, + 0.0248861201107502, + 0.014955629594624043, + 0.008195566944777966, + 0.01804865524172783, + 0.013762217946350574, + 0.013059801422059536, + 0.01228857971727848, + 0.01861351914703846, + 0.011485698632895947, + 0.0346500538289547, + 0.015495489351451397, + 0.019807390868663788, + 0.012674340046942234, + 0.004726402927190065, + 0.004226019140332937, + 0.0031656299252063036, + 0.0028931160923093557, + 0.00544442655518651, + 0.007843293249607086, + 0.0042631132528185844, + 0.004731722641736269, + 0.008831373415887356, + 0.004782804287970066, + 0.0041701821610331535, + 0.005150657147169113, + 0.009142346680164337, + 0.008042254485189915, + 0.007658130954951048, + 0.005068296566605568, + 0.009524243883788586, + 0.004947124980390072, + 0.006136528681963682, + 0.004846089985221624, + 0.008244100958108902, + 0.008183641359210014, + 0.005661824718117714, + 0.007366198115050793, + 0.06422004848718643, + 0.01314006932079792, + 0.010988076217472553, + 0.0084664486348629, + 0.019780589267611504, + 0.040757179260253906, + 0.01507309079170227, + 0.03796418383717537, + 0.011841709725558758, + 0.015430550090968609, + 0.01228845864534378, + 0.020169265568256378, + 0.025569148361682892, + 0.012358656153082848, + 0.019999999552965164, + 0.01195074524730444, + 0.0038298547733575106, + 0.004014585632830858, + 0.003894225927069783, + 0.004165006335824728, + 0.006905578076839447, + 0.0034789799246937037, + 0.004793011117726564, + 0.003183345776051283, + 0.007884622551500797, + 0.007586831692606211, + 0.007085956633090973, + 0.007531031966209412, + 0.015156382694840431, + 0.007943930104374886, + 0.009186359122395515, + 0.00547951553016901, + 0.015275698155164719, + 0.009059803560376167, + 0.017259543761610985, + 0.014744516462087631, + 0.014963468536734581, + 0.010659257881343365, + 0.015046332962810993, + 0.012625140137970448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.023429686203598976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082093119621277 + ], + "opposite_logprob_tokens": [ + -4.141082286834717 + ], + "logprob_reward": -0.016082093119621277, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082093119621277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10199842602014542, + 0.058957017958164215, + 0.4705767035484314, + 0.08888327330350876, + 0.04818151518702507, + 0.11869339644908905, + 0.032331932336091995, + 0.010477820411324501, + 0.0534510537981987, + 0.016448823735117912 + ], + "credit_map": [ + 0.017406761646270752, + 0.00842718593776226, + 0.025009019300341606, + 0.007723784539848566, + 0.010094565339386463, + 0.011952740140259266, + 0.010286703705787659, + 0.011097666807472706, + 0.005151928402483463, + 0.004957607947289944, + 0.00744711235165596, + 0.008181445300579071, + 0.009370150975883007, + 0.005324719939380884, + 0.012492784298956394, + 0.0060312687419354916, + 0.06837163120508194, + 0.05636664852499962, + 0.037773944437503815, + 0.03584375977516174, + 0.06779280304908752, + 0.1061968058347702, + 0.040925078094005585, + 0.05730609595775604, + 0.010885648429393768, + 0.0070084454491734505, + 0.015521922148764133, + 0.009582065977156162, + 0.0077343774028122425, + 0.0073576923459768295, + 0.01892317272722721, + 0.01186995580792427, + 0.010069923475384712, + 0.0034231350291520357, + 0.00659434637054801, + 0.005066493060439825, + 0.005167911760509014, + 0.004563883412629366, + 0.005072778556495905, + 0.008223043754696846, + 0.012717395089566708, + 0.008712908253073692, + 0.02535608410835266, + 0.015798721462488174, + 0.01149331871420145, + 0.009861012920737267, + 0.016877036541700363, + 0.01787692680954933, + 0.0032972691114991903, + 0.0025206664577126503, + 0.004919812548905611, + 0.005658470559865236, + 0.0035080844536423683, + 0.0027525268960744143, + 0.005304921884089708, + 0.004370180424302816, + 0.0013747907942160964, + 0.0013078078627586365, + 0.0010238972026854753, + 0.0013167532160878181, + 0.002194751985371113, + 0.0011627099011093378, + 0.0014229484368115664, + 0.0006741619436070323, + 0.008990809321403503, + 0.006244496908038855, + 0.00514150969684124, + 0.005569095723330975, + 0.008997543714940548, + 0.008194586262106895, + 0.006474693771451712, + 0.003838320728391409, + 0.002643887884914875, + 0.0014295614091679454, + 0.002350600901991129, + 0.002264481270685792, + 0.001895085908472538, + 0.0016750211361795664, + 0.0022610528394579887, + 0.0019291334319859743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.016082093119621277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1764465570449829, + 0.12688247859477997, + 0.19713404774665833, + 0.11089751869440079, + 0.10645810514688492, + 0.13528582453727722, + 0.048731785267591476, + 0.027668245136737823, + 0.03835021331906319, + 0.03214523568749428 + ], + "credit_map": [ + 0.026273800060153008, + 0.01793244667351246, + 0.030667461454868317, + 0.030180254951119423, + 0.013128239661455154, + 0.014937328174710274, + 0.023132342845201492, + 0.02019469253718853, + 0.015865832567214966, + 0.011002147570252419, + 0.017599456012248993, + 0.01164632011204958, + 0.02332441136240959, + 0.015181380324065685, + 0.01899927482008934, + 0.013263644650578499, + 0.027127908542752266, + 0.02313394472002983, + 0.018334364518523216, + 0.017758075147867203, + 0.027161234989762306, + 0.04241464287042618, + 0.01718239113688469, + 0.024021489545702934, + 0.014589673839509487, + 0.007900794968008995, + 0.016625069081783295, + 0.011193300597369671, + 0.01825045607984066, + 0.012327221222221851, + 0.01661183498799801, + 0.01339916791766882, + 0.01591777615249157, + 0.009554383344948292, + 0.015387538820505142, + 0.010017065331339836, + 0.013872920535504818, + 0.01533707045018673, + 0.012371299788355827, + 0.01400004979223013, + 0.024597279727458954, + 0.009551065973937511, + 0.021507445722818375, + 0.014001153409481049, + 0.013590903952717781, + 0.016694018617272377, + 0.014215844683349133, + 0.021128110587596893, + 0.0060878051444888115, + 0.005373553838580847, + 0.005046238657087088, + 0.0073677124455571175, + 0.007143144030123949, + 0.0042016818188130856, + 0.008122979663312435, + 0.005388668738305569, + 0.0027519427239894867, + 0.002567006042227149, + 0.0032942802645266056, + 0.003664771094918251, + 0.004427856300026178, + 0.002682314720004797, + 0.004866091534495354, + 0.003413981758058071, + 0.003712290432304144, + 0.004551009740680456, + 0.0034827201161533594, + 0.004644372500479221, + 0.008963754400610924, + 0.004231610335409641, + 0.005722925998270512, + 0.0030415281653404236, + 0.004097821656614542, + 0.0036298634950071573, + 0.004747970029711723, + 0.004187845624983311, + 0.005327467806637287, + 0.0033912700600922108, + 0.00443224748596549, + 0.0023307488299906254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204402819275856 + ], + "opposite_logprob_tokens": [ + -2.812044143676758 + ], + "logprob_reward": -0.06204402819275856, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204402819275856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19688548147678375, + 0.1334732323884964, + 0.09646227955818176, + 0.13686811923980713, + 0.1755194067955017, + 0.08149188756942749, + 0.05125412344932556, + 0.04835529252886772, + 0.048758041113615036, + 0.03093216009438038 + ], + "credit_map": [ + 0.020090240985155106, + 0.016001049429178238, + 0.022350303828716278, + 0.037660352885723114, + 0.030135754495859146, + 0.021522575989365578, + 0.03259284049272537, + 0.016532354056835175, + 0.011091609485447407, + 0.016171079128980637, + 0.015354135073721409, + 0.016344141215085983, + 0.024471724405884743, + 0.013303433544933796, + 0.02350689098238945, + 0.01323021948337555, + 0.00972621887922287, + 0.010748141445219517, + 0.014083500020205975, + 0.011755267158150673, + 0.01252997387200594, + 0.008468441665172577, + 0.01604076288640499, + 0.013109972700476646, + 0.01563609391450882, + 0.0098934480920434, + 0.027372226119041443, + 0.015184997580945492, + 0.016103601083159447, + 0.01130391750484705, + 0.02248174138367176, + 0.018892090767621994, + 0.030033960938453674, + 0.015615923330187798, + 0.018773095682263374, + 0.015578540042042732, + 0.02363114431500435, + 0.026397252455353737, + 0.01991037093102932, + 0.025579113513231277, + 0.020851904526352882, + 0.006072500720620155, + 0.005413747858256102, + 0.004652861971408129, + 0.012270528823137283, + 0.011144371703267097, + 0.006077195983380079, + 0.015008771792054176, + 0.004513885360211134, + 0.005393035709857941, + 0.007156587205827236, + 0.007494464050978422, + 0.008288288488984108, + 0.005250639282166958, + 0.00704533327370882, + 0.006111889146268368, + 0.004121399950236082, + 0.005661309231072664, + 0.004907754249870777, + 0.007503024302423, + 0.009878421202301979, + 0.004831959027796984, + 0.008038176223635674, + 0.003413248108699918, + 0.00786252785474062, + 0.004949052818119526, + 0.004203892778605223, + 0.005877206567674875, + 0.007601488381624222, + 0.0071805682964622974, + 0.006747029721736908, + 0.004336273763328791, + 0.0038240738213062286, + 0.003030543215572834, + 0.0025336756370961666, + 0.005234934855252504, + 0.004817702807486057, + 0.003185961628332734, + 0.0058792089112102985, + 0.002426058752462268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.06204402819275856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808743134140968 + ], + "opposite_logprob_tokens": [ + -3.2880873680114746 + ], + "logprob_reward": -0.03808743134140968, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808743134140968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11424407362937927, + 0.09416884928941727, + 0.16164085268974304, + 0.33668968081474304, + 0.11173257976770401, + 0.06716520339250565, + 0.05637740716338158, + 0.018302666023373604, + 0.02120986022055149, + 0.018468838185071945 + ], + "credit_map": [ + 0.010898111388087273, + 0.011410078033804893, + 0.013399539515376091, + 0.0164622999727726, + 0.0157550610601902, + 0.01625855639576912, + 0.014234057627618313, + 0.015826372429728508, + 0.00934853870421648, + 0.008798948489129543, + 0.011766890063881874, + 0.012338812462985516, + 0.01426235307008028, + 0.011677267961204052, + 0.01577596552670002, + 0.010200077667832375, + 0.020197873935103416, + 0.01993100717663765, + 0.0174409169703722, + 0.018688974902033806, + 0.015116109512746334, + 0.03277556225657463, + 0.018036963418126106, + 0.019453449174761772, + 0.035154759883880615, + 0.019834520295262337, + 0.0751991719007492, + 0.039089418947696686, + 0.02310543693602085, + 0.02966579608619213, + 0.06328893452882767, + 0.0513516329228878, + 0.017739450559020042, + 0.00930777471512556, + 0.012315123341977596, + 0.010217870585620403, + 0.01350343506783247, + 0.02030763030052185, + 0.013024069368839264, + 0.015317214652895927, + 0.010575780645012856, + 0.00713355140760541, + 0.008938813582062721, + 0.007143889553844929, + 0.007776283193379641, + 0.0072221048176288605, + 0.008646012283861637, + 0.009728765115141869, + 0.005461071152240038, + 0.00624192925170064, + 0.006394393742084503, + 0.00901259295642376, + 0.009011253714561462, + 0.005536061245948076, + 0.008594865910708904, + 0.0061252363957464695, + 0.002189378021284938, + 0.0020653768442571163, + 0.001744435983709991, + 0.002420959295704961, + 0.0034864479675889015, + 0.0021867691539227962, + 0.0029922688845545053, + 0.0012170295231044292, + 0.002682003192603588, + 0.002293064258992672, + 0.0023920994717627764, + 0.0024113149847835302, + 0.0032825456000864506, + 0.002899367595091462, + 0.0033214257564395666, + 0.0019280395936220884, + 0.0018120550084859133, + 0.0014841725351288915, + 0.0015911469236016273, + 0.0020889185834676027, + 0.004805802833288908, + 0.0020237911958247423, + 0.003468448296189308, + 0.0011945025762543082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03808743134140968 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, medium shot, shot from above, focuses with powerful visual intensity on a young deity with flowing black hair, clad in an intricately crafted suit of crimson lotus armor, standing proudly amidst a devastating mountain torrent. The armor, seemingly forged from flaming lotus petals, each blade a deep crimson, gleaming with a faint golden hue at its edges, its surface shimmering with the ripples of water swept by the rapids. Beneath a sky thick with dark clouds and thunder, the dynamic, realistic play of light and shadow shifts constantly, each lightning flash illuminating his resolute and stern young face and fearless eyes that seem to hold a star. The turbulent torrent crashes upon him with unstoppable force, yet he remains unmoved, as steady as a rock. At this moment, the camera captures a stunning slow-motion particle effect: the massive splash of water from the impact is broken into countless individual droplets, which solidify in the air like suspended crystals and shimmering diamond dust. Each droplet clearly reflects the cold light of the lightning and the fiery red glow of the armor, creating a striking contrast between cold and warm. The entire scene, imbued with unparalleled tension in this extreme contrast between movement and stillness, showcases the absolute majesty and composure of the deity against the violent forces of nature.", + "seed": 16548, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_0/index_4989_seed_16548_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": -0.026448157383129, + "vlm_logprob": -0.026448157383129, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.026448157383129, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00815750751644373 + ], + "opposite_logprob_tokens": [ + -5.133157730102539 + ], + "logprob_reward": -0.00815750751644373, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00815750751644373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23654673993587494, + 0.08710793405771255, + 0.1435960829257965, + 0.04841718077659607, + 0.1110321506857872, + 0.03933080658316612, + 0.08330493420362473, + 0.09144214540719986, + 0.055944960564374924, + 0.10327703505754471 + ], + "credit_map": [ + 0.04642263799905777, + 0.019412163645029068, + 0.05609780550003052, + 0.01845189929008484, + 0.02851572260260582, + 0.025865312665700912, + 0.0202710572630167, + 0.021510157734155655, + 0.010899445042014122, + 0.009609486907720566, + 0.008554999716579914, + 0.008724626153707504, + 0.010277044028043747, + 0.014905205927789211, + 0.012785864062607288, + 0.011351266875863075, + 0.017697148025035858, + 0.013236364349722862, + 0.016490068286657333, + 0.026081016287207603, + 0.011641128920018673, + 0.018983295187354088, + 0.02463972009718418, + 0.014827361330389977, + 0.007163574453443289, + 0.0054859803058207035, + 0.005862791556864977, + 0.00936987902969122, + 0.005473626311868429, + 0.004155567381531, + 0.006109048146754503, + 0.004796721041202545, + 0.01311811339110136, + 0.00929480791091919, + 0.013796955347061157, + 0.02775823138654232, + 0.010580749250948429, + 0.008807486854493618, + 0.017731431871652603, + 0.00994438212364912, + 0.005991431884467602, + 0.004052396863698959, + 0.0050095911137759686, + 0.006909522693604231, + 0.004731356631964445, + 0.0022680629044771194, + 0.006418935023248196, + 0.003949512727558613, + 0.024282166734337807, + 0.00487445667386055, + 0.027438677847385406, + 0.0069451360031962395, + 0.005467540584504604, + 0.004052537959069014, + 0.006584285292774439, + 0.00366014800965786, + 0.018801197409629822, + 0.006160188000649214, + 0.021906940266489983, + 0.010184038430452347, + 0.012698431499302387, + 0.0063741677440702915, + 0.009726116433739662, + 0.0055910805240273476, + 0.009936032816767693, + 0.0038448814302682877, + 0.010278141126036644, + 0.005494230426847935, + 0.011984694749116898, + 0.004020245745778084, + 0.0066973064094781876, + 0.0036894320510327816, + 0.02810773439705372, + 0.00843453872948885, + 0.025295237079262733, + 0.007225099951028824, + 0.010686037130653858, + 0.006139387376606464, + 0.010109182447195053, + 0.007279824465513229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.00815750751644373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036208562087267637 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036208562087267637, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036208562087267637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13754291832447052, + 0.2679271399974823, + 0.049107711762189865, + 0.06155101954936981, + 0.10809893906116486, + 0.045134287327528, + 0.03212399780750275, + 0.06791761517524719, + 0.05402626097202301, + 0.1765700876712799 + ], + "credit_map": [ + 0.028637604787945747, + 0.014377683401107788, + 0.03285255283117294, + 0.012025425210595131, + 0.012480087578296661, + 0.014124573208391666, + 0.010177258402109146, + 0.012867740355432034, + 0.04988640174269676, + 0.02113257348537445, + 0.02655915915966034, + 0.035975538194179535, + 0.026701562106609344, + 0.03126126527786255, + 0.042544201016426086, + 0.03386645019054413, + 0.007641112431883812, + 0.003745904890820384, + 0.00529493298381567, + 0.009035909548401833, + 0.0035910361912101507, + 0.007469139993190765, + 0.007398287765681744, + 0.0049313888885080814, + 0.007200679741799831, + 0.006112902425229549, + 0.00982632115483284, + 0.01730009727180004, + 0.0044687348417937756, + 0.0038758632726967335, + 0.007908019237220287, + 0.004858399275690317, + 0.01856519840657711, + 0.010950284078717232, + 0.01325168926268816, + 0.020756816491484642, + 0.009144186973571777, + 0.012080936692655087, + 0.011972308158874512, + 0.011377528309822083, + 0.006833304185420275, + 0.004117405973374844, + 0.005592092871665955, + 0.009270570240914822, + 0.004789769183844328, + 0.002896103309467435, + 0.0076968371868133545, + 0.00393820321187377, + 0.008060218766331673, + 0.0025943368673324585, + 0.009155339561402798, + 0.002480415627360344, + 0.0035087382420897484, + 0.002008729847148061, + 0.0028284303843975067, + 0.001487787812948227, + 0.013264122419059277, + 0.005958925932645798, + 0.013933631591498852, + 0.010011394508183002, + 0.007819650694727898, + 0.004066389985382557, + 0.007749858312308788, + 0.005113638937473297, + 0.015036129392683506, + 0.002605670830234885, + 0.015788866207003593, + 0.0038768653757870197, + 0.007209450472146273, + 0.002290313597768545, + 0.004753004759550095, + 0.0024659635964781046, + 0.05905063822865486, + 0.009226052090525627, + 0.06065330281853676, + 0.008610889315605164, + 0.010568248108029366, + 0.009611228480935097, + 0.009753710590302944, + 0.00909601990133524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0036208562087267637 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781880304217339 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781880304217339, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781880304217339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1612991988658905, + 0.12053891271352768, + 0.17255614697933197, + 0.07628308981657028, + 0.06182708218693733, + 0.06400558352470398, + 0.08421476930379868, + 0.12581278383731842, + 0.05271579325199127, + 0.08074667304754257 + ], + "credit_map": [ + 0.0295358058065176, + 0.012894696556031704, + 0.031539835035800934, + 0.012431503273546696, + 0.011200751177966595, + 0.027165593579411507, + 0.02030618116259575, + 0.016224833205342293, + 0.023616651073098183, + 0.012983824126422405, + 0.014706969261169434, + 0.0128951296210289, + 0.01250283420085907, + 0.011826246976852417, + 0.016730230301618576, + 0.015277019701898098, + 0.02190248668193817, + 0.021780353039503098, + 0.0220783744007349, + 0.03200212121009827, + 0.014572950080037117, + 0.019580379128456116, + 0.02191086672246456, + 0.01872861757874489, + 0.012704513035714626, + 0.008036814630031586, + 0.009545858949422836, + 0.013421888463199139, + 0.009655061177909374, + 0.006513956002891064, + 0.008991280570626259, + 0.0074137113988399506, + 0.006632679607719183, + 0.009206508286297321, + 0.007368982769548893, + 0.014004002325236797, + 0.007196355611085892, + 0.004043941851705313, + 0.007393931038677692, + 0.005980677902698517, + 0.012302159331738949, + 0.0057508801110088825, + 0.009240583516657352, + 0.01308598555624485, + 0.005647442769259214, + 0.003853587433695793, + 0.009125657379627228, + 0.004999286495149136, + 0.025823164731264114, + 0.00414563762024045, + 0.026335768401622772, + 0.004420371726155281, + 0.010973810218274593, + 0.004611100070178509, + 0.004344669636338949, + 0.00356024201028049, + 0.03381466120481491, + 0.0075758276507258415, + 0.036810725927352905, + 0.010785712860524654, + 0.012278584763407707, + 0.006055937148630619, + 0.011352146975696087, + 0.007139191497117281, + 0.013509713113307953, + 0.0033866134472191334, + 0.014469082467257977, + 0.00440922100096941, + 0.006215496454387903, + 0.002538552274927497, + 0.005296709481626749, + 0.0028903980273753405, + 0.025038570165634155, + 0.005279684439301491, + 0.023722177371382713, + 0.005097128450870514, + 0.005401052068918943, + 0.005446215625852346, + 0.005542566534131765, + 0.005219276528805494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.009781880304217339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042683362960815 + ], + "opposite_logprob_tokens": [ + -2.812042713165283 + ], + "logprob_reward": -0.062042683362960815, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042683362960815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37316834926605225, + 0.06527351588010788, + 0.1475103199481964, + 0.0924445167183876, + 0.08701185882091522, + 0.04943562671542168, + 0.035847801715135574, + 0.058527953922748566, + 0.053040869534015656, + 0.03773914650082588 + ], + "credit_map": [ + 0.07921095192432404, + 0.024100882932543755, + 0.09011542797088623, + 0.027133837342262268, + 0.02348317764699459, + 0.050904467701911926, + 0.03510262072086334, + 0.04311700165271759, + 0.008382407017052174, + 0.005409063305705786, + 0.006547903176397085, + 0.008012205362319946, + 0.008717449381947517, + 0.008628135547041893, + 0.009962443262338638, + 0.00961391068994999, + 0.02030070126056671, + 0.01454063318669796, + 0.01777978613972664, + 0.03003520518541336, + 0.010091773234307766, + 0.017309371381998062, + 0.019327975809574127, + 0.018124885857105255, + 0.012213767506182194, + 0.010841875337064266, + 0.011013529263436794, + 0.016431139782071114, + 0.011046462692320347, + 0.005715546198189259, + 0.010981311090290546, + 0.014200887642800808, + 0.008375401608645916, + 0.008613832294940948, + 0.010093986056745052, + 0.018019577488303185, + 0.009901837445795536, + 0.007016296498477459, + 0.01354085560888052, + 0.011450069956481457, + 0.007075564935803413, + 0.004531456623226404, + 0.006729273125529289, + 0.009881716221570969, + 0.005882759112864733, + 0.002728146966546774, + 0.007183543872088194, + 0.005423168186098337, + 0.008005676791071892, + 0.002565559931099415, + 0.008376115933060646, + 0.0023684741463512182, + 0.007161453366279602, + 0.002489960752427578, + 0.00300481915473938, + 0.0018757445504888892, + 0.009832273237407207, + 0.0054686167277395725, + 0.012505472637712955, + 0.007919310592114925, + 0.006819969043135643, + 0.004348142072558403, + 0.007343926001340151, + 0.004290246870368719, + 0.01662057265639305, + 0.002652493305504322, + 0.017559664323925972, + 0.0035291751846671104, + 0.004260951187461615, + 0.0016694143414497375, + 0.004557479172945023, + 0.0021911158692091703, + 0.007833817973732948, + 0.0024676218163222075, + 0.008221730589866638, + 0.003961576148867607, + 0.003638974390923977, + 0.003082146868109703, + 0.003850834211334586, + 0.004682445432990789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.062042683362960815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048637859523296356 + ], + "opposite_logprob_tokens": [ + -3.048637866973877 + ], + "logprob_reward": -0.048637859523296356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048637859523296356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1455170065164566, + 0.1358257234096527, + 0.1800437718629837, + 0.14348411560058594, + 0.14611473679542542, + 0.023143302649259567, + 0.049601733684539795, + 0.10208478569984436, + 0.0173676535487175, + 0.056817181408405304 + ], + "credit_map": [ + 0.022786656394600868, + 0.02618538774549961, + 0.02565835602581501, + 0.01669415459036827, + 0.012035010382533073, + 0.01669309101998806, + 0.011687442660331726, + 0.013776886276900768, + 0.017728479579091072, + 0.01463723462074995, + 0.012727665714919567, + 0.018505072221159935, + 0.014025361277163029, + 0.012850687839090824, + 0.02591053396463394, + 0.019440675154328346, + 0.023385951295495033, + 0.01968279480934143, + 0.02304690331220627, + 0.03742852434515953, + 0.014735412783920765, + 0.017993222922086716, + 0.023270683363080025, + 0.020500270649790764, + 0.021970264613628387, + 0.014996635727584362, + 0.02172282710671425, + 0.03205278888344765, + 0.01123382244259119, + 0.011241193860769272, + 0.017378272488713264, + 0.012888289988040924, + 0.01702098548412323, + 0.014637078158557415, + 0.01639372482895851, + 0.034583184868097305, + 0.012286653742194176, + 0.011602560989558697, + 0.024174928665161133, + 0.015415593050420284, + 0.0036127781495451927, + 0.001834554597735405, + 0.003022401127964258, + 0.0036402761470526457, + 0.0023021141532808542, + 0.001904580625705421, + 0.0042834714986383915, + 0.0025431246031075716, + 0.009626252576708794, + 0.004034509416669607, + 0.009185408242046833, + 0.006008544471114874, + 0.005593347828835249, + 0.005243393126875162, + 0.004706236068159342, + 0.005204038694500923, + 0.02880997397005558, + 0.006422704551368952, + 0.031208399683237076, + 0.007418766152113676, + 0.009465680457651615, + 0.005188023671507835, + 0.007892794907093048, + 0.005678434856235981, + 0.0022897475864738226, + 0.0017897391226142645, + 0.0038939504884183407, + 0.0020716500002890825, + 0.0027721121441572905, + 0.0013574904296547174, + 0.0015986587386578321, + 0.0015943036414682865, + 0.012374292127788067, + 0.003903017845004797, + 0.013160820119082928, + 0.005522006656974554, + 0.005831572227180004, + 0.004570582415908575, + 0.006316670682281256, + 0.005138214211910963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.048637859523296356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, medium shot, shot from above, focuses with powerful visual intensity on a young deity with flowing black hair, clad in an intricately crafted suit of crimson lotus armor, standing proudly amidst a devastating mountain torrent. The armor, seemingly forged from flaming lotus petals, each blade a deep crimson, gleaming with a faint golden hue at its edges, its surface shimmering with the ripples of water swept by the rapids. Beneath a sky thick with dark clouds and thunder, the dynamic, realistic play of light and shadow shifts constantly, each lightning flash illuminating his resolute and stern young face and fearless eyes that seem to hold a star. The turbulent torrent crashes upon him with unstoppable force, yet he remains unmoved, as steady as a rock. At this moment, the camera captures a stunning slow-motion particle effect: the massive splash of water from the impact is broken into countless individual droplets, which solidify in the air like suspended crystals and shimmering diamond dust. Each droplet clearly reflects the cold light of the lightning and the fiery red glow of the armor, creating a striking contrast between cold and warm. The entire scene, imbued with unparalleled tension in this extreme contrast between movement and stillness, showcases the absolute majesty and composure of the deity against the violent forces of nature.", + "seed": 16549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_0/index_4989_seed_16549_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": -0.025798268243670465, + "vlm_logprob": -0.025798268243670465, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.025798268243670465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01946432515978813 + ], + "opposite_logprob_tokens": [ + -4.269464492797852 + ], + "logprob_reward": -0.01946432515978813, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01946432515978813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2058834284543991, + 0.03796286880970001, + 0.09362883865833282, + 0.15781685709953308, + 0.042248789221048355, + 0.039291925728321075, + 0.03984702005982399, + 0.040969666093587875, + 0.16858607530593872, + 0.17376458644866943 + ], + "credit_map": [ + 0.024817373603582382, + 0.020533010363578796, + 0.028502242639660835, + 0.015850773081183434, + 0.02319266088306904, + 0.0366973839700222, + 0.019998470321297646, + 0.03629150986671448, + 0.006967584136873484, + 0.005478534381836653, + 0.003858050564303994, + 0.003461990738287568, + 0.004394892603158951, + 0.004502936266362667, + 0.004596607759594917, + 0.004702274687588215, + 0.012652040459215641, + 0.009269153699278831, + 0.009465817362070084, + 0.009376013651490211, + 0.011143903248012066, + 0.01827598176896572, + 0.013921177014708519, + 0.00952475517988205, + 0.020602848380804062, + 0.015363352373242378, + 0.026701418682932854, + 0.023597192019224167, + 0.015098410658538342, + 0.016837801784276962, + 0.02832133136689663, + 0.01129449624568224, + 0.004963572137057781, + 0.004659918136894703, + 0.0042096772231161594, + 0.006513710133731365, + 0.005553793162107468, + 0.005100407171994448, + 0.006833592429757118, + 0.004414118826389313, + 0.005481590982526541, + 0.003652868326753378, + 0.006299483124166727, + 0.0051414319314062595, + 0.003541836980730295, + 0.0032956956420093775, + 0.008352179080247879, + 0.003526839194819331, + 0.007706278003752232, + 0.003712752601131797, + 0.0041303811594843864, + 0.005562537349760532, + 0.004056951496750116, + 0.004394807852804661, + 0.0066787744872272015, + 0.0036045368760824203, + 0.007617754861712456, + 0.004167343024164438, + 0.0051427860744297504, + 0.00626413244754076, + 0.003859534626826644, + 0.004086662549525499, + 0.005815918557345867, + 0.0040155332535505295, + 0.014588928781449795, + 0.017421573400497437, + 0.020635565742850304, + 0.03132801502943039, + 0.016786746680736542, + 0.01579408347606659, + 0.03309596702456474, + 0.01893519051373005, + 0.017023300752043724, + 0.018061861395835876, + 0.02203972451388836, + 0.0182599276304245, + 0.02789483778178692, + 0.018228823319077492, + 0.023700153455138206, + 0.028555946424603462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01946432515978813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036259638145565987 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.0036259638145565987, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036259638145565987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11659873276948929, + 0.1327255517244339, + 0.09863311797380447, + 0.1910587102174759, + 0.05913221091032028, + 0.07925280183553696, + 0.05303877592086792, + 0.0845743864774704, + 0.04778023809194565, + 0.13720542192459106 + ], + "credit_map": [ + 0.032819636166095734, + 0.012578612193465233, + 0.018126120790839195, + 0.01128510944545269, + 0.012860383838415146, + 0.008279399015009403, + 0.01152409240603447, + 0.009125389158725739, + 0.04011097550392151, + 0.016574181616306305, + 0.018223833292722702, + 0.009217708371579647, + 0.011622131802141666, + 0.011383556760847569, + 0.012416760437190533, + 0.013176404871046543, + 0.017864173278212547, + 0.0087638720870018, + 0.012196526862680912, + 0.0069372751750051975, + 0.0096209067851305, + 0.02402392029762268, + 0.008868982084095478, + 0.010357470251619816, + 0.032534174621105194, + 0.025664567947387695, + 0.030114823952317238, + 0.014690154232084751, + 0.021158581599593163, + 0.03597772866487503, + 0.014593135565519333, + 0.016325561329722404, + 0.009187228977680206, + 0.005248615518212318, + 0.012177628464996815, + 0.006743097677826881, + 0.007682078052312136, + 0.006371721159666777, + 0.007267430890351534, + 0.004454413428902626, + 0.019218841567635536, + 0.005991313140839338, + 0.016097821295261383, + 0.009946741163730621, + 0.00624006986618042, + 0.008071704767644405, + 0.008369951508939266, + 0.005316364578902721, + 0.011172065511345863, + 0.004972288850694895, + 0.005939240101724863, + 0.006654737051576376, + 0.005408423021435738, + 0.007288759108632803, + 0.007163115311414003, + 0.004440146964043379, + 0.013285711407661438, + 0.006091061979532242, + 0.010672084055840969, + 0.013518263585865498, + 0.00853952020406723, + 0.007752942852675915, + 0.011724308133125305, + 0.012990504503250122, + 0.007218681275844574, + 0.0034117521718144417, + 0.007791317068040371, + 0.006682938430458307, + 0.005283169448375702, + 0.0037788189947605133, + 0.005078782793134451, + 0.008534780703485012, + 0.012685847468674183, + 0.012573289684951305, + 0.0197602566331625, + 0.0186700951308012, + 0.01365517359226942, + 0.013720342889428139, + 0.022828515619039536, + 0.023311922326683998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.0036259638145565987 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06338043510913849, + 0.03285723179578781, + 0.20455937087535858, + 0.11896618455648422, + 0.039783962070941925, + 0.049258455634117126, + 0.03137153014540672, + 0.10721148550510406, + 0.2876295745372772, + 0.06498175859451294 + ], + "credit_map": [ + 0.01174604706466198, + 0.009409320540726185, + 0.007700304966419935, + 0.006780030205845833, + 0.007985198870301247, + 0.006350435316562653, + 0.006965750828385353, + 0.006443358492106199, + 0.006559488829225302, + 0.0035322231706231833, + 0.0037385609466582537, + 0.0029540720861405134, + 0.004778061527758837, + 0.00317421299405396, + 0.004750180523842573, + 0.0033704328816384077, + 0.030060771852731705, + 0.03527506813406944, + 0.019940996542572975, + 0.021864600479602814, + 0.02310858480632305, + 0.033214814960956573, + 0.02694706991314888, + 0.01414749026298523, + 0.022289272397756577, + 0.010698284953832626, + 0.013847808353602886, + 0.006634049117565155, + 0.009695250540971756, + 0.03470762446522713, + 0.008379791863262653, + 0.012714105658233166, + 0.0061856722459197044, + 0.0041345409117639065, + 0.00481631001457572, + 0.004485718905925751, + 0.00533463666215539, + 0.006329949013888836, + 0.00448852451518178, + 0.0040086121298372746, + 0.011135793291032314, + 0.0029070975724607706, + 0.010306330397725105, + 0.004955172538757324, + 0.003503581276163459, + 0.0044245547614991665, + 0.007576048374176025, + 0.004449883475899696, + 0.007031810004264116, + 0.0034481598995625973, + 0.0036065359599888325, + 0.004195413552224636, + 0.0035615572705864906, + 0.004115364048629999, + 0.0029283110052347183, + 0.0024843791034072638, + 0.008917550556361675, + 0.006936103571206331, + 0.01549487840384245, + 0.013232736848294735, + 0.008646154776215553, + 0.012419942766427994, + 0.017736824229359627, + 0.02382729947566986, + 0.025025175884366035, + 0.027332747355103493, + 0.03823300078511238, + 0.0569218210875988, + 0.027705594897270203, + 0.025206947699189186, + 0.052634868770837784, + 0.034569449722766876, + 0.005017678253352642, + 0.006302088964730501, + 0.009557468816637993, + 0.010376163758337498, + 0.006904599256813526, + 0.00468905083835125, + 0.013294060714542866, + 0.008840653114020824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043070822954177856 + ], + "opposite_logprob_tokens": [ + -3.1680707931518555 + ], + "logprob_reward": -0.043070822954177856, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043070822954177856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10806484520435333, + 0.05603078380227089, + 0.12898783385753632, + 0.11322605609893799, + 0.12073668837547302, + 0.1394623965024948, + 0.029649706557393074, + 0.06993149220943451, + 0.1423131376504898, + 0.09159708023071289 + ], + "credit_map": [ + 0.013845635578036308, + 0.014628062956035137, + 0.01341264694929123, + 0.011371047236025333, + 0.017107857391238213, + 0.011182144284248352, + 0.012802209705114365, + 0.01371524203568697, + 0.00792616792023182, + 0.004384378902614117, + 0.006019879132509232, + 0.006774818059056997, + 0.006470755208283663, + 0.005813744850456715, + 0.009835121221840382, + 0.008805916644632816, + 0.017542459070682526, + 0.0191646721214056, + 0.018867451697587967, + 0.013708577491343021, + 0.016751043498516083, + 0.019679652526974678, + 0.013058160431683064, + 0.010215814225375652, + 0.020224550738930702, + 0.020861733704805374, + 0.01961093209683895, + 0.008506678976118565, + 0.012024014256894588, + 0.01673172600567341, + 0.007438306696712971, + 0.007828117348253727, + 0.016389846801757812, + 0.010309810750186443, + 0.017940061166882515, + 0.013422020711004734, + 0.013155021704733372, + 0.020801793783903122, + 0.017511073499917984, + 0.01120705995708704, + 0.03471820056438446, + 0.008296613581478596, + 0.03808336332440376, + 0.01647908054292202, + 0.007510831579566002, + 0.012606787495315075, + 0.013093938119709492, + 0.00867357850074768, + 0.006632653065025806, + 0.0027054098900407553, + 0.003714718157425523, + 0.004750750493258238, + 0.002646326320245862, + 0.003643724136054516, + 0.003515297546982765, + 0.0020408297423273325, + 0.006992831826210022, + 0.005330657120794058, + 0.009502534754574299, + 0.009693522937595844, + 0.0062947398982942104, + 0.007272666320204735, + 0.011173428036272526, + 0.013671115040779114, + 0.012628061696887016, + 0.014798211865127087, + 0.017169782891869545, + 0.02509237267076969, + 0.015001807361841202, + 0.01360574085265398, + 0.026969609782099724, + 0.01704755239188671, + 0.006917348597198725, + 0.007713041268289089, + 0.016325870528817177, + 0.011810758151113987, + 0.007786461152136326, + 0.007012450601905584, + 0.010475922375917435, + 0.023555226624011993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.043070822954177856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young deity with flowing black hair wearing crimson lotus armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character standing amidst a mountain torrent with dark clouds and lightning in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character remain unmoved and steady while the water crashes upon them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene shot from an elevated angle as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show a slow-motion particle effect where water droplets solidify in the air like crystals?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863763228058815 + ], + "opposite_logprob_tokens": [ + -3.048637628555298 + ], + "logprob_reward": -0.04863763228058815, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863763228058815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13744746148586273, + 0.05013136565685272, + 0.207699716091156, + 0.136557474732399, + 0.0308513380587101, + 0.04130125790834427, + 0.03607970103621483, + 0.05131757631897926, + 0.17420145869255066, + 0.13441263139247894 + ], + "credit_map": [ + 0.0217131320387125, + 0.01771165430545807, + 0.018012864515185356, + 0.010731066577136517, + 0.016399594023823738, + 0.01737179607152939, + 0.012558274902403355, + 0.022949080914258957, + 0.010873142629861832, + 0.00682963477447629, + 0.006319172214716673, + 0.004132342990487814, + 0.005365689285099506, + 0.004297402687370777, + 0.006657609716057777, + 0.00565637182444334, + 0.03941386565566063, + 0.017342325299978256, + 0.028627613559365273, + 0.0118815703317523, + 0.015231628902256489, + 0.05649558827280998, + 0.015425649471580982, + 0.02328147552907467, + 0.02390783280134201, + 0.012036914937198162, + 0.021566160023212433, + 0.009356840513646603, + 0.013116319663822651, + 0.031576115638017654, + 0.012491722591221333, + 0.012505562976002693, + 0.004268835298717022, + 0.0031330985948443413, + 0.004967174492776394, + 0.003813717979937792, + 0.0038179310504347086, + 0.003447058377787471, + 0.003746211528778076, + 0.003657311201095581, + 0.008592573925852776, + 0.003831499721854925, + 0.008551263250410557, + 0.00429088668897748, + 0.003958831541240215, + 0.003820979967713356, + 0.005335389170795679, + 0.002919835038483143, + 0.005382178351283073, + 0.0029936411883682013, + 0.004140681587159634, + 0.0053631276823580265, + 0.003630861407145858, + 0.004347372800111771, + 0.006736104842275381, + 0.003485732711851597, + 0.008647559210658073, + 0.004006048198789358, + 0.00732852378860116, + 0.007798219099640846, + 0.0033501440193504095, + 0.004805300384759903, + 0.0074792602099478245, + 0.007902523502707481, + 0.024758726358413696, + 0.013096771202981472, + 0.024372899904847145, + 0.029221680015325546, + 0.01380099356174469, + 0.018359528854489326, + 0.020046234130859375, + 0.030544621869921684, + 0.01588745228946209, + 0.009705837815999985, + 0.019791115075349808, + 0.020423268899321556, + 0.009884233586490154, + 0.013082156889140606, + 0.022485530003905296, + 0.0231530349701643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.04863763228058815 + } + ], + "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/0000058/group_1/index_tensor([8819, 8819])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_tensor([8819, 8819])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0d6a1e559ddc5d875d554129df905626a76ff119 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_tensor([8819, 8819])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro close-up, captured from a horizontal angle, plunges us into a microscopic world dominated by intense black and white. A massive praying mantis lurks on a broad leaf, its angular body bathed in a pure, luminous white, half shrouded in a bottomless, inky shadow, illuminated by the harsh overhead light. Its triangular head shifts slightly, its enormous compound eyes reflecting the light with an inorganic, obsidian sheen. Its body remains motionless, like a meticulously crafted sculpture of death. Not far below, an unsuspecting fly rests on the leaf's edge, its fine hairs clearly visible in the strong light as it leisurely cleans its forelegs, oblivious to its impending doom. Suddenly, the mantis's folded, spiky forelimbs, like a wound-up trap, flung out with lightning speed. The movement was swift and decisive, leaving a blurry white afterimage in the air. It instantly bridged the gap between the two and opened with unerring precision, seizing the fragile body of the fly. The entire image, rendered in the texture of a high-contrast black-and-white film, the interplay of light and shadow imbues this microscopic struggle for survival with immense drama and visual impact, creating an atmosphere of intense concentration and primal instinct.", + "seed": 10698, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_8819_seed_10698_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": -0.22706825844943523, + "vlm_logprob": -0.22706825844943523, + "vlm_yes_no_margin": 3.450000011920929, + "vlm_token_credit": -0.22706825844943523, + "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 an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0184539332985878 + ], + "opposite_logprob_tokens": [ + -4.643454074859619 + ], + "logprob_reward": -0.0184539332985878, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0184539332985878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20163145661354065, + 0.0639645978808403, + 0.07219850271940231, + 0.07912319898605347, + 0.0451255701482296, + 0.0836491659283638, + 0.13458779454231262, + 0.09118510037660599, + 0.17783308029174805, + 0.050701532512903214 + ], + "credit_map": [ + 0.029374582692980766, + 0.018599865958094597, + 0.02924520894885063, + 0.02112104929983616, + 0.019036054611206055, + 0.027823418378829956, + 0.024028152227401733, + 0.0324031300842762, + 0.011543274857103825, + 0.0064095729030668736, + 0.008927654474973679, + 0.007370227016508579, + 0.007050703279674053, + 0.007319123949855566, + 0.009165355935692787, + 0.006178686860948801, + 0.010243823751807213, + 0.007091921288520098, + 0.010279451496899128, + 0.006473307963460684, + 0.006501290015876293, + 0.009265128523111343, + 0.007308108266443014, + 0.015035468153655529, + 0.011199806816875935, + 0.007745224982500076, + 0.010661199688911438, + 0.006071063689887524, + 0.007952549494802952, + 0.007097065914422274, + 0.019819527864456177, + 0.008576758205890656, + 0.007371343672275543, + 0.005812428891658783, + 0.006005971692502499, + 0.003397245891392231, + 0.009425302036106586, + 0.003992524463683367, + 0.00424273731186986, + 0.0048780133947730064, + 0.013013238087296486, + 0.006631182972341776, + 0.011581497266888618, + 0.008478940464556217, + 0.016058998182415962, + 0.009822600521147251, + 0.011668767780065536, + 0.006393933668732643, + 0.013565300963819027, + 0.010228198021650314, + 0.014527260325849056, + 0.016433250159025192, + 0.02292342111468315, + 0.01623871736228466, + 0.02870902605354786, + 0.011962621472775936, + 0.017130937427282333, + 0.0069430312141776085, + 0.012571757659316063, + 0.009391705505549908, + 0.00905616395175457, + 0.010410796850919724, + 0.011019687168300152, + 0.01466101873666048, + 0.02191889099776745, + 0.01638142764568329, + 0.019793838262557983, + 0.020008625462651253, + 0.01690656878054142, + 0.031530581414699554, + 0.022241471335291862, + 0.02905169501900673, + 0.009608684107661247, + 0.0035803101491183043, + 0.007097940426319838, + 0.006613537203520536, + 0.004263663198798895, + 0.006174841895699501, + 0.00743410037830472, + 0.005928453058004379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0184539332985878 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099145747721195 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099145747721195, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099145747721195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26638302206993103, + 0.10260609537363052, + 0.058260977268218994, + 0.06546727567911148, + 0.11748958379030228, + 0.05009123310446739, + 0.06034298241138458, + 0.04232291877269745, + 0.18874648213386536, + 0.04828941076993942 + ], + "credit_map": [ + 0.0416717529296875, + 0.03409296274185181, + 0.03765011578798294, + 0.029768355190753937, + 0.02627924457192421, + 0.03604981303215027, + 0.027160221710801125, + 0.0337105467915535, + 0.01901586726307869, + 0.012571612372994423, + 0.015751060098409653, + 0.013404729776084423, + 0.010819585993885994, + 0.012473884969949722, + 0.009317969903349876, + 0.009251389652490616, + 0.009036283940076828, + 0.007603736594319344, + 0.007955201901495457, + 0.008236967027187347, + 0.005242288578301668, + 0.006887906230986118, + 0.0061669060960412025, + 0.007131680846214294, + 0.012373806908726692, + 0.007732339669018984, + 0.010849017649888992, + 0.007431848905980587, + 0.005954831838607788, + 0.009357462637126446, + 0.006004744209349155, + 0.005763218272477388, + 0.02554185315966606, + 0.012787574902176857, + 0.017546899616718292, + 0.011732293292880058, + 0.017080847173929214, + 0.012292434461414814, + 0.010088001377880573, + 0.01041968259960413, + 0.005909669678658247, + 0.004561087116599083, + 0.007394210435450077, + 0.006075499579310417, + 0.01104008685797453, + 0.006113220937550068, + 0.0054918862879276276, + 0.003505573607981205, + 0.007859221659600735, + 0.006177784409373999, + 0.011085625737905502, + 0.008210767060518265, + 0.007998734712600708, + 0.009327824227511883, + 0.00485182972624898, + 0.004831200931221247, + 0.004967100918292999, + 0.004134980496019125, + 0.004644645377993584, + 0.003854538779705763, + 0.00507738022133708, + 0.008208829909563065, + 0.004322127904742956, + 0.007113313302397728, + 0.026665400713682175, + 0.02046968601644039, + 0.023970521986484528, + 0.02787279710173607, + 0.01938471756875515, + 0.026807324960827827, + 0.026171354576945305, + 0.017404675483703613, + 0.008159104734659195, + 0.004238687921315432, + 0.0070050605572760105, + 0.0068167634308338165, + 0.0032718225847929716, + 0.007279058452695608, + 0.0070879096165299416, + 0.004431001376360655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.011099145747721195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866942763328552 + ], + "opposite_logprob_tokens": [ + -3.0486693382263184 + ], + "logprob_reward": -0.04866942763328552, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866942763328552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17652054131031036, + 0.08982527256011963, + 0.05853038653731346, + 0.07160945981740952, + 0.06475304812192917, + 0.1150808334350586, + 0.21157504618167877, + 0.0678514689207077, + 0.07700749486684799, + 0.0672464668750763 + ], + "credit_map": [ + 0.03451325744390488, + 0.019166164100170135, + 0.02476615272462368, + 0.021741380915045738, + 0.016861673444509506, + 0.02157234586775303, + 0.02177482284605503, + 0.016124757006764412, + 0.014206166379153728, + 0.008639984764158726, + 0.011729137971997261, + 0.010753419250249863, + 0.00961772445589304, + 0.009835905395448208, + 0.014035450294613838, + 0.011007482185959816, + 0.008831317536532879, + 0.005661490373313427, + 0.008154346607625484, + 0.006524159573018551, + 0.0071943714283406734, + 0.006876756902784109, + 0.005991317797452211, + 0.009296626783907413, + 0.014155995100736618, + 0.00846831128001213, + 0.010209317319095135, + 0.006145316641777754, + 0.006268977653235197, + 0.006893439218401909, + 0.006780543830245733, + 0.012687558308243752, + 0.012579886242747307, + 0.008189107291400433, + 0.008987336419522762, + 0.007950118742883205, + 0.009898172691464424, + 0.0060382927767932415, + 0.005688172299414873, + 0.00542195700109005, + 0.012358264066278934, + 0.018625101074576378, + 0.014637758955359459, + 0.015223748981952667, + 0.015231933444738388, + 0.015193830244243145, + 0.013476012274622917, + 0.010334188118577003, + 0.021417561918497086, + 0.030435023829340935, + 0.030092915520071983, + 0.02213788963854313, + 0.022859999909996986, + 0.03851713240146637, + 0.030062217265367508, + 0.016052301973104477, + 0.011749805882573128, + 0.0073846750892698765, + 0.008356954902410507, + 0.008753682486712933, + 0.005661432631313801, + 0.007377253845334053, + 0.009588479064404964, + 0.008979177102446556, + 0.010140452533960342, + 0.00541625265032053, + 0.009005645290017128, + 0.009720399044454098, + 0.011082873679697514, + 0.012194275856018066, + 0.00937424972653389, + 0.0100733432918787, + 0.013351863250136375, + 0.005855249240994453, + 0.010145751759409904, + 0.009017457254230976, + 0.005224400199949741, + 0.00828856136649847, + 0.007418611086905003, + 0.007944571785628796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.04866942763328552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031963707879185677 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031963707879185677, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031963707879185677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14723195135593414, + 0.06509483605623245, + 0.07045397907495499, + 0.13143299520015717, + 0.04286788031458855, + 0.07562466710805893, + 0.08306162059307098, + 0.14808593690395355, + 0.1479271501302719, + 0.08821893483400345 + ], + "credit_map": [ + 0.020358920097351074, + 0.017707139253616333, + 0.020394761115312576, + 0.02038314938545227, + 0.017557986080646515, + 0.01602025143802166, + 0.016568120568990707, + 0.018241623416543007, + 0.010409140028059483, + 0.005587863270193338, + 0.00956766027957201, + 0.007548439782112837, + 0.010282227769494057, + 0.0059616840444505215, + 0.008822770789265633, + 0.006915045436471701, + 0.014599848538637161, + 0.005968272220343351, + 0.009250547736883163, + 0.006758225616067648, + 0.006945346482098103, + 0.006813836749643087, + 0.009309017099440098, + 0.010808881372213364, + 0.022874291986227036, + 0.018178094178438187, + 0.01773732528090477, + 0.019182264804840088, + 0.013187197037041187, + 0.010579564608633518, + 0.0178932286798954, + 0.011801033280789852, + 0.009393036365509033, + 0.005975135136395693, + 0.005958407185971737, + 0.003989613149315119, + 0.006531950086355209, + 0.003780858125537634, + 0.004084930755198002, + 0.003153951372951269, + 0.009341664612293243, + 0.007901967503130436, + 0.009511270560324192, + 0.007695682346820831, + 0.015057493932545185, + 0.011388778686523438, + 0.010342689231038094, + 0.0043851230293512344, + 0.010534215718507767, + 0.01539431419223547, + 0.012391449883580208, + 0.012619241140782833, + 0.0073700868524611, + 0.009184090420603752, + 0.009711393155157566, + 0.005856834817677736, + 0.025127343833446503, + 0.013415494933724403, + 0.01839006133377552, + 0.013128434307873249, + 0.012772560119628906, + 0.017581302672624588, + 0.01444779522716999, + 0.033222947269678116, + 0.018727995455265045, + 0.01021864265203476, + 0.015446377918124199, + 0.01813461072742939, + 0.01589014381170273, + 0.027968868613243103, + 0.022167110815644264, + 0.019373396411538124, + 0.016107255592942238, + 0.008886676281690598, + 0.01696031168103218, + 0.011117218993604183, + 0.008023655042052269, + 0.00998988188803196, + 0.009420620277523994, + 0.0077133155427873135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0031963707879185677 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.053922414779663 + ], + "opposite_logprob_tokens": [ + -0.42892247438430786 + ], + "logprob_reward": -1.053922414779663, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.053922414779663, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0856042429804802, + 0.026373347267508507, + 0.025953885167837143, + 0.04771588742733002, + 0.03230344504117966, + 0.0325162410736084, + 0.08369666337966919, + 0.21604782342910767, + 0.3921622633934021, + 0.05762611702084541 + ], + "credit_map": [ + 0.016027024015784264, + 0.008041518740355968, + 0.011704579927027225, + 0.009269095957279205, + 0.008262207731604576, + 0.008882245980203152, + 0.01381622813642025, + 0.009601349011063576, + 0.0048750233836472034, + 0.0024512961972504854, + 0.003652779618278146, + 0.0032312124967575073, + 0.00277409958653152, + 0.002398219658061862, + 0.0040528192184865475, + 0.0029378996696323156, + 0.0043683527037501335, + 0.002441820688545704, + 0.003728637006133795, + 0.0030495200771838427, + 0.0024398034438490868, + 0.0030279953498393297, + 0.003698652843013406, + 0.003199104918166995, + 0.009851142764091492, + 0.006022040266543627, + 0.006976379547268152, + 0.005172033794224262, + 0.004737435840070248, + 0.003853739472106099, + 0.006633319891989231, + 0.0044698030687868595, + 0.004500010050833225, + 0.0033506823237985373, + 0.0039745839312672615, + 0.0035499893128871918, + 0.004537573084235191, + 0.0038648357149213552, + 0.003831930458545685, + 0.0046938443556427956, + 0.00458961958065629, + 0.003815405536442995, + 0.0046997577883303165, + 0.005377467721700668, + 0.004981796257197857, + 0.0032389690168201923, + 0.0037008689250797033, + 0.0021123618353158236, + 0.00911442656069994, + 0.008309573866426945, + 0.009546687826514244, + 0.01410903874784708, + 0.007751401513814926, + 0.010964123532176018, + 0.016046294942498207, + 0.007855123840272427, + 0.03480201214551926, + 0.020034976303577423, + 0.02787444368004799, + 0.017720265313982964, + 0.016966398805379868, + 0.02504606731235981, + 0.0207922775298357, + 0.0528114028275013, + 0.05119306966662407, + 0.037001755088567734, + 0.0450722798705101, + 0.04494341090321541, + 0.04704012721776962, + 0.05465894564986229, + 0.049198947846889496, + 0.06305376440286636, + 0.012200743891298771, + 0.004490888677537441, + 0.009492015466094017, + 0.009027711115777493, + 0.0036873635835945606, + 0.005685650743544102, + 0.007718388922512531, + 0.005323357414454222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -1.053922414779663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a horizontal angle, plunges us into a microscopic world dominated by intense black and white. A massive praying mantis lurks on a broad leaf, its angular body bathed in a pure, luminous white, half shrouded in a bottomless, inky shadow, illuminated by the harsh overhead light. Its triangular head shifts slightly, its enormous compound eyes reflecting the light with an inorganic, obsidian sheen. Its body remains motionless, like a meticulously crafted sculpture of death. Not far below, an unsuspecting fly rests on the leaf's edge, its fine hairs clearly visible in the strong light as it leisurely cleans its forelegs, oblivious to its impending doom. Suddenly, the mantis's folded, spiky forelimbs, like a wound-up trap, flung out with lightning speed. The movement was swift and decisive, leaving a blurry white afterimage in the air. It instantly bridged the gap between the two and opened with unerring precision, seizing the fragile body of the fly. The entire image, rendered in the texture of a high-contrast black-and-white film, the interplay of light and shadow imbues this microscopic struggle for survival with immense drama and visual impact, creating an atmosphere of intense concentration and primal instinct.", + "seed": 10699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_8819_seed_10699_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": -0.11235809517093003, + "vlm_logprob": -0.11235809517093003, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.11235809517093003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020880404859781265 + ], + "opposite_logprob_tokens": [ + -4.645880222320557 + ], + "logprob_reward": -0.020880404859781265, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.020880404859781265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1611209362745285, + 0.08358612656593323, + 0.076480932533741, + 0.03618490323424339, + 0.039209529757499695, + 0.05708898603916168, + 0.18780124187469482, + 0.18026162683963776, + 0.10186655074357986, + 0.07639913260936737 + ], + "credit_map": [ + 0.016614967957139015, + 0.03110618330538273, + 0.014279243536293507, + 0.019523298367857933, + 0.03205520659685135, + 0.015701521188020706, + 0.01926836557686329, + 0.01257215067744255, + 0.007768620271235704, + 0.011855003423988819, + 0.006854950916022062, + 0.010849480517208576, + 0.010367035865783691, + 0.0127736646682024, + 0.013371000066399574, + 0.009746374562382698, + 0.008275774307549, + 0.009709823876619339, + 0.008003532886505127, + 0.012393420562148094, + 0.009394205175340176, + 0.012170708738267422, + 0.01028374396264553, + 0.006249731872230768, + 0.003600307274609804, + 0.00491880439221859, + 0.0027064094319939613, + 0.003923031035810709, + 0.005845499690622091, + 0.004653255920857191, + 0.005869477055966854, + 0.004668121691793203, + 0.0045923590660095215, + 0.0037824225146323442, + 0.0032108346931636333, + 0.004794824402779341, + 0.00437571806833148, + 0.004634692799299955, + 0.008756125345826149, + 0.005062556825578213, + 0.008887016214430332, + 0.005926546175032854, + 0.003951088525354862, + 0.0048018465749919415, + 0.006339469458907843, + 0.014576363377273083, + 0.007769737858325243, + 0.004836920648813248, + 0.021803133189678192, + 0.0209239162504673, + 0.01206315029412508, + 0.01260354183614254, + 0.030007751658558846, + 0.03919626772403717, + 0.030310437083244324, + 0.020893074572086334, + 0.020730067044496536, + 0.03427140787243843, + 0.014326684176921844, + 0.010491644963622093, + 0.0191289484500885, + 0.03603874891996384, + 0.02895679511129856, + 0.01631736196577549, + 0.016356151551008224, + 0.015293247997760773, + 0.011328507214784622, + 0.007862428203225136, + 0.01404801569879055, + 0.014776571653783321, + 0.01363700907677412, + 0.008564624935388565, + 0.010413880459964275, + 0.007178719155490398, + 0.007952127605676651, + 0.006814072839915752, + 0.016165325418114662, + 0.009953899309039116, + 0.009313728660345078, + 0.008607390336692333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.020880404859781265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005278341472148895 + ], + "opposite_logprob_tokens": [ + -5.25527811050415 + ], + "logprob_reward": -0.005278341472148895, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005278341472148895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19272558391094208, + 0.054358117282390594, + 0.07125452905893326, + 0.06526004523038864, + 0.11108312010765076, + 0.11726519465446472, + 0.0797036737203598, + 0.16827555000782013, + 0.08382266014814377, + 0.05625157803297043 + ], + "credit_map": [ + 0.019134096801280975, + 0.045776400715112686, + 0.01894155889749527, + 0.020541856065392494, + 0.023146798834204674, + 0.022029491141438484, + 0.0248103067278862, + 0.018345076590776443, + 0.005214963108301163, + 0.008314864709973335, + 0.004620121326297522, + 0.006025542505085468, + 0.00786285474896431, + 0.005917277652770281, + 0.011551426723599434, + 0.004851063713431358, + 0.0069241952151060104, + 0.007727826479822397, + 0.00581161817535758, + 0.007081052288413048, + 0.012183396145701408, + 0.009914199821650982, + 0.013926027342677116, + 0.007686218246817589, + 0.007198499981313944, + 0.006681820843368769, + 0.0052359686233103275, + 0.00532712833955884, + 0.008100458420813084, + 0.009607752785086632, + 0.014783037826418877, + 0.008325374685227871, + 0.01677057333290577, + 0.009664691984653473, + 0.010612770915031433, + 0.010440400801599026, + 0.017992034554481506, + 0.01849796064198017, + 0.014958278276026249, + 0.012146401219069958, + 0.01468341238796711, + 0.016468139365315437, + 0.008640037849545479, + 0.011512183584272861, + 0.016024500131607056, + 0.019993752241134644, + 0.020198630169034004, + 0.00974453054368496, + 0.010413013398647308, + 0.011026165448129177, + 0.008319927379488945, + 0.007716051302850246, + 0.010730663314461708, + 0.012646750546991825, + 0.010362958535552025, + 0.008488141931593418, + 0.020163429901003838, + 0.032879892736673355, + 0.014774384908378124, + 0.014808673411607742, + 0.015584128908813, + 0.03239152953028679, + 0.022798901423811913, + 0.014874604530632496, + 0.010551932267844677, + 0.017425784841179848, + 0.00878065824508667, + 0.0073189991526305676, + 0.007366131991147995, + 0.012371432036161423, + 0.012464798986911774, + 0.007542917039245367, + 0.009266890585422516, + 0.010945048183202744, + 0.004723832476884127, + 0.004281431902199984, + 0.007107060868293047, + 0.006990950088948011, + 0.007144298870116472, + 0.0057920655235648155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.005278341472148895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.202078714966774, + 0.04946031793951988, + 0.033548396080732346, + 0.09027037024497986, + 0.13061361014842987, + 0.10649926215410233, + 0.09473472833633423, + 0.13046741485595703, + 0.11979125440120697, + 0.042535945773124695 + ], + "credit_map": [ + 0.017974594607949257, + 0.025926919654011726, + 0.023442039266228676, + 0.029310517013072968, + 0.04010460525751114, + 0.019958756864070892, + 0.027720589190721512, + 0.017640691250562668, + 0.004785505589097738, + 0.006081452593207359, + 0.004454930312931538, + 0.0061360434629023075, + 0.007630972657352686, + 0.006196916103363037, + 0.007517862133681774, + 0.006656637880951166, + 0.0033132508397102356, + 0.0031828496139496565, + 0.003844600636512041, + 0.005482269451022148, + 0.004283777438104153, + 0.0053675370290875435, + 0.004446735605597496, + 0.0036273759324103594, + 0.008938094601035118, + 0.008764170110225677, + 0.006840549409389496, + 0.008814805187284946, + 0.01335034891963005, + 0.010650446638464928, + 0.02290916256606579, + 0.010002793744206429, + 0.014713097363710403, + 0.010723691433668137, + 0.01072496734559536, + 0.011660801246762276, + 0.032333679497241974, + 0.01882779411971569, + 0.01919308304786682, + 0.012436489574611187, + 0.012357084080576897, + 0.015699394047260284, + 0.0111179668456316, + 0.012860791757702827, + 0.011192313395440578, + 0.018853001296520233, + 0.01648884080350399, + 0.00792987272143364, + 0.011572254821658134, + 0.009757154621183872, + 0.012071839533746243, + 0.009032594971358776, + 0.011064549908041954, + 0.01527326088398695, + 0.016040876507759094, + 0.00992219801992178, + 0.018990488722920418, + 0.010984035208821297, + 0.009750358760356903, + 0.009372192434966564, + 0.017255134880542755, + 0.02056710049510002, + 0.02971155010163784, + 0.013836555182933807, + 0.01775585673749447, + 0.017933836206793785, + 0.016213439404964447, + 0.010975438170135021, + 0.01065074373036623, + 0.01850922405719757, + 0.01716136373579502, + 0.010591356083750725, + 0.006353882607072592, + 0.003864363767206669, + 0.0047324043698608875, + 0.0036669571418315172, + 0.006617506500333548, + 0.006605201400816441, + 0.005088058765977621, + 0.0056075737811625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.008645840920507908 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036176491994410753 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036176491994410753, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036176491994410753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11149027198553085, + 0.0740620419383049, + 0.033110350370407104, + 0.046154461801052094, + 0.107578344643116, + 0.09589386731386185, + 0.11439653486013412, + 0.22577650845050812, + 0.1592152863740921, + 0.03232235834002495 + ], + "credit_map": [ + 0.012136065401136875, + 0.010638536885380745, + 0.009412708692252636, + 0.009779563173651695, + 0.026842502877116203, + 0.01580771990120411, + 0.018017755821347237, + 0.008855420164763927, + 0.0077980593778193, + 0.007635502144694328, + 0.004727781750261784, + 0.007898789830505848, + 0.009192065335810184, + 0.01294323243200779, + 0.013625440187752247, + 0.010241164825856686, + 0.0034951369743794203, + 0.0036583140026777983, + 0.0031211324967443943, + 0.005524631589651108, + 0.004900822881609201, + 0.004823682364076376, + 0.005036345683038235, + 0.0025502860080450773, + 0.004697707016021013, + 0.0048943692818284035, + 0.0036724070087075233, + 0.004424841143190861, + 0.006921135354787111, + 0.006633814889937639, + 0.010178295895457268, + 0.0047318902797997, + 0.014441902749240398, + 0.009327736683189869, + 0.010773176327347755, + 0.009436821565032005, + 0.019616737961769104, + 0.015245103277266026, + 0.019930556416511536, + 0.008806315250694752, + 0.009432932361960411, + 0.015451690182089806, + 0.007874764502048492, + 0.011914748698472977, + 0.011609489098191261, + 0.012513917870819569, + 0.019425110891461372, + 0.007671212777495384, + 0.014582209289073944, + 0.014128035865724087, + 0.015610075555741787, + 0.011289549060165882, + 0.014404786750674248, + 0.01807240955531597, + 0.014441793784499168, + 0.011867676861584187, + 0.029162893071770668, + 0.026581980288028717, + 0.014780689962208271, + 0.017651988193392754, + 0.028171755373477936, + 0.041730690747499466, + 0.04520045965909958, + 0.022496061399579048, + 0.017828119918704033, + 0.025996102020144463, + 0.019776931032538414, + 0.01498932670801878, + 0.016494611278176308, + 0.022631553933024406, + 0.026861149817705154, + 0.014637484215199947, + 0.004548280965536833, + 0.0028677836526185274, + 0.0033966563642024994, + 0.004475998226553202, + 0.004135006573051214, + 0.004355643410235643, + 0.003976603504270315, + 0.004566385876387358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0036176491994410753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.523368239402771 + ], + "opposite_logprob_tokens": [ + -0.898368239402771 + ], + "logprob_reward": -0.523368239402771, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.523368239402771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2615128457546234, + 0.09878022223711014, + 0.04848342388868332, + 0.040754131972789764, + 0.10685007274150848, + 0.04113269969820976, + 0.06874854862689972, + 0.1930045187473297, + 0.06905777007341385, + 0.07167577743530273 + ], + "credit_map": [ + 0.03164304792881012, + 0.05173945426940918, + 0.02451729029417038, + 0.030638452619314194, + 0.032808005809783936, + 0.025766776874661446, + 0.04055563732981682, + 0.023844201117753983, + 0.009912754409015179, + 0.01225935760885477, + 0.007473141420632601, + 0.009827734902501106, + 0.016727007925510406, + 0.012164315208792686, + 0.018552076071500778, + 0.011863835155963898, + 0.005964428652077913, + 0.006589284632354975, + 0.004219706170260906, + 0.0056419167667627335, + 0.00674430001527071, + 0.006656713783740997, + 0.0074159931391477585, + 0.0052510821260511875, + 0.005013858899474144, + 0.004759966395795345, + 0.0030162958428263664, + 0.0038889050483703613, + 0.0068550193682312965, + 0.004792730789631605, + 0.008191701956093311, + 0.004235656466335058, + 0.009228212758898735, + 0.009404611773788929, + 0.007822888903319836, + 0.008672993630170822, + 0.023115573450922966, + 0.015424688346683979, + 0.02344818040728569, + 0.009732918813824654, + 0.00585003150627017, + 0.0028368132188916206, + 0.003277831943705678, + 0.00391063978895545, + 0.005651966203004122, + 0.01152131799608469, + 0.004826008807867765, + 0.0032580906990915537, + 0.006663863081485033, + 0.00750434584915638, + 0.007589645683765411, + 0.007296000607311726, + 0.01104584988206625, + 0.010829516686499119, + 0.011131009086966515, + 0.006688314490020275, + 0.021776478737592697, + 0.03602674603462219, + 0.011956336908042431, + 0.015078959986567497, + 0.020346272736787796, + 0.04162804037332535, + 0.02707028202712536, + 0.019121408462524414, + 0.007323444355279207, + 0.01566348969936371, + 0.0063733914867043495, + 0.00659343833103776, + 0.006976269651204348, + 0.009366529062390327, + 0.01016200054436922, + 0.006599207408726215, + 0.009500249288976192, + 0.01025962270796299, + 0.007168341428041458, + 0.009463392198085785, + 0.011400226503610611, + 0.010207101702690125, + 0.007230416405946016, + 0.006446428131312132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.523368239402771 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a horizontal angle, plunges us into a microscopic world dominated by intense black and white. A massive praying mantis lurks on a broad leaf, its angular body bathed in a pure, luminous white, half shrouded in a bottomless, inky shadow, illuminated by the harsh overhead light. Its triangular head shifts slightly, its enormous compound eyes reflecting the light with an inorganic, obsidian sheen. Its body remains motionless, like a meticulously crafted sculpture of death. Not far below, an unsuspecting fly rests on the leaf's edge, its fine hairs clearly visible in the strong light as it leisurely cleans its forelegs, oblivious to its impending doom. Suddenly, the mantis's folded, spiky forelimbs, like a wound-up trap, flung out with lightning speed. The movement was swift and decisive, leaving a blurry white afterimage in the air. It instantly bridged the gap between the two and opened with unerring precision, seizing the fragile body of the fly. The entire image, rendered in the texture of a high-contrast black-and-white film, the interplay of light and shadow imbues this microscopic struggle for survival with immense drama and visual impact, creating an atmosphere of intense concentration and primal instinct.", + "seed": 10700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_8819_seed_10700_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": -0.08353615382220596, + "vlm_logprob": -0.08353615382220596, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.08353615382220596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01360242534428835 + ], + "opposite_logprob_tokens": [ + -5.138602256774902 + ], + "logprob_reward": -0.01360242534428835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01360242534428835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0831710696220398, + 0.03084947168827057, + 0.1560346782207489, + 0.058052707463502884, + 0.0734841376543045, + 0.06105060502886772, + 0.13339892029762268, + 0.1956665813922882, + 0.03585164248943329, + 0.17244017124176025 + ], + "credit_map": [ + 0.008990741334855556, + 0.009348757565021515, + 0.010060463100671768, + 0.0102404048666358, + 0.008904334157705307, + 0.015006516128778458, + 0.007690323982387781, + 0.012929526157677174, + 0.0035012473817914724, + 0.0032412009313702583, + 0.0036333452444523573, + 0.0032691520173102617, + 0.0035609614569693804, + 0.005522707011550665, + 0.00341053563170135, + 0.004710322245955467, + 0.01646936871111393, + 0.016479171812534332, + 0.015134617686271667, + 0.02546085976064205, + 0.014412946067750454, + 0.028319166973233223, + 0.01779298484325409, + 0.02196555957198143, + 0.00508621521294117, + 0.008534464053809643, + 0.00630980683490634, + 0.007371852174401283, + 0.0065389834344387054, + 0.008480137214064598, + 0.006811985746026039, + 0.008919263258576393, + 0.006468151230365038, + 0.010804872028529644, + 0.007458197884261608, + 0.008076689206063747, + 0.011490533128380775, + 0.011653219349682331, + 0.007857397198677063, + 0.00967507902532816, + 0.00695394491776824, + 0.0066562616266310215, + 0.008068528026342392, + 0.007747696712613106, + 0.006571215111762285, + 0.011053383350372314, + 0.005716658663004637, + 0.008282916620373726, + 0.014792883768677711, + 0.013688717037439346, + 0.019504927098751068, + 0.01702875830233097, + 0.016336796805262566, + 0.01488780602812767, + 0.017502259463071823, + 0.019656777381896973, + 0.020000245422124863, + 0.026497725397348404, + 0.014273515902459621, + 0.03748390078544617, + 0.025402691215276718, + 0.02589731104671955, + 0.017473392188549042, + 0.02863779291510582, + 0.004792875610291958, + 0.004197912756353617, + 0.0054244669154286385, + 0.0029390077106654644, + 0.005133002996444702, + 0.005848919507116079, + 0.0035225101746618748, + 0.003992946352809668, + 0.03009568713605404, + 0.016994334757328033, + 0.01668613590300083, + 0.01671203039586544, + 0.020145561546087265, + 0.016215307638049126, + 0.0207939725369215, + 0.03479713574051857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.01360242534428835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076459851115942 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.0076459851115942, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076459851115942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10272636264562607, + 0.08047613501548767, + 0.10177382826805115, + 0.03707488626241684, + 0.059683170169591904, + 0.06517255306243896, + 0.1335403323173523, + 0.1687488853931427, + 0.061592862010002136, + 0.18921099603176117 + ], + "credit_map": [ + 0.011296811513602734, + 0.011953388340771198, + 0.012919108383357525, + 0.02172233723104, + 0.009640702977776527, + 0.014141908846795559, + 0.0089667784050107, + 0.012085331603884697, + 0.008699977770447731, + 0.008118286728858948, + 0.00807909481227398, + 0.014655714854598045, + 0.0065393634140491486, + 0.01567157730460167, + 0.006959471385926008, + 0.011752650141716003, + 0.010851174592971802, + 0.0131610706448555, + 0.009346677921712399, + 0.01451614685356617, + 0.010932480916380882, + 0.01654873974621296, + 0.009407170116901398, + 0.017010366544127464, + 0.004745787475258112, + 0.005299689248204231, + 0.004002934787422419, + 0.004404471255838871, + 0.0038020338397473097, + 0.005017934367060661, + 0.004423476755619049, + 0.005378559231758118, + 0.007866248488426208, + 0.013584921136498451, + 0.005781327839940786, + 0.006248618010431528, + 0.006137497257441282, + 0.006953698582947254, + 0.004849783610552549, + 0.008261073380708694, + 0.007196796126663685, + 0.008854534476995468, + 0.012109960429370403, + 0.0072169313207268715, + 0.006622711196541786, + 0.009910983964800835, + 0.00512916361913085, + 0.008131473325192928, + 0.015541746281087399, + 0.010350031778216362, + 0.0170888714492321, + 0.015010496601462364, + 0.017863014712929726, + 0.017217399552464485, + 0.014948665164411068, + 0.02552010864019394, + 0.016707707196474075, + 0.02198447659611702, + 0.012707913294434547, + 0.03169393166899681, + 0.023116040974855423, + 0.021679185330867767, + 0.019355829805135727, + 0.021503804251551628, + 0.009355246089398861, + 0.008646841160953045, + 0.00931521411985159, + 0.005304235033690929, + 0.006945459637790918, + 0.010970617644488811, + 0.00564080523326993, + 0.005414443090558052, + 0.028735514730215073, + 0.019405823200941086, + 0.05537441372871399, + 0.01807478815317154, + 0.014479707926511765, + 0.016394274309277534, + 0.01498800702393055, + 0.021758457645773888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.0076459851115942 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632971741259098 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007632971741259098, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632971741259098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1376982480287552, + 0.03403244912624359, + 0.1919555813074112, + 0.08713188022375107, + 0.09510362148284912, + 0.13022413849830627, + 0.0744154155254364, + 0.13222020864486694, + 0.06194053962826729, + 0.05527793988585472 + ], + "credit_map": [ + 0.015215370804071426, + 0.01515821274369955, + 0.016497323289513588, + 0.018015749752521515, + 0.01359360758215189, + 0.023201363161206245, + 0.014263874851167202, + 0.021752741187810898, + 0.003712864825502038, + 0.004365286324173212, + 0.0031836852431297302, + 0.007391107734292746, + 0.00282755377702415, + 0.00450345641002059, + 0.003049601335078478, + 0.004998891614377499, + 0.01877923123538494, + 0.022226963192224503, + 0.0212996955960989, + 0.022743333131074905, + 0.018925581127405167, + 0.037649769335985184, + 0.020906705409288406, + 0.029424304142594337, + 0.007991021499037743, + 0.016256552189588547, + 0.0069241295568645, + 0.009371213614940643, + 0.009952569380402565, + 0.016659557819366455, + 0.009134995751082897, + 0.010841836221516132, + 0.0105959577485919, + 0.01372692920267582, + 0.01204090565443039, + 0.015625402331352234, + 0.00953446514904499, + 0.013840976171195507, + 0.00837219413369894, + 0.011366788297891617, + 0.01329919882118702, + 0.014433472417294979, + 0.020089194178581238, + 0.013989717699587345, + 0.0128806596621871, + 0.026509733870625496, + 0.01215269323438406, + 0.01686946488916874, + 0.010070297867059708, + 0.005867864470928907, + 0.008596987463533878, + 0.010095611214637756, + 0.010636669583618641, + 0.008646262809634209, + 0.010019400157034397, + 0.010482324287295341, + 0.014822863973677158, + 0.01927897147834301, + 0.010820812545716763, + 0.021104339510202408, + 0.01839563623070717, + 0.017277570441365242, + 0.01696477085351944, + 0.013555241748690605, + 0.00862698070704937, + 0.007164000533521175, + 0.008915532380342484, + 0.006477201823145151, + 0.007908085361123085, + 0.010171252302825451, + 0.005038667004555464, + 0.007638821844011545, + 0.00990532897412777, + 0.006119624245911837, + 0.007639740128070116, + 0.006884563714265823, + 0.007486462593078613, + 0.005802069325000048, + 0.00598346721380949, + 0.005456683691591024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.007632971741259098 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017156180692836642 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.0017156180692836642, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017156180692836642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06682540476322174, + 0.0450945645570755, + 0.1896115392446518, + 0.01724187284708023, + 0.07560258358716965, + 0.13134893774986267, + 0.10558061301708221, + 0.2184792160987854, + 0.07701997458934784, + 0.07319530844688416 + ], + "credit_map": [ + 0.009711408987641335, + 0.0053926510736346245, + 0.008728126995265484, + 0.008758406154811382, + 0.007447928190231323, + 0.007644400000572205, + 0.007486815098673105, + 0.011655661277472973, + 0.004159842152148485, + 0.004111232236027718, + 0.008004271425306797, + 0.005274196155369282, + 0.005254795774817467, + 0.007192065939307213, + 0.004781665280461311, + 0.006316495593637228, + 0.018540676683187485, + 0.022351616993546486, + 0.020816760137677193, + 0.023375948891043663, + 0.020000772550702095, + 0.0350753590464592, + 0.021766481921076775, + 0.0276839267462492, + 0.0026893180329352617, + 0.001649190322495997, + 0.002203575801104307, + 0.0021695950999855995, + 0.0019109302666038275, + 0.001741610118187964, + 0.0017096608644351363, + 0.0031679936219006777, + 0.005191884934902191, + 0.01744355820119381, + 0.006457111332565546, + 0.01057666540145874, + 0.005769791081547737, + 0.015318939462304115, + 0.006243547424674034, + 0.008601081557571888, + 0.007946671918034554, + 0.04296572506427765, + 0.005892226472496986, + 0.011655653826892376, + 0.007187509909272194, + 0.04293452948331833, + 0.005101810209453106, + 0.007664806209504604, + 0.01352866180241108, + 0.008269095793366432, + 0.011133071966469288, + 0.014163683168590069, + 0.01194009743630886, + 0.013049842789769173, + 0.013145669363439083, + 0.020350487902760506, + 0.02095509134232998, + 0.03284153342247009, + 0.01878133788704872, + 0.0495452918112278, + 0.024877022951841354, + 0.02697090059518814, + 0.019790595397353172, + 0.02471744269132614, + 0.010704558342695236, + 0.00963891763240099, + 0.009908687323331833, + 0.006108742672950029, + 0.012666520662605762, + 0.012811047025024891, + 0.007822892628610134, + 0.007358603645116091, + 0.011995552107691765, + 0.007372521795332432, + 0.009232506155967712, + 0.007325885817408562, + 0.008039970882236958, + 0.006215720437467098, + 0.009412730112671852, + 0.013600416481494904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.0017156180692836642 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870837688446045 + ], + "opposite_logprob_tokens": [ + -1.1370837688446045 + ], + "logprob_reward": -0.3870837688446045, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870837688446045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13318094611167908, + 0.028380827978253365, + 0.04413510859012604, + 0.06779699772596359, + 0.05853583663702011, + 0.10227660089731216, + 0.319269061088562, + 0.08950857818126678, + 0.06680034846067429, + 0.09011569619178772 + ], + "credit_map": [ + 0.015858937054872513, + 0.01342601515352726, + 0.017398351803421974, + 0.02427585981786251, + 0.01148578617721796, + 0.016127891838550568, + 0.012732310220599174, + 0.02187580056488514, + 0.004782531410455704, + 0.0028531416319310665, + 0.0028807686176151037, + 0.003989795222878456, + 0.0025246767327189445, + 0.004462548065930605, + 0.002515839645639062, + 0.004371528513729572, + 0.004243350587785244, + 0.004852751269936562, + 0.005947299767285585, + 0.006532938219606876, + 0.004038261249661446, + 0.005488788243383169, + 0.005110620055347681, + 0.007921104319393635, + 0.007018475793302059, + 0.012088198214769363, + 0.006704321131110191, + 0.007756841368973255, + 0.006443969905376434, + 0.010884860530495644, + 0.007280638441443443, + 0.0096196960657835, + 0.005385300610214472, + 0.012451084330677986, + 0.007095093373209238, + 0.009584616869688034, + 0.00498014222830534, + 0.008062237873673439, + 0.004530700389295816, + 0.006446667946875095, + 0.010491813533008099, + 0.008826235309243202, + 0.024410992860794067, + 0.012873271480202675, + 0.008091655559837818, + 0.014201180078089237, + 0.009956496767699718, + 0.013424968346953392, + 0.04581441730260849, + 0.02773454785346985, + 0.030766138806939125, + 0.0389561764895916, + 0.04022045433521271, + 0.03699488937854767, + 0.040859345346689224, + 0.05792310833930969, + 0.010515048168599606, + 0.015547489747405052, + 0.008074161596596241, + 0.015603571198880672, + 0.008314702659845352, + 0.01230119913816452, + 0.008497041650116444, + 0.0106553565710783, + 0.009371291846036911, + 0.009416486136615276, + 0.01187821477651596, + 0.00620671920478344, + 0.00988724734634161, + 0.007837035693228245, + 0.006037408020347357, + 0.006165945902466774, + 0.01370164193212986, + 0.007904290221631527, + 0.008958986960351467, + 0.007151938509196043, + 0.00814883504062891, + 0.01023764256387949, + 0.011530554853379726, + 0.02248181775212288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.3870837688446045 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up, captured from a horizontal angle, plunges us into a microscopic world dominated by intense black and white. A massive praying mantis lurks on a broad leaf, its angular body bathed in a pure, luminous white, half shrouded in a bottomless, inky shadow, illuminated by the harsh overhead light. Its triangular head shifts slightly, its enormous compound eyes reflecting the light with an inorganic, obsidian sheen. Its body remains motionless, like a meticulously crafted sculpture of death. Not far below, an unsuspecting fly rests on the leaf's edge, its fine hairs clearly visible in the strong light as it leisurely cleans its forelegs, oblivious to its impending doom. Suddenly, the mantis's folded, spiky forelimbs, like a wound-up trap, flung out with lightning speed. The movement was swift and decisive, leaving a blurry white afterimage in the air. It instantly bridged the gap between the two and opened with unerring precision, seizing the fragile body of the fly. The entire image, rendered in the texture of a high-contrast black-and-white film, the interplay of light and shadow imbues this microscopic struggle for survival with immense drama and visual impact, creating an atmosphere of intense concentration and primal instinct.", + "seed": 10701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_1/index_8819_seed_10701_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": -0.04825291221495718, + "vlm_logprob": -0.04825291221495718, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.04825291221495718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022021645680069923 + ], + "opposite_logprob_tokens": [ + -5.147021770477295 + ], + "logprob_reward": -0.022021645680069923, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.022021645680069923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18613576889038086, + 0.15985366702079773, + 0.15357297658920288, + 0.0932469591498375, + 0.05804228410124779, + 0.055588990449905396, + 0.054108187556266785, + 0.03878531977534294, + 0.06873563677072525, + 0.13193020224571228 + ], + "credit_map": [ + 0.03415562957525253, + 0.020281100645661354, + 0.025966331362724304, + 0.019742965698242188, + 0.018152404576539993, + 0.02095731534063816, + 0.025420106947422028, + 0.02145990915596485, + 0.02700166217982769, + 0.02105576917529106, + 0.01943475753068924, + 0.018130576238036156, + 0.020194921642541885, + 0.015973757952451706, + 0.02306254394352436, + 0.014999658800661564, + 0.024699781090021133, + 0.01777052693068981, + 0.018765676766633987, + 0.019953522831201553, + 0.01828509010374546, + 0.01530234981328249, + 0.02342129684984684, + 0.015374721959233284, + 0.014699295163154602, + 0.012976938858628273, + 0.011690930463373661, + 0.010310519486665726, + 0.011612871661782265, + 0.009163678623735905, + 0.01280955970287323, + 0.009983154945075512, + 0.009367517195641994, + 0.006885484792292118, + 0.008719216100871563, + 0.006300916895270348, + 0.008185802027583122, + 0.005857046227902174, + 0.00626587588340044, + 0.0064604198560118675, + 0.0073023526929318905, + 0.00923925545066595, + 0.0072429366409778595, + 0.006616923026740551, + 0.006874666083604097, + 0.00566214369609952, + 0.0064310855232179165, + 0.00621962221339345, + 0.004725240170955658, + 0.008163285441696644, + 0.006470559164881706, + 0.005771953612565994, + 0.0064310976304113865, + 0.006003613583743572, + 0.007357263471931219, + 0.009185168892145157, + 0.0031022222246974707, + 0.007235888857394457, + 0.00479903956875205, + 0.00410916842520237, + 0.005237235687673092, + 0.0036010039038956165, + 0.005190998315811157, + 0.005509757902473211, + 0.007360938470810652, + 0.009889780543744564, + 0.00447176955640316, + 0.007430306635797024, + 0.009502516128122807, + 0.006130824331194162, + 0.014008263126015663, + 0.00994123425334692, + 0.021396251395344734, + 0.02473538927733898, + 0.010886707343161106, + 0.016499772667884827, + 0.011319278739392757, + 0.014294812455773354, + 0.01934688724577427, + 0.013451085425913334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.022021645680069923 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890708714723587 + ], + "opposite_logprob_tokens": [ + -2.4640707969665527 + ], + "logprob_reward": -0.0890708714723587, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890708714723587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45079126954078674, + 0.11984819173812866, + 0.08118252456188202, + 0.056325431913137436, + 0.03262941539287567, + 0.06911963224411011, + 0.03128696233034134, + 0.03781822696328163, + 0.018627813085913658, + 0.1023705005645752 + ], + "credit_map": [ + 0.07044728100299835, + 0.051491495221853256, + 0.05524851381778717, + 0.05163780599832535, + 0.04993651062250137, + 0.04740678891539574, + 0.07706405967473984, + 0.04755883663892746, + 0.013803957961499691, + 0.019037077203392982, + 0.012956274673342705, + 0.012238147668540478, + 0.01690233312547207, + 0.011411676183342934, + 0.020609036087989807, + 0.012889685109257698, + 0.013338610529899597, + 0.010012837126851082, + 0.009922533296048641, + 0.01020839624106884, + 0.008823615498840809, + 0.00810379721224308, + 0.01159414742141962, + 0.009178584441542625, + 0.007183958776295185, + 0.00819780956953764, + 0.005922019015997648, + 0.006991254631429911, + 0.007555206771939993, + 0.005515165161341429, + 0.008119556121528149, + 0.006840463262051344, + 0.0033804154954850674, + 0.003930939827114344, + 0.0048482418060302734, + 0.005492046475410461, + 0.0039042686112225056, + 0.0033595028799027205, + 0.0036972095258533955, + 0.004016790073364973, + 0.011346670798957348, + 0.009540222585201263, + 0.008418958634138107, + 0.00793071836233139, + 0.007312508299946785, + 0.007085834164172411, + 0.010068015195429325, + 0.00741670373827219, + 0.003025348763912916, + 0.0033204162027686834, + 0.003933778498321772, + 0.004906781017780304, + 0.0037362503353506327, + 0.0034960045013576746, + 0.0035485068801790476, + 0.005319875665009022, + 0.003963015973567963, + 0.008280329406261444, + 0.0032674327958375216, + 0.0038188721518963575, + 0.007809692528098822, + 0.0030931588262319565, + 0.004616193939000368, + 0.002969532273709774, + 0.0014370151329785585, + 0.0029974125791341066, + 0.0012773449998348951, + 0.0019138422794640064, + 0.0017479361267760396, + 0.0022211011964827776, + 0.003228466957807541, + 0.003804695326834917, + 0.010518557392060757, + 0.014039617031812668, + 0.013150511309504509, + 0.01773373782634735, + 0.012827597558498383, + 0.009632292203605175, + 0.012917803600430489, + 0.01155038457363844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0890708714723587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019450333202257752 + ], + "opposite_logprob_tokens": [ + -6.2519450187683105 + ], + "logprob_reward": -0.0019450333202257752, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019450333202257752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11585148423910141, + 0.049341052770614624, + 0.028787778690457344, + 0.044820625334978104, + 0.05264467000961304, + 0.07161607593297958, + 0.040915634483098984, + 0.079954594373703, + 0.26529285311698914, + 0.2507753074169159 + ], + "credit_map": [ + 0.017415283247828484, + 0.014714783988893032, + 0.012081318534910679, + 0.014246274717152119, + 0.010629862546920776, + 0.012077792547643185, + 0.02079705521464348, + 0.013889100402593613, + 0.007445965893566608, + 0.006298167631030083, + 0.005167023744434118, + 0.007486289367079735, + 0.005746607203036547, + 0.00506687443703413, + 0.006404552608728409, + 0.005725574214011431, + 0.002873533871024847, + 0.005309646483510733, + 0.002634287578985095, + 0.0037398217245936394, + 0.004308765288442373, + 0.0027562719769775867, + 0.004147606901824474, + 0.0030178430024534464, + 0.0044679963029921055, + 0.007235312834382057, + 0.005097576417028904, + 0.005578306969255209, + 0.005754471756517887, + 0.0043766382150352, + 0.007097831927239895, + 0.005212489049881697, + 0.006058693863451481, + 0.010508034378290176, + 0.0050451308488845825, + 0.005507894791662693, + 0.006743597332388163, + 0.004865773022174835, + 0.008304078131914139, + 0.005611465312540531, + 0.008473057299852371, + 0.01074823085218668, + 0.00834814365953207, + 0.009184351190924644, + 0.009182445704936981, + 0.007492886856198311, + 0.008315496146678925, + 0.009871457703411579, + 0.005682960618287325, + 0.004380078054964542, + 0.005331556778401136, + 0.004834353923797607, + 0.004015441983938217, + 0.00461036991328001, + 0.006440318189561367, + 0.0056205508299171925, + 0.009816168807446957, + 0.013698635622859001, + 0.008191795088350773, + 0.008347170427441597, + 0.012392127886414528, + 0.0072569637559354305, + 0.012999267317354679, + 0.007252457085996866, + 0.020567648112773895, + 0.059835657477378845, + 0.025715308263897896, + 0.029452253133058548, + 0.04420307278633118, + 0.026201803237199783, + 0.03772654011845589, + 0.021590542048215866, + 0.03627008944749832, + 0.04569287970662117, + 0.02029699832201004, + 0.035070836544036865, + 0.028751494362950325, + 0.028831636533141136, + 0.03305330500006676, + 0.022808043286204338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0019450333202257752 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011830481234937906 + ], + "opposite_logprob_tokens": [ + -6.751183032989502 + ], + "logprob_reward": -0.0011830481234937906, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011830481234937906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04204534739255905, + 0.05873947963118553, + 0.05479496717453003, + 0.016926642507314682, + 0.03869107365608215, + 0.04813347011804581, + 0.06237361580133438, + 0.07313539832830429, + 0.33967649936676025, + 0.2654834985733032 + ], + "credit_map": [ + 0.005592518951743841, + 0.005744922906160355, + 0.00478165689855814, + 0.0044740187004208565, + 0.004237127024680376, + 0.00471663661301136, + 0.006357343401759863, + 0.006141121033579111, + 0.008606174029409885, + 0.00861087255179882, + 0.00688310619443655, + 0.006938947830349207, + 0.006762349512428045, + 0.005821105558425188, + 0.008587325923144817, + 0.006529600825160742, + 0.008373602293431759, + 0.0064825559966266155, + 0.005954963620752096, + 0.007517828606069088, + 0.006098009645938873, + 0.00556303421035409, + 0.00864036101847887, + 0.0061646136455237865, + 0.0020404362585395575, + 0.002488445257768035, + 0.0019095649477094412, + 0.002292770892381668, + 0.0026740352623164654, + 0.0016950435237959027, + 0.002156956121325493, + 0.001669392455369234, + 0.005053156055510044, + 0.0064318859949707985, + 0.003905801335349679, + 0.0040952665731310844, + 0.005025241989642382, + 0.003692256985232234, + 0.005852416157722473, + 0.004635049495846033, + 0.005507182329893112, + 0.007205071859061718, + 0.005791834555566311, + 0.006252423394471407, + 0.006039545405656099, + 0.004951950162649155, + 0.006099993828684092, + 0.006285469047725201, + 0.010203738696873188, + 0.00949934683740139, + 0.006633067969232798, + 0.006766562350094318, + 0.007065249606966972, + 0.0064145480282604694, + 0.008988205343484879, + 0.006802898366004229, + 0.009250358678400517, + 0.012842684984207153, + 0.0069733355194330215, + 0.007791618350893259, + 0.011211042292416096, + 0.006869700737297535, + 0.011977762915194035, + 0.0062188925221562386, + 0.02706141583621502, + 0.07335282117128372, + 0.0372217521071434, + 0.033682260662317276, + 0.05762583762407303, + 0.032823916524648666, + 0.05010061711072922, + 0.027807893231511116, + 0.042096272110939026, + 0.040988482534885406, + 0.02599618211388588, + 0.04226919636130333, + 0.0280238576233387, + 0.029077405110001564, + 0.03164360672235489, + 0.025388525798916817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0011830481234937906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene an extreme macro close-up of a praying mantis and a fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style high-contrast black and white with intense shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the mantis initially motionless on a broad leaf while a fly cleans its forelegs nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mantis suddenly strike with lightning speed to seize the fly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle horizontal?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270439624786377 + ], + "opposite_logprob_tokens": [ + -2.1270439624786377 + ], + "logprob_reward": -0.1270439624786377, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270439624786377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16664789617061615, + 0.13811084628105164, + 0.06985485553741455, + 0.11795838177204132, + 0.06957492977380753, + 0.10084328800439835, + 0.10309212654829025, + 0.024233244359493256, + 0.03890475258231163, + 0.17077969014644623 + ], + "credit_map": [ + 0.021795129403471947, + 0.022234661504626274, + 0.020198168233036995, + 0.019437074661254883, + 0.018105482682585716, + 0.018178043887019157, + 0.025838619098067284, + 0.020860716700553894, + 0.02080167829990387, + 0.014942210167646408, + 0.016581332311034203, + 0.01821596547961235, + 0.016228873282670975, + 0.0142463194206357, + 0.02139602228999138, + 0.015698451548814774, + 0.010184607468545437, + 0.007237397599965334, + 0.00800367258489132, + 0.010994755662977695, + 0.008692548610270023, + 0.006920614745467901, + 0.01073948759585619, + 0.007081769406795502, + 0.017668861895799637, + 0.017388613894581795, + 0.01337179820984602, + 0.012327450327575207, + 0.013553446158766747, + 0.01130419597029686, + 0.017058223485946655, + 0.015285799279808998, + 0.008638943545520306, + 0.00928353052586317, + 0.008365265093743801, + 0.006870044860988855, + 0.00874574389308691, + 0.006355843972414732, + 0.012486225925385952, + 0.008829331956803799, + 0.0171199943870306, + 0.012968864291906357, + 0.013520126231014729, + 0.011038427241146564, + 0.009896447882056236, + 0.009896609000861645, + 0.014846529811620712, + 0.011556286364793777, + 0.015803411602973938, + 0.01137516275048256, + 0.016925254836678505, + 0.010488100349903107, + 0.01133484486490488, + 0.009853006340563297, + 0.014579473063349724, + 0.01273287646472454, + 0.0035140083637088537, + 0.0030446455348283052, + 0.0032670393120497465, + 0.0025437602307647467, + 0.0028757110703736544, + 0.002410810673609376, + 0.0032237041741609573, + 0.003353564301505685, + 0.005368917249143124, + 0.004647323861718178, + 0.0033853882923722267, + 0.00412145396694541, + 0.006011880002915859, + 0.0036596625577658415, + 0.006248535122722387, + 0.005461593624204397, + 0.023024756461381912, + 0.024192027747631073, + 0.023137275129556656, + 0.021298848092556, + 0.015990808606147766, + 0.01766694150865078, + 0.01706051453948021, + 0.02840852178633213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.1270439624786377 + } + ], + "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/0000058/group_10/index_tensor([7859, 7859])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_tensor([7859, 7859])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cbad9638dffc177d0ad5f25a37360c634d530d4f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_tensor([7859, 7859])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "motionless, fixed long shot captures a mirror-like lake, perfectly calm at dusk. The distant mountains merge into deep blue silhouettes against the horizon. In the very center of the lake, the air begins to twist and flicker violently, like a heat wave, as the silhouette of a massive ghostly ship slowly crystallizes from the void. This translucent vessel, its hull seemingly composed of solidified light, allows for a glimpse into its interior. Its hull is crisscrossed with an incredibly complex and dense mechanical structure inspired by classic Akira animations—countless pipes of varying sizes, exposed cable bundles, and layers of hydraulics and fins intertwined, creating a grotesque yet sophisticated industrial aesthetic. As it slowly materialized, its flickering outline stretched and deformed in the air, emitting a cold, faint blue glow. The light penetrated the hull and cast an equally distorted and flickering, yet unusually clear reflection on the calm lake. The entire grand and tranquil scene was shrouded in an awe-inspiring sense of mystery, as if a behemoth from another dimension was steadily and calmly descending into this world.", + "seed": 9772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_7859_seed_9772_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": -0.01055880666244775, + "vlm_logprob": -0.01055880666244775, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.01055880666244775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.025836575776338577 + ], + "opposite_logprob_tokens": [ + -3.90083646774292 + ], + "logprob_reward": -0.025836575776338577, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025836575776338577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38373836874961853, + 0.040828123688697815, + 0.04450656846165657, + 0.04449102655053139, + 0.0491144061088562, + 0.053457990288734436, + 0.039205025881528854, + 0.0838107168674469, + 0.10216701030731201, + 0.1586807817220688 + ], + "credit_map": [ + 0.036329738795757294, + 0.08294743299484253, + 0.034206800162792206, + 0.040742114186286926, + 0.06308025866746902, + 0.04507600888609886, + 0.037408359348773956, + 0.04394766315817833, + 0.0038718548603355885, + 0.005596840754151344, + 0.0030946373008191586, + 0.005663271527737379, + 0.009653650224208832, + 0.00489531084895134, + 0.004439899232238531, + 0.003612656146287918, + 0.0026494842022657394, + 0.005699077155441046, + 0.0014893574407324195, + 0.006290385965257883, + 0.011522217653691769, + 0.007504674606025219, + 0.006978256162256002, + 0.002373115625232458, + 0.003145457012578845, + 0.006270366720855236, + 0.001930972095578909, + 0.007678696420043707, + 0.007740024011582136, + 0.006065453868359327, + 0.008811410516500473, + 0.002848645206540823, + 0.0041973041370511055, + 0.007950868457555771, + 0.0037506294902414083, + 0.006729028653353453, + 0.0074670324102044106, + 0.006704746745526791, + 0.007444709073752165, + 0.004870086908340454, + 0.00439252145588398, + 0.006516107823699713, + 0.003698893589898944, + 0.008741755969822407, + 0.009349052794277668, + 0.007614586036652327, + 0.006877662148326635, + 0.006267409306019545, + 0.004953875672072172, + 0.005975424777716398, + 0.002606999594718218, + 0.006502150092273951, + 0.004771423060446978, + 0.005029173102229834, + 0.0057692937552928925, + 0.00359668442979455, + 0.006424728315323591, + 0.01869923062622547, + 0.007431245408952236, + 0.01055422704666853, + 0.012629588134586811, + 0.011677402071654797, + 0.010918800719082355, + 0.005475485697388649, + 0.0121912257745862, + 0.014089340344071388, + 0.008400938473641872, + 0.012082582339644432, + 0.011729435995221138, + 0.014515126124024391, + 0.015166422352194786, + 0.013991942629218102, + 0.02151169627904892, + 0.018780682235956192, + 0.017418643459677696, + 0.015948839485645294, + 0.016902152448892593, + 0.021427448838949203, + 0.02048690989613533, + 0.026204418390989304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.025836575776338577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.020594101399183273 + ], + "opposite_logprob_tokens": [ + -3.8955941200256348 + ], + "logprob_reward": -0.020594101399183273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594101399183273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27460071444511414, + 0.03204184025526047, + 0.0736798644065857, + 0.04089485481381416, + 0.027779335156083107, + 0.04786938056349754, + 0.0370514802634716, + 0.018361175432801247, + 0.08751043677330017, + 0.36021098494529724 + ], + "credit_map": [ + 0.02457721158862114, + 0.03340626880526543, + 0.03695034608244896, + 0.04243830218911171, + 0.03235245868563652, + 0.025611869990825653, + 0.030423026531934738, + 0.04884118214249611, + 0.0021005249582231045, + 0.006421777885407209, + 0.001943419803865254, + 0.0052143423818051815, + 0.007108645513653755, + 0.003122430294752121, + 0.003692281898111105, + 0.0024384167045354843, + 0.004973101429641247, + 0.011720013804733753, + 0.0015366862062364817, + 0.006221545394510031, + 0.02483914978802204, + 0.01642361842095852, + 0.005637910682708025, + 0.0023278279695659876, + 0.0029623156879097223, + 0.004010783042758703, + 0.003946478478610516, + 0.007638279814273119, + 0.005222815554589033, + 0.004483629483729601, + 0.007610498461872339, + 0.005020051263272762, + 0.0027520516887307167, + 0.0028611202724277973, + 0.0022706782910972834, + 0.003421858185902238, + 0.0038431487046182156, + 0.005390367936342955, + 0.0032509060110896826, + 0.003989201504737139, + 0.004895109683275223, + 0.006410367786884308, + 0.003883389988914132, + 0.005681653972715139, + 0.00844889972358942, + 0.007278294302523136, + 0.0048142834566533566, + 0.006457378622144461, + 0.005194356665015221, + 0.00423820223659277, + 0.003798237070441246, + 0.004439161159098148, + 0.004161552991718054, + 0.004491198342293501, + 0.004063495900481939, + 0.006665273569524288, + 0.0017971991328522563, + 0.003730772528797388, + 0.0014987331815063953, + 0.0023874978069216013, + 0.002711457433179021, + 0.0025712864007800817, + 0.0023385046515613794, + 0.0013257210375741124, + 0.012849216349422932, + 0.009748579934239388, + 0.009328505024313927, + 0.007938980124890804, + 0.01091124676167965, + 0.011152994818985462, + 0.011193222366273403, + 0.014387687668204308, + 0.05622316524386406, + 0.03724125027656555, + 0.04656219854950905, + 0.030987832695245743, + 0.03248073160648346, + 0.04927309602499008, + 0.03246233984827995, + 0.07498031854629517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.020594101399183273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.0013404440833255649 + ], + "opposite_logprob_tokens": [ + -6.626340389251709 + ], + "logprob_reward": -0.0013404440833255649, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013404440833255649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1546887457370758, + 0.07431007921695709, + 0.2600092589855194, + 0.08648350834846497, + 0.05102934688329697, + 0.11715786159038544, + 0.04902908205986023, + 0.04822298511862755, + 0.06430120021104813, + 0.09476795047521591 + ], + "credit_map": [ + 0.013881094753742218, + 0.029977811500430107, + 0.00941119622439146, + 0.024701735004782677, + 0.02617962658405304, + 0.021423902362585068, + 0.015501819550991058, + 0.013611541129648685, + 0.004318417981266975, + 0.012573489919304848, + 0.004057052079588175, + 0.014349563047289848, + 0.009491849690675735, + 0.007847469300031662, + 0.016495520249009132, + 0.005176709033548832, + 0.014079131186008453, + 0.028978334739804268, + 0.0070432983338832855, + 0.018121736124157906, + 0.08910223096609116, + 0.0711417943239212, + 0.02228391170501709, + 0.00925879180431366, + 0.008737782016396523, + 0.010089527815580368, + 0.00876331701874733, + 0.01714303530752659, + 0.011148970574140549, + 0.010305491276085377, + 0.011790843680500984, + 0.008504534140229225, + 0.003870234824717045, + 0.005194654688239098, + 0.002266752766445279, + 0.009315971285104752, + 0.009102949872612953, + 0.008971752598881721, + 0.009625442326068878, + 0.002681580837816, + 0.007481045089662075, + 0.014604154974222183, + 0.0065321545116603374, + 0.02077341638505459, + 0.02512608841061592, + 0.01907378062605858, + 0.015572682954370975, + 0.007994532585144043, + 0.0029335638973861933, + 0.006268611643463373, + 0.0021476668771356344, + 0.00915514212101698, + 0.007441686000674963, + 0.007154783699661493, + 0.011344343423843384, + 0.0025832816027104855, + 0.004451632034033537, + 0.006264482159167528, + 0.0035590198822319508, + 0.005152468103915453, + 0.00620263209566474, + 0.006103907246142626, + 0.008328389376401901, + 0.008160448633134365, + 0.005742385983467102, + 0.011061906814575195, + 0.004754133988171816, + 0.00782006699591875, + 0.0094669284299016, + 0.0076603880152106285, + 0.009506269358098507, + 0.008289117366075516, + 0.007413245737552643, + 0.015401122160255909, + 0.005032717250287533, + 0.013143940828740597, + 0.015376029536128044, + 0.01156806480139494, + 0.020379777997732162, + 0.006453048903495073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0013404440833255649 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.0022023487836122513 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022023487836122513, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022023487836122513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16344864666461945, + 0.026608653366565704, + 0.22699134051799774, + 0.12252248823642731, + 0.0551404245197773, + 0.08704019337892532, + 0.026212042197585106, + 0.08531630784273148, + 0.06810760498046875, + 0.13861232995986938 + ], + "credit_map": [ + 0.012574801221489906, + 0.024406176060438156, + 0.011327375657856464, + 0.020502695813775063, + 0.04149219021201134, + 0.0240030474960804, + 0.015064130537211895, + 0.014078219421207905, + 0.0013421312905848026, + 0.005084686446934938, + 0.0015291738091036677, + 0.003491903655230999, + 0.00404516514390707, + 0.004785536788403988, + 0.0045372918248176575, + 0.0017927647568285465, + 0.007399712689220905, + 0.018645131960511208, + 0.002762463642284274, + 0.005596838425844908, + 0.10173693299293518, + 0.08044161647558212, + 0.006295288447290659, + 0.00411335239186883, + 0.009230491705238819, + 0.01810631714761257, + 0.006785588804632425, + 0.01579112745821476, + 0.02827080897986889, + 0.019666846841573715, + 0.017935553565621376, + 0.006735754199326038, + 0.004721393343061209, + 0.006674484815448523, + 0.004647077061235905, + 0.007548467721790075, + 0.009023950435221195, + 0.009967475198209286, + 0.00735783064737916, + 0.005199747625738382, + 0.006583269219845533, + 0.011981263756752014, + 0.005393047817051411, + 0.012697200290858746, + 0.016273818910121918, + 0.011595156043767929, + 0.014530574902892113, + 0.007985862903296947, + 0.0023495834320783615, + 0.0025338681880384684, + 0.0027571539394557476, + 0.004938158206641674, + 0.003193809185177088, + 0.0034788022749125957, + 0.004356898833066225, + 0.0026037669740617275, + 0.006494310684502125, + 0.01422231923788786, + 0.006591533310711384, + 0.012308996170759201, + 0.013245762325823307, + 0.010508359409868717, + 0.015171161852777004, + 0.00677386624738574, + 0.005052678287029266, + 0.009279882535338402, + 0.006543678231537342, + 0.008954820223152637, + 0.010947278700768948, + 0.00854714959859848, + 0.012559092603623867, + 0.006223026663064957, + 0.009575816802680492, + 0.020619431510567665, + 0.00763546722009778, + 0.01806619204580784, + 0.03169161081314087, + 0.017729198560118675, + 0.026200300082564354, + 0.007094296161085367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.0022023487836122513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.002820563269779086 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.002820563269779086, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820563269779086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3007504642009735, + 0.04310827702283859, + 0.19966775178909302, + 0.07814213633537292, + 0.02294519543647766, + 0.04536989703774452, + 0.028494592756032944, + 0.11074497550725937, + 0.0326821506023407, + 0.13809457421302795 + ], + "credit_map": [ + 0.024835363030433655, + 0.0665140450000763, + 0.02167876809835434, + 0.030469004064798355, + 0.06336268782615662, + 0.04428310692310333, + 0.022980131208896637, + 0.026627356186509132, + 0.0022138208150863647, + 0.006124670151621103, + 0.004301686771214008, + 0.006122251972556114, + 0.004858058877289295, + 0.00443246029317379, + 0.009644897654652596, + 0.005410430021584034, + 0.008662438951432705, + 0.019041189923882484, + 0.004187596030533314, + 0.010642691515386105, + 0.07819923758506775, + 0.06230528652667999, + 0.01211563590914011, + 0.00451368372887373, + 0.007762073073536158, + 0.008519889786839485, + 0.005760038271546364, + 0.009290640242397785, + 0.01654025912284851, + 0.015688832849264145, + 0.009408988058567047, + 0.005171417258679867, + 0.0018410640768706799, + 0.002862151712179184, + 0.001477475045248866, + 0.0030157954897731543, + 0.004637995734810829, + 0.003075164742767811, + 0.0036288199480623007, + 0.0024067284539341927, + 0.005314038600772619, + 0.005144950468093157, + 0.0045503186993300915, + 0.005053030326962471, + 0.00825347751379013, + 0.006100628525018692, + 0.005731094162911177, + 0.0052223606035113335, + 0.0022490567062050104, + 0.003915577195584774, + 0.001186059438623488, + 0.0049913120456039906, + 0.005154157057404518, + 0.005088416859507561, + 0.004284820519387722, + 0.0016251932829618454, + 0.008627279661595821, + 0.0194466020911932, + 0.01018471922725439, + 0.014226041734218597, + 0.016502995043992996, + 0.014159316197037697, + 0.018225518986582756, + 0.009372502565383911, + 0.002680594567209482, + 0.0056381928734481335, + 0.002386948559433222, + 0.004695389419794083, + 0.004531786311417818, + 0.005420972127467394, + 0.004142916761338711, + 0.0031853485852479935, + 0.017254238948225975, + 0.01653805933892727, + 0.016726091504096985, + 0.010691042058169842, + 0.02297757938504219, + 0.01733035035431385, + 0.014012381434440613, + 0.022564832121133804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.002820563269779086 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "motionless, fixed long shot captures a mirror-like lake, perfectly calm at dusk. The distant mountains merge into deep blue silhouettes against the horizon. In the very center of the lake, the air begins to twist and flicker violently, like a heat wave, as the silhouette of a massive ghostly ship slowly crystallizes from the void. This translucent vessel, its hull seemingly composed of solidified light, allows for a glimpse into its interior. Its hull is crisscrossed with an incredibly complex and dense mechanical structure inspired by classic Akira animations—countless pipes of varying sizes, exposed cable bundles, and layers of hydraulics and fins intertwined, creating a grotesque yet sophisticated industrial aesthetic. As it slowly materialized, its flickering outline stretched and deformed in the air, emitting a cold, faint blue glow. The light penetrated the hull and cast an equally distorted and flickering, yet unusually clear reflection on the calm lake. The entire grand and tranquil scene was shrouded in an awe-inspiring sense of mystery, as if a behemoth from another dimension was steadily and calmly descending into this world.", + "seed": 9773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_7859_seed_9773_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": -0.02059795695822686, + "vlm_logprob": -0.02059795695822686, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.02059795695822686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.08755872398614883 + ], + "opposite_logprob_tokens": [ + -2.5875587463378906 + ], + "logprob_reward": -0.08755872398614883, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08755872398614883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1673102080821991, + 0.09068804234266281, + 0.13649138808250427, + 0.029362449422478676, + 0.03849274292588234, + 0.07310248911380768, + 0.16330769658088684, + 0.2499610334634781, + 0.03837379068136215, + 0.012910126708447933 + ], + "credit_map": [ + 0.009219382889568806, + 0.03239326551556587, + 0.021672919392585754, + 0.017598912119865417, + 0.029417745769023895, + 0.013527573086321354, + 0.017466653138399124, + 0.026013772934675217, + 0.007164703216403723, + 0.011333838105201721, + 0.01387822162359953, + 0.01429529394954443, + 0.011557692661881447, + 0.009228051640093327, + 0.01116908248513937, + 0.012061164714396, + 0.003358828369528055, + 0.05886765941977501, + 0.0027736895717680454, + 0.003957866225391626, + 0.05402492359280586, + 0.004981609992682934, + 0.004784842487424612, + 0.0037419754080474377, + 0.001811145688407123, + 0.004601994063705206, + 0.0031619935762137175, + 0.004392002243548632, + 0.0046041193418204784, + 0.0031562475487589836, + 0.004346847999840975, + 0.0032881018705666065, + 0.0019883750937879086, + 0.007914024405181408, + 0.0036225977819412947, + 0.004393903538584709, + 0.007132751401513815, + 0.004026189912110567, + 0.005313768517225981, + 0.004101132974028587, + 0.0046164230443537235, + 0.019028840586543083, + 0.004858245607465506, + 0.006855615880340338, + 0.007673170883208513, + 0.005856673698872328, + 0.019389918074011803, + 0.00482360552996397, + 0.01170363649725914, + 0.013611453585326672, + 0.012724820524454117, + 0.02088591642677784, + 0.024824470281600952, + 0.015805242583155632, + 0.053825993090867996, + 0.009926166385412216, + 0.030721217393875122, + 0.01865543983876705, + 0.059901583939790726, + 0.02896326780319214, + 0.030602918937802315, + 0.02045898325741291, + 0.030084678903222084, + 0.03057294711470604, + 0.0031192379537969828, + 0.003971002995967865, + 0.004207202233374119, + 0.0065727438777685165, + 0.006620059255510569, + 0.0035684150643646717, + 0.004877956118434668, + 0.005437172949314117, + 0.001091649173758924, + 0.001383660244755447, + 0.0010422723134979606, + 0.0017523664282634854, + 0.002766386140137911, + 0.0011404816759750247, + 0.0017735771834850311, + 0.0019597343634814024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.08755872398614883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.0028484982904046774 + ], + "opposite_logprob_tokens": [ + -5.8778486251831055 + ], + "logprob_reward": -0.0028484982904046774, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028484982904046774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040788307785987854, + 0.07495519518852234, + 0.1471571922302246, + 0.015108867548406124, + 0.0191558338701725, + 0.4966728389263153, + 0.15599942207336426, + 0.030989937484264374, + 0.010794560424983501, + 0.008377824909985065 + ], + "credit_map": [ + 0.003857963252812624, + 0.007172944489866495, + 0.006834936328232288, + 0.004682505968958139, + 0.0049842484295368195, + 0.004490804858505726, + 0.003365178359672427, + 0.005399724002927542, + 0.0021524946205317974, + 0.029356520622968674, + 0.002723763696849346, + 0.0031737315002828836, + 0.027803385630249977, + 0.0033592504914849997, + 0.003786944318562746, + 0.002599098486825824, + 0.003979751840233803, + 0.06204435974359512, + 0.003802351653575897, + 0.004412201698869467, + 0.056199610233306885, + 0.007020444143563509, + 0.004312048200517893, + 0.00538642005994916, + 0.0011989990016445518, + 0.004092575516551733, + 0.0014846358681097627, + 0.0011647742940112948, + 0.0024353968910872936, + 0.0015853527002036572, + 0.0014822673983871937, + 0.0016648642485961318, + 0.0016958791529759765, + 0.0029417353216558695, + 0.0016927077667787671, + 0.0018734325421974063, + 0.003929219674319029, + 0.00267593702301383, + 0.002419332042336464, + 0.0019275863887742162, + 0.022749995812773705, + 0.18177665770053864, + 0.019713006913661957, + 0.029621809720993042, + 0.045834559947252274, + 0.022060440853238106, + 0.1535700559616089, + 0.02134629525244236, + 0.015508187003433704, + 0.012232508510351181, + 0.007746835704892874, + 0.014730643481016159, + 0.02293688803911209, + 0.012479783967137337, + 0.06158476322889328, + 0.008779800496995449, + 0.0038123158738017082, + 0.0018290207954123616, + 0.007362066302448511, + 0.002546570962294936, + 0.0036559375002980232, + 0.002209430793300271, + 0.005501278676092625, + 0.004073313903063536, + 0.0008822229574434459, + 0.001167035079561174, + 0.0010219061514362693, + 0.002360566286370158, + 0.0018876640824601054, + 0.0010031559504568577, + 0.0015048994682729244, + 0.000967109517659992, + 0.000624229374807328, + 0.0009717126959003508, + 0.0010253498330712318, + 0.001517617260105908, + 0.0012617146130651236, + 0.0006827773177064955, + 0.001007556333206594, + 0.0012868671910837293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.0028484982904046774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.0011922164121642709 + ], + "opposite_logprob_tokens": [ + -6.751192092895508 + ], + "logprob_reward": -0.0011922164121642709, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011922164121642709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08937858045101166, + 0.06847521662712097, + 0.065301813185215, + 0.05649818480014801, + 0.06400542706251144, + 0.39210042357444763, + 0.11169265955686569, + 0.06571361422538757, + 0.039426811039447784, + 0.04740724712610245 + ], + "credit_map": [ + 0.006287876050919294, + 0.012667357921600342, + 0.012980921193957329, + 0.007045376114547253, + 0.015737634152173996, + 0.008418655954301357, + 0.012423950247466564, + 0.01381680928170681, + 0.0049103256314992905, + 0.018512021750211716, + 0.007303126621991396, + 0.005815507378429174, + 0.011835540644824505, + 0.005386256147176027, + 0.007054814137518406, + 0.007657624781131744, + 0.0025668181478977203, + 0.020736465230584145, + 0.003084380179643631, + 0.003691896563395858, + 0.02301357127726078, + 0.004292368423193693, + 0.004608264658600092, + 0.0033080505672842264, + 0.004761555697768927, + 0.008168335072696209, + 0.008536571636795998, + 0.0067140175960958, + 0.008825967088341713, + 0.00577877601608634, + 0.006412562448531389, + 0.007300397381186485, + 0.004182333126664162, + 0.013163610361516476, + 0.005726957693696022, + 0.006642551627010107, + 0.012988213449716568, + 0.006960456725209951, + 0.008125375024974346, + 0.006215925328433514, + 0.01622336357831955, + 0.15051411092281342, + 0.014005713164806366, + 0.02053147181868553, + 0.03680847957730293, + 0.014838729053735733, + 0.1234726831316948, + 0.01570584438741207, + 0.012718729674816132, + 0.011811330914497375, + 0.013851918280124664, + 0.01748288981616497, + 0.015380045399069786, + 0.008746244944632053, + 0.02236037701368332, + 0.009341122582554817, + 0.005760239902883768, + 0.007988248951733112, + 0.0059020426124334335, + 0.010040219873189926, + 0.012169984169304371, + 0.008426477201282978, + 0.009527236223220825, + 0.005899166222661734, + 0.0034016952849924564, + 0.004240401554852724, + 0.003044252982363105, + 0.005195038393139839, + 0.01043013297021389, + 0.003920088987797499, + 0.004450546111911535, + 0.0047446563839912415, + 0.0034728404134511948, + 0.0038069700822234154, + 0.004324508365243673, + 0.006772079970687628, + 0.010010521858930588, + 0.0065625873394310474, + 0.005960477981716394, + 0.006497264839708805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.0011922164121642709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.004648118279874325 + ], + "opposite_logprob_tokens": [ + -5.379648208618164 + ], + "logprob_reward": -0.004648118279874325, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648118279874325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06663703918457031, + 0.011556464247405529, + 0.10119279474020004, + 0.019796201959252357, + 0.012024483643472195, + 0.5621638298034668, + 0.08155177533626556, + 0.11077883094549179, + 0.015268668532371521, + 0.019029946997761726 + ], + "credit_map": [ + 0.004614288453012705, + 0.014149865135550499, + 0.00851528998464346, + 0.007260807324200869, + 0.012117533013224602, + 0.005487495567649603, + 0.005163081921637058, + 0.009328678250312805, + 0.0008901481050997972, + 0.0018092490499839187, + 0.0015477316919714212, + 0.0014218442374840379, + 0.0020802824292331934, + 0.0012321661924943328, + 0.0012299888767302036, + 0.0013450534315779805, + 0.0023756108712404966, + 0.04170110076665878, + 0.001987296622246504, + 0.002725461032241583, + 0.04129602015018463, + 0.003928383346647024, + 0.0034464553464204073, + 0.003732468234375119, + 0.0013255091616883874, + 0.003355376422405243, + 0.0021569596137851477, + 0.0022460410837084055, + 0.0038053414318710566, + 0.002281379420310259, + 0.0023306207731366158, + 0.0022949748672544956, + 0.0008631125674583018, + 0.002298310399055481, + 0.0012344946153461933, + 0.0012172037968412042, + 0.0025164957623928785, + 0.0011665852507576346, + 0.001387649797834456, + 0.0013406318612396717, + 0.025564968585968018, + 0.20661622285842896, + 0.02283809520304203, + 0.032621756196022034, + 0.05333007499575615, + 0.02458527684211731, + 0.1726313978433609, + 0.023976003751158714, + 0.009277522563934326, + 0.010690866969525814, + 0.006537301000207663, + 0.010059922002255917, + 0.019095081835985184, + 0.009062113240361214, + 0.01084659993648529, + 0.005982371978461742, + 0.012836462818086147, + 0.00934706348925829, + 0.02458583004772663, + 0.016075072810053825, + 0.0180436410009861, + 0.009067223407328129, + 0.010588519275188446, + 0.010235017165541649, + 0.0012633318547159433, + 0.0018354147905483842, + 0.0013169004814699292, + 0.0028859225567430258, + 0.0021518447902053595, + 0.0013817149447277188, + 0.0029307850636541843, + 0.0015027538174763322, + 0.0017753943102434278, + 0.0020179457496851683, + 0.0018128582742065191, + 0.003916776739060879, + 0.003905867924913764, + 0.001382089452818036, + 0.0023219087161123753, + 0.0018971061799675226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.004648118279874325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.0067422278225421906 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.0067422278225421906, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067422278225421906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18358924984931946, + 0.08594531565904617, + 0.2262735813856125, + 0.08938632160425186, + 0.03301236033439636, + 0.06038081645965576, + 0.17286206781864166, + 0.09208039194345474, + 0.0238114632666111, + 0.032658450305461884 + ], + "credit_map": [ + 0.013499777764081955, + 0.03131496533751488, + 0.024949194863438606, + 0.023376109078526497, + 0.02803877741098404, + 0.01342405378818512, + 0.01955145038664341, + 0.02943491004407406, + 0.005585351958870888, + 0.019231677055358887, + 0.008318286389112473, + 0.011958418413996696, + 0.014227480627596378, + 0.007394823711365461, + 0.01156605500727892, + 0.007663218770176172, + 0.00831778347492218, + 0.10165149718523026, + 0.006122682709246874, + 0.007392283994704485, + 0.07982506603002548, + 0.009193047881126404, + 0.007051306311041117, + 0.006719897501170635, + 0.005987148266285658, + 0.0126062436029315, + 0.009542912244796753, + 0.012168033979833126, + 0.013615482486784458, + 0.008032635785639286, + 0.016503026708960533, + 0.01093082595616579, + 0.0029356111772358418, + 0.004581267945468426, + 0.003460401901975274, + 0.0033577056601643562, + 0.005818204954266548, + 0.0036075604148209095, + 0.005382321309298277, + 0.0038692839443683624, + 0.0030416296795010567, + 0.017231401056051254, + 0.004804505035281181, + 0.00493596913293004, + 0.006921625230461359, + 0.0041177161037921906, + 0.01584303192794323, + 0.0034849338699132204, + 0.018158933147788048, + 0.01381121575832367, + 0.009260362945497036, + 0.01651948131620884, + 0.025179201737046242, + 0.015690965577960014, + 0.06429342925548553, + 0.009948465041816235, + 0.006675661075860262, + 0.008110035210847855, + 0.011047760024666786, + 0.007569282781332731, + 0.011128212325274944, + 0.013115040957927704, + 0.01545772422105074, + 0.018976669758558273, + 0.0020769331604242325, + 0.002354492899030447, + 0.002613551216199994, + 0.004576259292662144, + 0.004285089671611786, + 0.0018924946198239923, + 0.0032497693318873644, + 0.0027628729585558176, + 0.0039684162475168705, + 0.003247709944844246, + 0.003639033529907465, + 0.005320374388247728, + 0.005996298510581255, + 0.0033846416044980288, + 0.003991532605141401, + 0.003110440680757165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0067422278225421906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "motionless, fixed long shot captures a mirror-like lake, perfectly calm at dusk. The distant mountains merge into deep blue silhouettes against the horizon. In the very center of the lake, the air begins to twist and flicker violently, like a heat wave, as the silhouette of a massive ghostly ship slowly crystallizes from the void. This translucent vessel, its hull seemingly composed of solidified light, allows for a glimpse into its interior. Its hull is crisscrossed with an incredibly complex and dense mechanical structure inspired by classic Akira animations—countless pipes of varying sizes, exposed cable bundles, and layers of hydraulics and fins intertwined, creating a grotesque yet sophisticated industrial aesthetic. As it slowly materialized, its flickering outline stretched and deformed in the air, emitting a cold, faint blue glow. The light penetrated the hull and cast an equally distorted and flickering, yet unusually clear reflection on the calm lake. The entire grand and tranquil scene was shrouded in an awe-inspiring sense of mystery, as if a behemoth from another dimension was steadily and calmly descending into this world.", + "seed": 9774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_7859_seed_9774_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": -0.009599547483958304, + "vlm_logprob": -0.009599547483958304, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.009599547483958304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.024928538128733635 + ], + "opposite_logprob_tokens": [ + -4.024928569793701 + ], + "logprob_reward": -0.024928538128733635, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024928538128733635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2545371949672699, + 0.08447689563035965, + 0.10655099153518677, + 0.14475010335445404, + 0.08974949270486832, + 0.05248004198074341, + 0.043403495103120804, + 0.04928094148635864, + 0.02724895253777504, + 0.1475219577550888 + ], + "credit_map": [ + 0.021439626812934875, + 0.03586484119296074, + 0.027402127161622047, + 0.03596369922161102, + 0.029285239055752754, + 0.025103745982050896, + 0.04391265660524368, + 0.03556524217128754, + 0.007153340615332127, + 0.020730363205075264, + 0.005512900184839964, + 0.010244165547192097, + 0.015001555904746056, + 0.009458165615797043, + 0.009312143549323082, + 0.007064265199005604, + 0.009521915577352047, + 0.040976282209157944, + 0.004219035152345896, + 0.007608886808156967, + 0.02733934484422207, + 0.005062964279204607, + 0.005820321384817362, + 0.006002241279929876, + 0.01204985473304987, + 0.03981447219848633, + 0.01054868008941412, + 0.012407384812831879, + 0.026161717250943184, + 0.015360800549387932, + 0.01204199530184269, + 0.016365190967917442, + 0.006574354600161314, + 0.02753806672990322, + 0.005150358658283949, + 0.010882463306188583, + 0.019556140527129173, + 0.0065016974695026875, + 0.008992942981421947, + 0.004553471226245165, + 0.00367626640945673, + 0.0056518083438277245, + 0.005881491582840681, + 0.006148935295641422, + 0.007697663735598326, + 0.004740456119179726, + 0.006029220763593912, + 0.012654202990233898, + 0.0054670157842338085, + 0.012674454599618912, + 0.003293125657364726, + 0.006503285374492407, + 0.006033849436789751, + 0.0020082942210137844, + 0.005582390818744898, + 0.001841078046709299, + 0.0028963065706193447, + 0.009179268963634968, + 0.003363402560353279, + 0.010734899900853634, + 0.006037483923137188, + 0.0031937528401613235, + 0.009960721246898174, + 0.0039151073433458805, + 0.0014968229224905372, + 0.004922672640532255, + 0.0013599112862721086, + 0.0038657882250845432, + 0.007980930618941784, + 0.0020029956940561533, + 0.004150611814111471, + 0.0014692199183627963, + 0.004319885745644569, + 0.0572621151804924, + 0.0029032668098807335, + 0.004864079877734184, + 0.059922173619270325, + 0.0054451702162623405, + 0.008795560337603092, + 0.00400970596820116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.024928538128733635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.009805961512029171 + ], + "opposite_logprob_tokens": [ + -4.634806156158447 + ], + "logprob_reward": -0.009805961512029171, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805961512029171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29775798320770264, + 0.1404344141483307, + 0.13921743631362915, + 0.025049256160855293, + 0.07726136595010757, + 0.02790430188179016, + 0.06951413303613663, + 0.04195312410593033, + 0.019438061863183975, + 0.1614699363708496 + ], + "credit_map": [ + 0.022068504244089127, + 0.04556375741958618, + 0.027305761352181435, + 0.03053705208003521, + 0.044781286269426346, + 0.04245064780116081, + 0.03874703869223595, + 0.04630396142601967, + 0.012302789837121964, + 0.02075229585170746, + 0.01271762978285551, + 0.02244272641837597, + 0.022574327886104584, + 0.020939094945788383, + 0.01707615703344345, + 0.011629394255578518, + 0.012781440280377865, + 0.038465119898319244, + 0.00867455080151558, + 0.00900054071098566, + 0.02781454101204872, + 0.020215608179569244, + 0.0114526292309165, + 0.01081300713121891, + 0.0029282362665981054, + 0.0032848140690475702, + 0.002337160985916853, + 0.004425901919603348, + 0.003296863753348589, + 0.0022439381573349237, + 0.0032728391233831644, + 0.003259501652792096, + 0.007424668408930302, + 0.02340182103216648, + 0.005523893516510725, + 0.007506784982979298, + 0.013326257467269897, + 0.007906832732260227, + 0.0061080665327608585, + 0.006063042674213648, + 0.0025076481979340315, + 0.003958389163017273, + 0.002995372749865055, + 0.003917250782251358, + 0.004759938456118107, + 0.0028577656485140324, + 0.00303283566609025, + 0.003875100752338767, + 0.0054906997829675674, + 0.015424887649714947, + 0.004081813618540764, + 0.011330234818160534, + 0.010920976288616657, + 0.005696483887732029, + 0.012036113999783993, + 0.004532917868345976, + 0.003680128837004304, + 0.00592414103448391, + 0.003525024512782693, + 0.008444033563137054, + 0.005850417539477348, + 0.004612120334059, + 0.005352075677365065, + 0.004565184470266104, + 0.0014852429740130901, + 0.0028847772628068924, + 0.0012094946578145027, + 0.002062908373773098, + 0.0034209403675049543, + 0.0014370401622727513, + 0.005632985383272171, + 0.0013046726817265153, + 0.005128605756908655, + 0.05888228863477707, + 0.003988815471529961, + 0.006634472403675318, + 0.0604032501578331, + 0.0068954466842114925, + 0.014145138673484325, + 0.005391914397478104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.009805961512029171 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.001520549994893372 + ], + "opposite_logprob_tokens": [ + -6.50152063369751 + ], + "logprob_reward": -0.001520549994893372, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001520549994893372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10108467936515808, + 0.02884688600897789, + 0.0753689855337143, + 0.050285644829273224, + 0.06179226189851761, + 0.04279214143753052, + 0.2518509030342102, + 0.05179883539676666, + 0.06964743882417679, + 0.26653215289115906 + ], + "credit_map": [ + 0.005623107776045799, + 0.01855090819299221, + 0.007768783252686262, + 0.01936255395412445, + 0.01324465312063694, + 0.006224909797310829, + 0.021051565185189247, + 0.009258214384317398, + 0.001732841832563281, + 0.006050130818039179, + 0.0017237652791664004, + 0.00340634910389781, + 0.0072525059804320335, + 0.0028569039423018694, + 0.0035034706816077232, + 0.0023209217470139265, + 0.00688939681276679, + 0.02830824814736843, + 0.002947286469861865, + 0.005530516151338816, + 0.019014770165085793, + 0.0038813697174191475, + 0.004612513352185488, + 0.004184891004115343, + 0.004717667121440172, + 0.01253884844481945, + 0.003931344952434301, + 0.006746135652065277, + 0.009211125783622265, + 0.0038696543779224157, + 0.005479428451508284, + 0.003791444469243288, + 0.004882694222033024, + 0.012369286268949509, + 0.0039991410449147224, + 0.01137602236121893, + 0.009384512901306152, + 0.006440101657062769, + 0.00953720323741436, + 0.0038033099845051765, + 0.00337953120470047, + 0.006008104421198368, + 0.003696985775604844, + 0.008921677246689796, + 0.005295603536069393, + 0.0035240042489022017, + 0.006710251793265343, + 0.005255988799035549, + 0.02176312543451786, + 0.09211774170398712, + 0.012695975601673126, + 0.03589712083339691, + 0.036687254905700684, + 0.011430240236222744, + 0.02659372054040432, + 0.014665760099887848, + 0.003137943334877491, + 0.009335685521364212, + 0.003315466223284602, + 0.00829761940985918, + 0.0076096379198133945, + 0.00378023786470294, + 0.011167529970407486, + 0.005154720973223448, + 0.0036372109316289425, + 0.009463946335017681, + 0.003105940530076623, + 0.0101778544485569, + 0.020492058247327805, + 0.0030104268807917833, + 0.016160493716597557, + 0.003599517047405243, + 0.010370833799242973, + 0.10279498249292374, + 0.005504869855940342, + 0.012156516313552856, + 0.10112830996513367, + 0.011722293682396412, + 0.014787177555263042, + 0.008067196235060692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.001520549994893372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.00764184445142746 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.00764184445142746, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764184445142746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07810904830694199, + 0.05152582377195358, + 0.05758071690797806, + 0.04472727328538895, + 0.0703526958823204, + 0.08746007829904556, + 0.13496623933315277, + 0.12821680307388306, + 0.13004513084888458, + 0.21701613068580627 + ], + "credit_map": [ + 0.005076316185295582, + 0.015260999090969563, + 0.007062404416501522, + 0.011337035335600376, + 0.014668894000351429, + 0.005390624515712261, + 0.012061817571520805, + 0.007250957190990448, + 0.003240961115807295, + 0.007817089557647705, + 0.004683754872530699, + 0.0074144164100289345, + 0.0077002402395009995, + 0.006299328990280628, + 0.006796794477850199, + 0.007573236245661974, + 0.004558668006211519, + 0.013982926495373249, + 0.004977657459676266, + 0.005898420233279467, + 0.01051811408251524, + 0.005739033222198486, + 0.005858199670910835, + 0.006047697737812996, + 0.003932933788746595, + 0.008299889974296093, + 0.003779089078307152, + 0.008332609198987484, + 0.006871645804494619, + 0.0033661178313195705, + 0.0065036918967962265, + 0.0036412952467799187, + 0.008868098258972168, + 0.02317841164767742, + 0.004147524945437908, + 0.006363042630255222, + 0.01467054057866335, + 0.005216618999838829, + 0.004247189499437809, + 0.0036612695548683405, + 0.005857044830918312, + 0.011400552466511726, + 0.009559047408401966, + 0.011634605936706066, + 0.01173985656350851, + 0.00864443276077509, + 0.010269692167639732, + 0.01835484802722931, + 0.010057725012302399, + 0.03848319128155708, + 0.007620333693921566, + 0.024859175086021423, + 0.016173919662833214, + 0.005941139534115791, + 0.02467876859009266, + 0.007151986472308636, + 0.00827263668179512, + 0.02434174343943596, + 0.008152113296091557, + 0.02021457813680172, + 0.019737983122467995, + 0.011255408637225628, + 0.022610638290643692, + 0.013631717301905155, + 0.007643081247806549, + 0.014397037215530872, + 0.005681644193828106, + 0.01477153692394495, + 0.039045725017786026, + 0.006017719395458698, + 0.03696698322892189, + 0.005521404091268778, + 0.010578169487416744, + 0.08565956354141235, + 0.006132487207651138, + 0.009184270165860653, + 0.07961086183786392, + 0.008936028927564621, + 0.010908921249210835, + 0.006005828734487295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.00764184445142746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.004100843332707882 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100843332707882, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100843332707882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05602632835507393, + 0.020670494064688683, + 0.05808112770318985, + 0.07709237188100815, + 0.11991792172193527, + 0.04219998046755791, + 0.09714684635400772, + 0.026820015162229538, + 0.04374951124191284, + 0.45829537510871887 + ], + "credit_map": [ + 0.005406302399933338, + 0.008174472488462925, + 0.00657097389921546, + 0.006503491662442684, + 0.00877518579363823, + 0.005055336281657219, + 0.008579440414905548, + 0.0069611263461411, + 0.002297783037647605, + 0.003953427076339722, + 0.0015477845445275307, + 0.002286947565153241, + 0.003993908409029245, + 0.0019652601331472397, + 0.002663367660716176, + 0.0019620144739747047, + 0.005167827475816011, + 0.01789669506251812, + 0.002660507569089532, + 0.0045936983078718185, + 0.01203436590731144, + 0.005705005489289761, + 0.0054706004448235035, + 0.004552425350993872, + 0.00664555886760354, + 0.020141679793596268, + 0.0051202657632529736, + 0.01158127747476101, + 0.013161796145141125, + 0.0064070033840835094, + 0.008122200146317482, + 0.005912594497203827, + 0.011132507584989071, + 0.04384785518050194, + 0.005873701069504023, + 0.011960793286561966, + 0.027058979496359825, + 0.0060475836507976055, + 0.008361691609025002, + 0.005634806118905544, + 0.003586976556107402, + 0.005208131391555071, + 0.004255456849932671, + 0.0064850770868361, + 0.004843472503125668, + 0.00305226375348866, + 0.0066904774866998196, + 0.008078126236796379, + 0.00987232755869627, + 0.03976874053478241, + 0.004996434319764376, + 0.012294843792915344, + 0.01497290376573801, + 0.004288822878152132, + 0.005906987469643354, + 0.005045782774686813, + 0.0019627695437520742, + 0.004193076863884926, + 0.002363263163715601, + 0.00442022830247879, + 0.00393045786768198, + 0.0025695415679365396, + 0.005054323468357325, + 0.002326353918761015, + 0.0029474396724253893, + 0.008191204629838467, + 0.0029861354269087315, + 0.008801145479083061, + 0.007692408747971058, + 0.0025873929262161255, + 0.008175061084330082, + 0.002368723740801215, + 0.014725941233336926, + 0.18972283601760864, + 0.008992898277938366, + 0.012472990900278091, + 0.19520965218544006, + 0.017050597816705704, + 0.012300441041588783, + 0.007820013910531998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.004100843332707882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "motionless, fixed long shot captures a mirror-like lake, perfectly calm at dusk. The distant mountains merge into deep blue silhouettes against the horizon. In the very center of the lake, the air begins to twist and flicker violently, like a heat wave, as the silhouette of a massive ghostly ship slowly crystallizes from the void. This translucent vessel, its hull seemingly composed of solidified light, allows for a glimpse into its interior. Its hull is crisscrossed with an incredibly complex and dense mechanical structure inspired by classic Akira animations—countless pipes of varying sizes, exposed cable bundles, and layers of hydraulics and fins intertwined, creating a grotesque yet sophisticated industrial aesthetic. As it slowly materialized, its flickering outline stretched and deformed in the air, emitting a cold, faint blue glow. The light penetrated the hull and cast an equally distorted and flickering, yet unusually clear reflection on the calm lake. The entire grand and tranquil scene was shrouded in an awe-inspiring sense of mystery, as if a behemoth from another dimension was steadily and calmly descending into this world.", + "seed": 9775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_10/index_7859_seed_9775_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": -0.02381245684809983, + "vlm_logprob": -0.02381245684809983, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.02381245684809983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.09857726097106934 + ], + "opposite_logprob_tokens": [ + -2.4735772609710693 + ], + "logprob_reward": -0.09857726097106934, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09857726097106934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17148196697235107, + 0.3776344060897827, + 0.08011902868747711, + 0.05171998217701912, + 0.03186400979757309, + 0.03661538287997246, + 0.04553692042827606, + 0.04834531992673874, + 0.08643271028995514, + 0.07025031000375748 + ], + "credit_map": [ + 0.009670120663940907, + 0.07444600015878677, + 0.004510315600782633, + 0.004402236547321081, + 0.0376620776951313, + 0.03296628221869469, + 0.002963399514555931, + 0.004861537367105484, + 0.008761274628341198, + 0.13088887929916382, + 0.004487247206270695, + 0.010247640311717987, + 0.16583973169326782, + 0.0342632494866848, + 0.015101456083357334, + 0.008044918067753315, + 0.002427466679364443, + 0.02965741977095604, + 0.0017275280551984906, + 0.003958627115935087, + 0.03236904740333557, + 0.004782109521329403, + 0.003355441614985466, + 0.0018413932994008064, + 0.0044381157495081425, + 0.013003201223909855, + 0.0030670352280139923, + 0.0028654993511736393, + 0.012089021503925323, + 0.008051701821386814, + 0.004540320951491594, + 0.0036650854162871838, + 0.002075945260003209, + 0.012702960520982742, + 0.0011266226647421718, + 0.0013915698509663343, + 0.0037481067702174187, + 0.008329096250236034, + 0.0012897765263915062, + 0.0011999306734651327, + 0.0022415579296648502, + 0.0072862692177295685, + 0.002380745718255639, + 0.004478177521377802, + 0.005676864646375179, + 0.008280995301902294, + 0.0040985578671097755, + 0.0021722104866057634, + 0.003474213182926178, + 0.007523894309997559, + 0.003604297526180744, + 0.008684691973030567, + 0.007419601082801819, + 0.006335476879030466, + 0.0042321281507611275, + 0.004262614529579878, + 0.004157229792326689, + 0.008910167962312698, + 0.005926039069890976, + 0.004979428369551897, + 0.008027112111449242, + 0.006559011992067099, + 0.00415041483938694, + 0.005635916721075773, + 0.007936373353004456, + 0.011158579029142857, + 0.012515919283032417, + 0.009540878236293793, + 0.009261094033718109, + 0.012495109811425209, + 0.008579730056226254, + 0.01494502555578947, + 0.007700278423726559, + 0.010249723680317402, + 0.010808046907186508, + 0.008406544104218483, + 0.009590817615389824, + 0.00963061023503542, + 0.005814419128000736, + 0.008049871772527695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.09857726097106934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.001359372865408659 + ], + "opposite_logprob_tokens": [ + -6.626359462738037 + ], + "logprob_reward": -0.001359372865408659, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001359372865408659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3716895878314972, + 0.18693456053733826, + 0.10492618381977081, + 0.028888190165162086, + 0.05027994513511658, + 0.13049457967281342, + 0.025498755276203156, + 0.04739070683717728, + 0.030084071680903435, + 0.023813407868146896 + ], + "credit_map": [ + 0.02009442448616028, + 0.15407302975654602, + 0.008428316563367844, + 0.011385798454284668, + 0.0717751681804657, + 0.07481126487255096, + 0.016896311193704605, + 0.014225276187062263, + 0.007988246157765388, + 0.043475497514009476, + 0.005313176661729813, + 0.010697964578866959, + 0.07161512225866318, + 0.024904288351535797, + 0.016004258766770363, + 0.006936004851013422, + 0.0036444300785660744, + 0.03930278122425079, + 0.0025923901703208685, + 0.001716808183118701, + 0.04340840503573418, + 0.006777020171284676, + 0.003167514456436038, + 0.004316827282309532, + 0.00192917266394943, + 0.00491888914257288, + 0.0018832141067832708, + 0.0022250066976994276, + 0.009231659583747387, + 0.004331827163696289, + 0.0026465589180588722, + 0.0017218622379004955, + 0.0034021788742393255, + 0.01826377585530281, + 0.0027647404931485653, + 0.002420237986370921, + 0.010481156408786774, + 0.007940937764942646, + 0.002399690914899111, + 0.002607229631394148, + 0.005046560429036617, + 0.054961636662483215, + 0.002543790964409709, + 0.006319795269519091, + 0.023612005636096, + 0.030986353754997253, + 0.003671840997412801, + 0.0033525945618748665, + 0.002283022040501237, + 0.004272765014320612, + 0.002050290582701564, + 0.0033767230343073606, + 0.005043298006057739, + 0.004382358398288488, + 0.002035410376265645, + 0.002054888289421797, + 0.005938641726970673, + 0.009553352370858192, + 0.002371276495978236, + 0.00450116116553545, + 0.007503540720790625, + 0.01093544065952301, + 0.003673778148368001, + 0.002913517178967595, + 0.0028286778833717108, + 0.004238698165863752, + 0.0030193782877177, + 0.004673804622143507, + 0.005116403568536043, + 0.0033915876410901546, + 0.0034130823332816362, + 0.0034024366177618504, + 0.0024344855919480324, + 0.0028952183201909065, + 0.004366680514067411, + 0.003134683705866337, + 0.0026013769675046206, + 0.0028070814441889524, + 0.002593645127490163, + 0.002980237128213048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.001359372865408659 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.0017231153324246407 + ], + "opposite_logprob_tokens": [ + -6.376723289489746 + ], + "logprob_reward": -0.0017231153324246407, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017231153324246407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14062397181987762, + 0.22113795578479767, + 0.12935808300971985, + 0.05889435485005379, + 0.04774774610996246, + 0.06635444611310959, + 0.07186166197061539, + 0.059470321983098984, + 0.09250747412443161, + 0.11204394698143005 + ], + "credit_map": [ + 0.005168378818780184, + 0.04757678136229515, + 0.004373479168862104, + 0.007645006757229567, + 0.05334491655230522, + 0.010461587458848953, + 0.00704661151394248, + 0.005007219035178423, + 0.006521943490952253, + 0.06948380917310715, + 0.003031127853319049, + 0.007006052415817976, + 0.0952657163143158, + 0.024193175137043, + 0.010038778185844421, + 0.005597375798970461, + 0.0032403641380369663, + 0.051764313131570816, + 0.0017856205813586712, + 0.002188923303037882, + 0.05762217193841934, + 0.006563418544828892, + 0.004426481202244759, + 0.001766796805895865, + 0.0032932357862591743, + 0.019321851432323456, + 0.002434552414342761, + 0.004721440374851227, + 0.01525572594255209, + 0.006851139012724161, + 0.004630162846297026, + 0.002386250998824835, + 0.0035608410835266113, + 0.01573536917567253, + 0.002906693611294031, + 0.0035449655260890722, + 0.007724930066615343, + 0.00844417605549097, + 0.0030305238906294107, + 0.0028002483304589987, + 0.0042870137840509415, + 0.025577452033758163, + 0.001961062429472804, + 0.004265052266418934, + 0.0045723626390099525, + 0.020652735605835915, + 0.0031827448401600122, + 0.0018560272874310613, + 0.00589092168956995, + 0.0124783581122756, + 0.0047514792531728745, + 0.014048752374947071, + 0.015930593013763428, + 0.007468041498214006, + 0.006747030187398195, + 0.004546490032225847, + 0.005949718412011862, + 0.009164740331470966, + 0.005043812561780214, + 0.0065642292611300945, + 0.012670827098190784, + 0.007717014756053686, + 0.006279435008764267, + 0.006080546882003546, + 0.008575948886573315, + 0.014530746266245842, + 0.005443681497126818, + 0.011672000400722027, + 0.020884353667497635, + 0.01115412637591362, + 0.013894855044782162, + 0.006351771764457226, + 0.011429957114160061, + 0.0157000832259655, + 0.01343791838735342, + 0.01679426245391369, + 0.01828964613378048, + 0.01241512130945921, + 0.013374199159443378, + 0.010602762922644615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0017231153324246407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.0032011240255087614 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.0032011240255087614, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032011240255087614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3167795240879059, + 0.059255972504615784, + 0.10167351365089417, + 0.02816051058471203, + 0.03592580929398537, + 0.16369175910949707, + 0.03298144042491913, + 0.04423201084136963, + 0.06431841105222702, + 0.15298105776309967 + ], + "credit_map": [ + 0.00861390121281147, + 0.11590590327978134, + 0.005071126855909824, + 0.01428622379899025, + 0.12981826066970825, + 0.01783931814134121, + 0.0167330764234066, + 0.008511696942150593, + 0.002118853386491537, + 0.024640657007694244, + 0.001531572313979268, + 0.0013839732855558395, + 0.02340918406844139, + 0.002710449742153287, + 0.00177689827978611, + 0.0016843825578689575, + 0.002717128489166498, + 0.03874605894088745, + 0.0014787839027121663, + 0.002061923500150442, + 0.045931920409202576, + 0.006672744173556566, + 0.002803815295919776, + 0.0012611366109922528, + 0.0019848847296088934, + 0.007552003487944603, + 0.0012708931462839246, + 0.002276648534461856, + 0.008327843621373177, + 0.0029183842707425356, + 0.0025897545274347067, + 0.0012400980340316892, + 0.0028956602327525616, + 0.011072097346186638, + 0.00245377398096025, + 0.0027523641474545, + 0.005425706505775452, + 0.00653765769675374, + 0.0022649711463600397, + 0.0025235803332179785, + 0.005275479517877102, + 0.07485861331224442, + 0.0019085193052887917, + 0.0031544712837785482, + 0.015850907191634178, + 0.05721830576658249, + 0.0026047718711197376, + 0.002820702036842704, + 0.0025289098266512156, + 0.005571110639721155, + 0.002853936981409788, + 0.0044953348115086555, + 0.006090872921049595, + 0.0052847242914140224, + 0.003167941467836499, + 0.002988609718158841, + 0.005066159646958113, + 0.007992955856025219, + 0.00442053098231554, + 0.003999342676252127, + 0.00737124215811491, + 0.007197479717433453, + 0.0036999566946178675, + 0.004484341945499182, + 0.007564462721347809, + 0.008433066308498383, + 0.0069166505709290504, + 0.008048560470342636, + 0.010865154676139355, + 0.008486500009894371, + 0.005727487150579691, + 0.008276521228253841, + 0.01771637424826622, + 0.024595065042376518, + 0.020299158990383148, + 0.020501641556620598, + 0.02729051187634468, + 0.016374832019209862, + 0.012119673192501068, + 0.01408378779888153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0032011240255087614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 mirror-like lake that is perfectly calm at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are distant mountains visible as deep blue silhouettes against the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive ghostly ship slowly crystallize from the void in the center of the lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ship's hull translucent with a complex mechanical structure featuring pipes, cables, and fins inspired by classic Akira animations?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the ship emit a cold, faint blue glow that casts a distorted yet clear reflection on the lake 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": [ + -0.014201411046087742 + ], + "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": [ + 0.24014049768447876, + 0.08654946088790894, + 0.047355327755212784, + 0.04797790199518204, + 0.034686699509620667, + 0.23304370045661926, + 0.09524750709533691, + 0.04540542885661125, + 0.09267263859510422, + 0.07692079246044159 + ], + "credit_map": [ + 0.009867838583886623, + 0.10906507074832916, + 0.004285240080207586, + 0.004170780070126057, + 0.06563601642847061, + 0.03971070796251297, + 0.003113429294899106, + 0.004291438963264227, + 0.002537486841902137, + 0.036292411386966705, + 0.0014642288442701101, + 0.0021790768951177597, + 0.036796268075704575, + 0.0035701063461601734, + 0.001985235372558236, + 0.0017246537609025836, + 0.0019891695119440556, + 0.019945967942476273, + 0.0010675303637981415, + 0.0009285470587201416, + 0.01912611722946167, + 0.0021230606362223625, + 0.0013056814204901457, + 0.000869258597958833, + 0.0037304521538317204, + 0.018376292660832405, + 0.0023613127414137125, + 0.0017893878975883126, + 0.01258813589811325, + 0.0051468778401613235, + 0.001713635865598917, + 0.002271811943501234, + 0.002542750211432576, + 0.012655925005674362, + 0.001888708327896893, + 0.0019122163066640496, + 0.009161917492747307, + 0.003399715293198824, + 0.0013407344231382012, + 0.001784737454727292, + 0.008374858647584915, + 0.10907567292451859, + 0.0023128874599933624, + 0.004663200583308935, + 0.03588153049349785, + 0.06766440719366074, + 0.0018176939338445663, + 0.003253479488193989, + 0.005825783591717482, + 0.015629339963197708, + 0.006643296219408512, + 0.01813458651304245, + 0.021400539204478264, + 0.013677096925675869, + 0.007704619318246841, + 0.006232247222214937, + 0.004829629324376583, + 0.00851640198379755, + 0.0031665570568293333, + 0.004902758169919252, + 0.008112361654639244, + 0.007738681044429541, + 0.004106742795556784, + 0.004032301250845194, + 0.007320547942072153, + 0.01282651349902153, + 0.01119943242520094, + 0.012007361277937889, + 0.015124485827982426, + 0.011156963184475899, + 0.009472448378801346, + 0.013564891181886196, + 0.007740088272839785, + 0.009599232114851475, + 0.007046231534332037, + 0.013469449244439602, + 0.011294040828943253, + 0.009264428168535233, + 0.008517937734723091, + 0.009989388287067413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.014201411046087742 + } + ], + "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/0000058/group_11/index_tensor([5512, 5512])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_11/index_tensor([5512, 5512])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0b35d817ee3e75238f7ea4734fb3b367436a6621 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_11/index_tensor([5512, 5512])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level, medium shot frames a heavy, dark mahogany table in a dimly lit, classic study. The table's polished surface vaguely reflects the warm, concentrated glow of a green-shaded bank lamp, casting long, dancing shadows across the room. A stack of aged, off-white manuscripts rests placidly on the tabletop. Suddenly, an unseen force disrupts the stillness; the papers begin to flutter, then violently lift and swirl above the table, forming a chaotic vortex before dispersing and drifting onto the dark wooden floor. Simultaneously, the table's drawers, adorned with ornate, tarnished brass handles, begin to creak violently. In agitated and convulsive movements, they are flung open and then violently closed in a rapid, arrhythmic sequence, causing the entire sturdy furniture to tremble. The entire scene is imbued with a tense, supernatural atmosphere, the dramatic interplay of focused light and deep shadows enhancing the eerie, inexplicable animation of the inanimate objects.", + "seed": 23076, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_11/index_5512_seed_23076_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": -0.02693510092794895, + "vlm_logprob": -0.02693510092794895, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.02693510092794895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.004902129527181387 + ], + "opposite_logprob_tokens": [ + -7.129902362823486 + ], + "logprob_reward": -0.004902129527181387, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.004902129527181387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11131507903337479, + 0.055633362382650375, + 0.027777450159192085, + 0.04297332838177681, + 0.03511054441332817, + 0.05519271269440651, + 0.08940824866294861, + 0.1618914008140564, + 0.21559320390224457, + 0.20510469377040863 + ], + "credit_map": [ + 0.014260264113545418, + 0.01308682095259428, + 0.011369379237294197, + 0.012712415307760239, + 0.019486770033836365, + 0.01686367765069008, + 0.012374958023428917, + 0.011160795576870441, + 0.010177590884268284, + 0.005388712976127863, + 0.004527987912297249, + 0.005784465000033379, + 0.0067707872949540615, + 0.01252309326082468, + 0.006210861261934042, + 0.004249863792210817, + 0.00313781900331378, + 0.002699998440220952, + 0.0019552218727767467, + 0.002649902366101742, + 0.005784616805613041, + 0.004336402285844088, + 0.004434233997017145, + 0.002779257483780384, + 0.011826115660369396, + 0.00285545876249671, + 0.0019813072867691517, + 0.0018899482674896717, + 0.004543457180261612, + 0.014174036681652069, + 0.0026251813396811485, + 0.003077819710597396, + 0.003568457905203104, + 0.004880853462964296, + 0.0020028415601700544, + 0.002951675560325384, + 0.009293324314057827, + 0.003941768314689398, + 0.00524118822067976, + 0.0032304360065609217, + 0.010875764302909374, + 0.0048691099509596825, + 0.0038114390335977077, + 0.00493451114743948, + 0.004157980903983116, + 0.010264011099934578, + 0.003902964061126113, + 0.01237693801522255, + 0.008168268017470837, + 0.006587090436369181, + 0.014033096842467785, + 0.01494770310819149, + 0.009867741726338863, + 0.010063528083264828, + 0.011576260440051556, + 0.014164566993713379, + 0.011471052654087543, + 0.009626727551221848, + 0.02615148201584816, + 0.05298824608325958, + 0.010064154863357544, + 0.012378600426018238, + 0.025179212912917137, + 0.014031918719410896, + 0.015861663967370987, + 0.042240843176841736, + 0.026133492588996887, + 0.03800316900014877, + 0.02426350861787796, + 0.01545578334480524, + 0.030977876856923103, + 0.022656863555312157, + 0.015597768127918243, + 0.05024948716163635, + 0.0223222766071558, + 0.0371217206120491, + 0.03170973062515259, + 0.009657684713602066, + 0.021174412220716476, + 0.017271606251597404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.004902129527181387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.0067544239573180676 + ], + "opposite_logprob_tokens": [ + -5.006754398345947 + ], + "logprob_reward": -0.0067544239573180676, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067544239573180676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2949286699295044, + 0.09429414570331573, + 0.03957856073975563, + 0.05819988250732422, + 0.04201945662498474, + 0.06431780010461807, + 0.047774169594049454, + 0.09502772241830826, + 0.17765596508979797, + 0.08620361238718033 + ], + "credit_map": [ + 0.0335417315363884, + 0.03534082695841789, + 0.020298773422837257, + 0.02087528072297573, + 0.08201411366462708, + 0.031147755682468414, + 0.04368956759572029, + 0.028020618483424187, + 0.015478148125112057, + 0.010248624719679356, + 0.007780696265399456, + 0.008104086853563786, + 0.014123332686722279, + 0.015371400862932205, + 0.014585998840630054, + 0.008601849898695946, + 0.004276050720363855, + 0.003827448468655348, + 0.0036402069963514805, + 0.003812285838648677, + 0.008961481973528862, + 0.005238453857600689, + 0.0050112586468458176, + 0.004811369813978672, + 0.01387743093073368, + 0.00472211791202426, + 0.0028504100628197193, + 0.0033587045036256313, + 0.009335764683783054, + 0.017383167520165443, + 0.0030255464371293783, + 0.0036467460449784994, + 0.004271637182682753, + 0.007120959926396608, + 0.002414920600131154, + 0.0031333595979958773, + 0.012529981322586536, + 0.004742085002362728, + 0.004189135041087866, + 0.0036173781845718622, + 0.007964175194501877, + 0.004705017898231745, + 0.004174032248556614, + 0.005510208196938038, + 0.007824144326150417, + 0.01651690900325775, + 0.007488009054213762, + 0.010135307908058167, + 0.004717269446700811, + 0.003923185169696808, + 0.007204422727227211, + 0.008875035680830479, + 0.004791483283042908, + 0.005137898027896881, + 0.006603432819247246, + 0.006521441508084536, + 0.008973779156804085, + 0.007773616351187229, + 0.014015250839293003, + 0.020394161343574524, + 0.007792883552610874, + 0.009535943157970905, + 0.013233909383416176, + 0.013308179564774036, + 0.011525474488735199, + 0.03205712512135506, + 0.024446412920951843, + 0.04641933739185333, + 0.017962250858545303, + 0.01064941007643938, + 0.021838489919900894, + 0.012757464312016964, + 0.006037258077412844, + 0.01903507299721241, + 0.011430138722062111, + 0.015870902687311172, + 0.01256336085498333, + 0.004507394973188639, + 0.010653876699507236, + 0.006105611100792885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0067544239573180676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.03810567781329155 + ], + "opposite_logprob_tokens": [ + -3.2881057262420654 + ], + "logprob_reward": -0.03810567781329155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810567781329155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09017493575811386, + 0.0802989974617958, + 0.049571890383958817, + 0.034582797437906265, + 0.02467571571469307, + 0.05962054803967476, + 0.09408456832170486, + 0.13368967175483704, + 0.2790715992450714, + 0.15422934293746948 + ], + "credit_map": [ + 0.012488206848502159, + 0.009837820194661617, + 0.008426749147474766, + 0.008410647511482239, + 0.018683237954974174, + 0.011418115347623825, + 0.0127058494836092, + 0.00820431299507618, + 0.01347303669899702, + 0.0068812742829322815, + 0.004765546415001154, + 0.007913149893283844, + 0.012893793173134327, + 0.014988060109317303, + 0.012432853691279888, + 0.006951289251446724, + 0.008930227719247341, + 0.004235120490193367, + 0.0023503159172832966, + 0.004013329278677702, + 0.009924395941197872, + 0.010968472808599472, + 0.0052023036405444145, + 0.003947721794247627, + 0.007436148356646299, + 0.00337530137039721, + 0.0016124923713505268, + 0.001974656479433179, + 0.005743017885833979, + 0.00871175155043602, + 0.003445566864684224, + 0.00228386209346354, + 0.0026431456208229065, + 0.0030247909016907215, + 0.0018833137582987547, + 0.0024596319999545813, + 0.004540832247585058, + 0.0030504674650728703, + 0.004849793389439583, + 0.002223741030320525, + 0.006680778693407774, + 0.005448896437883377, + 0.00415956974029541, + 0.006835297215729952, + 0.008072097785770893, + 0.008081082254648209, + 0.006140061654150486, + 0.014202764257788658, + 0.00877438671886921, + 0.007920647040009499, + 0.013993848115205765, + 0.020677117630839348, + 0.009871716611087322, + 0.009068764746189117, + 0.010541209951043129, + 0.013236882165074348, + 0.011807196773588657, + 0.01457001082599163, + 0.020248936489224434, + 0.029451098293066025, + 0.013293281197547913, + 0.012839998118579388, + 0.01705450564622879, + 0.014424643479287624, + 0.021363750100135803, + 0.05303718522191048, + 0.034463830292224884, + 0.0700082927942276, + 0.02543286234140396, + 0.01850622519850731, + 0.03385017439723015, + 0.022409269586205482, + 0.010982625186443329, + 0.024258485063910484, + 0.02047375775873661, + 0.017247194424271584, + 0.0209260955452919, + 0.017946811392903328, + 0.017249520868062973, + 0.025144852697849274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.03810567781329155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.07896102219820023 + ], + "opposite_logprob_tokens": [ + -2.578961133956909 + ], + "logprob_reward": -0.07896102219820023, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896102219820023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04647419974207878, + 0.020650863647460938, + 0.030817775055766106, + 0.029108475893735886, + 0.019281568005681038, + 0.02962701953947544, + 0.126686230301857, + 0.10172073543071747, + 0.5138623714447021, + 0.08177080750465393 + ], + "credit_map": [ + 0.007114232052117586, + 0.004927982576191425, + 0.00487638870254159, + 0.004672599956393242, + 0.005706544499844313, + 0.0062692854553461075, + 0.007361080963164568, + 0.005546083208173513, + 0.002127269981428981, + 0.0020618995185941458, + 0.001766374334692955, + 0.0017139370320364833, + 0.004838974215090275, + 0.002490720711648464, + 0.002700550714507699, + 0.002951138187199831, + 0.0025894977152347565, + 0.002920133527368307, + 0.0026511007454246283, + 0.003539621364325285, + 0.007072860840708017, + 0.0041306838393211365, + 0.005004006903618574, + 0.0029098691884428263, + 0.005609630607068539, + 0.002723576035350561, + 0.001343423849903047, + 0.0026235959958285093, + 0.004492905456572771, + 0.006661284249275923, + 0.003601115895435214, + 0.002052943455055356, + 0.002259732922539115, + 0.002243440132588148, + 0.002055410761386156, + 0.0015205825911834836, + 0.004326685331761837, + 0.003046723548322916, + 0.0019354468677192926, + 0.0018935459665954113, + 0.0026543603744357824, + 0.0021476629190146923, + 0.0021727338898926973, + 0.00370239932090044, + 0.005212315358221531, + 0.003903619246557355, + 0.004912981763482094, + 0.004920948762446642, + 0.00903406459838152, + 0.00811882596462965, + 0.018793562427163124, + 0.022709712386131287, + 0.017494644969701767, + 0.013416556641459465, + 0.015230569057166576, + 0.02188829332590103, + 0.006609003990888596, + 0.007254326716065407, + 0.01734534464776516, + 0.031614042818546295, + 0.006882520858198404, + 0.006890172604471445, + 0.014694919809699059, + 0.010430406779050827, + 0.03162503242492676, + 0.1107252985239029, + 0.07126409560441971, + 0.144199937582016, + 0.0382552407681942, + 0.025523828342556953, + 0.05971932411193848, + 0.032549627125263214, + 0.00705077825114131, + 0.017583690583705902, + 0.009739120490849018, + 0.010277344845235348, + 0.012565641663968563, + 0.005826665088534355, + 0.009128617122769356, + 0.009598948061466217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.07896102219820023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.005952251143753529 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952251143753529, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952251143753529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09084859490394592, + 0.04034902900457382, + 0.06024038791656494, + 0.05398918315768242, + 0.04023894667625427, + 0.056865617632865906, + 0.09899693727493286, + 0.1412954479455948, + 0.302654892206192, + 0.11452094465494156 + ], + "credit_map": [ + 0.011842756532132626, + 0.010297724977135658, + 0.010489222593605518, + 0.010090640746057034, + 0.013398978859186172, + 0.009857511147856712, + 0.016280896961688995, + 0.008590862154960632, + 0.004539129324257374, + 0.0038164230063557625, + 0.0037512595299631357, + 0.004031846299767494, + 0.00849068071693182, + 0.004298878367990255, + 0.00783108826726675, + 0.00358972349204123, + 0.005650653969496489, + 0.005385927390307188, + 0.0047312104143202305, + 0.006121357902884483, + 0.01507788896560669, + 0.009236903861165047, + 0.008497131988406181, + 0.005539312493056059, + 0.012825039215385914, + 0.005301497410982847, + 0.0023846905678510666, + 0.002569109434261918, + 0.007928593084216118, + 0.01594069041311741, + 0.0039946576580405235, + 0.0030449058394879103, + 0.004128617700189352, + 0.004597092047333717, + 0.0025949175469577312, + 0.00435984879732132, + 0.009432519786059856, + 0.004997020121663809, + 0.006382977589964867, + 0.0037459535524249077, + 0.005974926985800266, + 0.006355050019919872, + 0.006509503349661827, + 0.007087156176567078, + 0.01038274634629488, + 0.005609212908893824, + 0.009782861918210983, + 0.005164163187146187, + 0.008583465591073036, + 0.008871749974787235, + 0.015315650030970573, + 0.017766335979104042, + 0.011046023108065128, + 0.009522481821477413, + 0.010860173031687737, + 0.01703106053173542, + 0.008913109079003334, + 0.015427663922309875, + 0.022911436855793, + 0.03448953479528427, + 0.01566535234451294, + 0.011245050467550755, + 0.018445882946252823, + 0.014197428710758686, + 0.015725361183285713, + 0.0504438616335392, + 0.0405859611928463, + 0.07655282318592072, + 0.02769472822546959, + 0.018243571743369102, + 0.04250258952379227, + 0.030906012281775475, + 0.007637705188244581, + 0.015779631212353706, + 0.014316692017018795, + 0.014403860084712505, + 0.017799511551856995, + 0.012961208820343018, + 0.012962283566594124, + 0.018660053610801697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.005952251143753529 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot frames a heavy, dark mahogany table in a dimly lit, classic study. The table's polished surface vaguely reflects the warm, concentrated glow of a green-shaded bank lamp, casting long, dancing shadows across the room. A stack of aged, off-white manuscripts rests placidly on the tabletop. Suddenly, an unseen force disrupts the stillness; the papers begin to flutter, then violently lift and swirl above the table, forming a chaotic vortex before dispersing and drifting onto the dark wooden floor. Simultaneously, the table's drawers, adorned with ornate, tarnished brass handles, begin to creak violently. In agitated and convulsive movements, they are flung open and then violently closed in a rapid, arrhythmic sequence, causing the entire sturdy furniture to tremble. The entire scene is imbued with a tense, supernatural atmosphere, the dramatic interplay of focused light and deep shadows enhancing the eerie, inexplicable animation of the inanimate objects.", + "seed": 23077, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_11/index_5512_seed_23077_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": -0.09082160796970129, + "vlm_logprob": -0.09082160796970129, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.09082160796970129, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.0069090514443814754 + ], + "opposite_logprob_tokens": [ + -5.881908893585205 + ], + "logprob_reward": -0.0069090514443814754, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0069090514443814754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0655064657330513, + 0.01835784502327442, + 0.008804273791611195, + 0.012735136784613132, + 0.024934072047472, + 0.03322996199131012, + 0.026297304779291153, + 0.056520797312259674, + 0.5195326209068298, + 0.23408150672912598 + ], + "credit_map": [ + 0.009712017141282558, + 0.009989028796553612, + 0.008301866240799427, + 0.0069234250113368034, + 0.007082059513777494, + 0.011377288028597832, + 0.0065780868753790855, + 0.005542695056647062, + 0.003930410370230675, + 0.0028915058355778456, + 0.0019109342247247696, + 0.0013378615258261561, + 0.0018153988057747483, + 0.002798195695504546, + 0.0017090312903746963, + 0.0019645069260150194, + 0.0022111383732408285, + 0.0016061223577708006, + 0.0010640175314620137, + 0.0006721238605678082, + 0.0008063848945312202, + 0.001065899501554668, + 0.000693743524607271, + 0.0006848443299531937, + 0.00261130603030324, + 0.002493114909157157, + 0.0017959247343242168, + 0.001289385138079524, + 0.0012125977082177997, + 0.0011762952199205756, + 0.001241438090801239, + 0.0009150748956017196, + 0.004535186570137739, + 0.0033568209037184715, + 0.0026404194068163633, + 0.0026763398200273514, + 0.003266645362600684, + 0.00415343651548028, + 0.001964479684829712, + 0.002340743551030755, + 0.004736745730042458, + 0.00436337711289525, + 0.002052254742011428, + 0.005250558257102966, + 0.004930269904434681, + 0.0031068939715623856, + 0.005278868600726128, + 0.0035109941381961107, + 0.003692462807521224, + 0.0037237219512462616, + 0.004522970877587795, + 0.0030516861006617546, + 0.002823718125000596, + 0.004373504314571619, + 0.0025834059342741966, + 0.0015258340863510966, + 0.007135478314012289, + 0.0065989107824862, + 0.00839381106197834, + 0.010077136568725109, + 0.004576758947223425, + 0.004944273270666599, + 0.008857231587171555, + 0.0059371995739638805, + 0.027377091348171234, + 0.027928799390792847, + 0.057321809232234955, + 0.19931188225746155, + 0.015732666477560997, + 0.016056355088949203, + 0.12143629044294357, + 0.05436772480607033, + 0.02993142604827881, + 0.018349958583712578, + 0.026109257712960243, + 0.026351775974035263, + 0.03180384263396263, + 0.02491619996726513, + 0.04528548941016197, + 0.031333547085523605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0069090514443814754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.011097023263573647 + ], + "opposite_logprob_tokens": [ + -4.511096954345703 + ], + "logprob_reward": -0.011097023263573647, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097023263573647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0486636757850647, + 0.02423637919127941, + 0.020842459052801132, + 0.02233806811273098, + 0.052342262119054794, + 0.07127372175455093, + 0.05348118394613266, + 0.028171466663479805, + 0.4630163013935089, + 0.2156345397233963 + ], + "credit_map": [ + 0.009250069968402386, + 0.006810130551457405, + 0.006406406871974468, + 0.005564406979829073, + 0.005296160466969013, + 0.005997291300445795, + 0.00524689769372344, + 0.004092310555279255, + 0.005374945234507322, + 0.00374378589913249, + 0.0030949446372687817, + 0.0022307897452265024, + 0.002246237825602293, + 0.002530410187318921, + 0.002236922737210989, + 0.002778341993689537, + 0.004458464216440916, + 0.0032836992759257555, + 0.002733982400968671, + 0.0022600635420531034, + 0.0020876836497336626, + 0.002140990225598216, + 0.002262512454763055, + 0.001615060493350029, + 0.0029963385313749313, + 0.004770784173160791, + 0.0025391310919076204, + 0.0021788286976516247, + 0.0028775546234101057, + 0.002798765432089567, + 0.0025946260429918766, + 0.001582037191838026, + 0.008552812039852142, + 0.0082181291654706, + 0.008333145640790462, + 0.005095251370221376, + 0.005357919726520777, + 0.007427820935845375, + 0.00500965490937233, + 0.004347524140030146, + 0.01268468052148819, + 0.00948277022689581, + 0.0063765705563127995, + 0.015047716908156872, + 0.00723197590559721, + 0.006587032228708267, + 0.008531059138476849, + 0.005331909283995628, + 0.004207627382129431, + 0.007441211957484484, + 0.003749826457351446, + 0.0044492874294519424, + 0.008398222737014294, + 0.010431144386529922, + 0.008587879128754139, + 0.006215980276465416, + 0.003025614656507969, + 0.0035830882843583822, + 0.004104695748537779, + 0.004611324053257704, + 0.0026436527259647846, + 0.0028201278764754534, + 0.004884794354438782, + 0.0024981650058180094, + 0.015024031512439251, + 0.02164301835000515, + 0.06080379709601402, + 0.19222617149353027, + 0.011251967400312424, + 0.009706446900963783, + 0.10452382266521454, + 0.04783700779080391, + 0.015602744184434414, + 0.018163787201046944, + 0.038661181926727295, + 0.029422873631119728, + 0.03119734488427639, + 0.03632901981472969, + 0.02556474879384041, + 0.020692812278866768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.011097023263573647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.043067168444395065 + ], + "opposite_logprob_tokens": [ + -3.168067216873169 + ], + "logprob_reward": -0.043067168444395065, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067168444395065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045732371509075165, + 0.012444543652236462, + 0.013248470611870289, + 0.013321766629815102, + 0.10104753077030182, + 0.02072303183376789, + 0.013095688074827194, + 0.06294988840818405, + 0.5732814073562622, + 0.1441553384065628 + ], + "credit_map": [ + 0.007380098104476929, + 0.00663044722750783, + 0.005192131269723177, + 0.004536536987870932, + 0.005536802113056183, + 0.007753447163850069, + 0.004869762808084488, + 0.0038331467658281326, + 0.0025829579681158066, + 0.0017846755217760801, + 0.0017709459643810987, + 0.001306167570874095, + 0.0011663902550935745, + 0.0014780687633901834, + 0.0014888810692355037, + 0.0008664567722007632, + 0.003210552968084812, + 0.001851010718382895, + 0.0018600999610498548, + 0.0014168140478432178, + 0.0012211122084408998, + 0.001595810055732727, + 0.0013366913190111518, + 0.0007563791587017477, + 0.003093419363722205, + 0.002229057950899005, + 0.0016916616586968303, + 0.0014337890315800905, + 0.0011581945000216365, + 0.0012250402942299843, + 0.001669490011408925, + 0.0008211142849177122, + 0.010859359987080097, + 0.01818026602268219, + 0.012608831748366356, + 0.016266684979200363, + 0.010472326539456844, + 0.014073201455175877, + 0.012297182343900204, + 0.006289676297456026, + 0.002785447984933853, + 0.003375765634700656, + 0.0017903585685417056, + 0.002178576774895191, + 0.0036101359874010086, + 0.0027458651456981897, + 0.0025032106786966324, + 0.0017336716409772635, + 0.0021320818923413754, + 0.0019185456912964582, + 0.0014516233932226896, + 0.0014048030134290457, + 0.001236295560374856, + 0.0018287233542650938, + 0.0019542237278074026, + 0.0011693919077515602, + 0.005665677133947611, + 0.00834945309907198, + 0.008951598778367043, + 0.01145416684448719, + 0.009114625863730907, + 0.006446312181651592, + 0.008286508731544018, + 0.0046815467067062855, + 0.017875516787171364, + 0.026924846693873405, + 0.07526638358831406, + 0.24043488502502441, + 0.010994354262948036, + 0.011104976758360863, + 0.13271474838256836, + 0.057965703308582306, + 0.010782278142869473, + 0.01244509406387806, + 0.020669996738433838, + 0.01509012933820486, + 0.021221181377768517, + 0.024750404059886932, + 0.01630120351910591, + 0.022895049303770065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.043067168444395065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.3870823085308075 + ], + "opposite_logprob_tokens": [ + -1.1370823383331299 + ], + "logprob_reward": -0.3870823085308075, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870823085308075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011690567247569561, + 0.009845954366028309, + 0.010772393085062504, + 0.01234439667314291, + 0.01856311969459057, + 0.0175777580589056, + 0.011131564155220985, + 0.03719693794846535, + 0.5924058556556702, + 0.278471440076828 + ], + "credit_map": [ + 0.0022046605590730906, + 0.002028814749792218, + 0.0013999583898112178, + 0.0009768083691596985, + 0.001226375112310052, + 0.0013603009283542633, + 0.0012446334585547447, + 0.0012490146327763796, + 0.0017866338603198528, + 0.001449672388844192, + 0.001091869198717177, + 0.000993887661024928, + 0.0012442887527868152, + 0.001481638173572719, + 0.0009442537557333708, + 0.0008537101675756276, + 0.002137203235179186, + 0.0016535919858142734, + 0.0011464626295492053, + 0.0009745314600877464, + 0.001333587453700602, + 0.001716216909699142, + 0.0008831511368043721, + 0.0009276482160203159, + 0.0017474874621257186, + 0.0019600416999310255, + 0.0013560015941038728, + 0.0012087812647223473, + 0.001614284235984087, + 0.0018496779957786202, + 0.0013059919001534581, + 0.0013021307531744242, + 0.00219273311085999, + 0.003325287951156497, + 0.001835040980949998, + 0.0029298707377165556, + 0.0017699202289804816, + 0.002613268792629242, + 0.002308520022779703, + 0.0015884792665019631, + 0.001957547152414918, + 0.002321183681488037, + 0.0012905356707051396, + 0.0024988704826682806, + 0.0028947661630809307, + 0.0025049210526049137, + 0.0021305205300450325, + 0.001979414839297533, + 0.0013719386188313365, + 0.001820608857087791, + 0.001595664769411087, + 0.0014846211997792125, + 0.0013736918335780501, + 0.0013607898727059364, + 0.0013975136680528522, + 0.0007267352193593979, + 0.004559115041047335, + 0.004671221133321524, + 0.004728762432932854, + 0.005882652476429939, + 0.0034874221310019493, + 0.0027805869467556477, + 0.006250943522900343, + 0.004836236592382193, + 0.019969817250967026, + 0.027897534891963005, + 0.08168375492095947, + 0.24844671785831451, + 0.01404933538287878, + 0.011829025112092495, + 0.1326424926519394, + 0.05588724464178085, + 0.021044166758656502, + 0.023416418582201004, + 0.04850231856107712, + 0.037432510405778885, + 0.041871797293424606, + 0.046936675906181335, + 0.03286590427160263, + 0.02640162780880928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.3870823085308075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.005952488165348768 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952488165348768, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952488165348768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08321874588727951, + 0.0453299842774868, + 0.03724781051278114, + 0.03754381090402603, + 0.055686820298433304, + 0.05330968648195267, + 0.040809266269207, + 0.19018782675266266, + 0.27022984623908997, + 0.18643620610237122 + ], + "credit_map": [ + 0.013255788013339043, + 0.012110365554690361, + 0.012389796786010265, + 0.010903076268732548, + 0.008224301040172577, + 0.009621072560548782, + 0.01098920963704586, + 0.00572514021769166, + 0.010427129454910755, + 0.006128505803644657, + 0.0068005784414708614, + 0.005614681635051966, + 0.004204758908599615, + 0.005465865600854158, + 0.0037051027175039053, + 0.0029833645094186068, + 0.008311033248901367, + 0.005315144546329975, + 0.005621135234832764, + 0.003971100319176912, + 0.0035064395051449537, + 0.004830213263630867, + 0.002905658446252346, + 0.0027870871126651764, + 0.006775311194360256, + 0.005346314050257206, + 0.004845567513257265, + 0.004104156047105789, + 0.004170743748545647, + 0.004962147679179907, + 0.004569871351122856, + 0.002769698388874531, + 0.006487544160336256, + 0.01040065847337246, + 0.005473625846207142, + 0.007924212142825127, + 0.005462652537971735, + 0.008215269073843956, + 0.005962188355624676, + 0.005760665517300367, + 0.0043714288622140884, + 0.007686355616897345, + 0.003992434591054916, + 0.004897743929177523, + 0.009861597791314125, + 0.007388443220406771, + 0.008728601969778538, + 0.006383081898093224, + 0.005407229065895081, + 0.006703146733343601, + 0.0059435926377773285, + 0.005524381063878536, + 0.004856285639107227, + 0.004862356465309858, + 0.00447661429643631, + 0.0030356610659509897, + 0.017608417198061943, + 0.025035085156559944, + 0.02384381741285324, + 0.02692156843841076, + 0.028027592226862907, + 0.02478460595011711, + 0.027316031977534294, + 0.01665070466697216, + 0.02141648903489113, + 0.02139621041715145, + 0.05612490326166153, + 0.08392398804426193, + 0.019775481894612312, + 0.01959552802145481, + 0.031070921570062637, + 0.016926316544413567, + 0.015109309926629066, + 0.01934216544032097, + 0.023548131808638573, + 0.017065828666090965, + 0.029073797166347504, + 0.024870188906788826, + 0.022906729951500893, + 0.034520067274570465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.005952488165348768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot frames a heavy, dark mahogany table in a dimly lit, classic study. The table's polished surface vaguely reflects the warm, concentrated glow of a green-shaded bank lamp, casting long, dancing shadows across the room. A stack of aged, off-white manuscripts rests placidly on the tabletop. Suddenly, an unseen force disrupts the stillness; the papers begin to flutter, then violently lift and swirl above the table, forming a chaotic vortex before dispersing and drifting onto the dark wooden floor. Simultaneously, the table's drawers, adorned with ornate, tarnished brass handles, begin to creak violently. In agitated and convulsive movements, they are flung open and then violently closed in a rapid, arrhythmic sequence, causing the entire sturdy furniture to tremble. The entire scene is imbued with a tense, supernatural atmosphere, the dramatic interplay of focused light and deep shadows enhancing the eerie, inexplicable animation of the inanimate objects.", + "seed": 23078, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_11/index_5512_seed_23078_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": -0.029821015242487193, + "vlm_logprob": -0.029821015242487193, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.029821015242487193, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.007742987480014563 + ], + "opposite_logprob_tokens": [ + -6.007742881774902 + ], + "logprob_reward": -0.007742987480014563, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007742987480014563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11560273915529251, + 0.06533801555633545, + 0.07853475213050842, + 0.058534953743219376, + 0.04616311937570572, + 0.15655934810638428, + 0.13197489082813263, + 0.07560211420059204, + 0.1585666835308075, + 0.11312340945005417 + ], + "credit_map": [ + 0.01727023907005787, + 0.02018097974359989, + 0.012671232223510742, + 0.009110096842050552, + 0.015226548537611961, + 0.026532605290412903, + 0.006670632399618626, + 0.007940397597849369, + 0.011641024611890316, + 0.012794041074812412, + 0.008083050139248371, + 0.005346299614757299, + 0.00855743046849966, + 0.008038321509957314, + 0.004176659509539604, + 0.006701183971017599, + 0.017659692093729973, + 0.006627819035202265, + 0.017604246735572815, + 0.0048484536819159985, + 0.006455908063799143, + 0.006563480477780104, + 0.006538369692862034, + 0.012236781418323517, + 0.008860738016664982, + 0.005360057577490807, + 0.013291982933878899, + 0.004084417596459389, + 0.006728375796228647, + 0.007551651448011398, + 0.0053455280140042305, + 0.007312198635190725, + 0.005343571305274963, + 0.00363745866343379, + 0.004347924143075943, + 0.00928545743227005, + 0.003229535883292556, + 0.003978047054260969, + 0.00855943188071251, + 0.007781695574522018, + 0.014802510850131512, + 0.009889193810522556, + 0.03167238086462021, + 0.025649376213550568, + 0.012635842896997929, + 0.01463018637150526, + 0.026643313467502594, + 0.020636556670069695, + 0.009908645413815975, + 0.015188831835985184, + 0.026505842804908752, + 0.017319565638899803, + 0.01775955781340599, + 0.013324753381311893, + 0.01580941490828991, + 0.016158288344740868, + 0.005463498178869486, + 0.009413797408342361, + 0.006896344944834709, + 0.010304685682058334, + 0.007539736572653055, + 0.010785043239593506, + 0.011736315675079823, + 0.01346268504858017, + 0.021413510665297508, + 0.028445929288864136, + 0.014690075069665909, + 0.02306763082742691, + 0.015089682303369045, + 0.022276898846030235, + 0.014175274409353733, + 0.019407683983445168, + 0.0074248905293643475, + 0.02300949953496456, + 0.009405750781297684, + 0.010089640505611897, + 0.01694992370903492, + 0.016454651951789856, + 0.01822483353316784, + 0.011564220301806927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.007742987480014563 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.01256160531193018 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.01256160531193018, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256160531193018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15372146666049957, + 0.10464822500944138, + 0.06647979468107224, + 0.12914571166038513, + 0.0954430028796196, + 0.048223789781332016, + 0.07971982657909393, + 0.16507990658283234, + 0.07021072506904602, + 0.08732759952545166 + ], + "credit_map": [ + 0.024462483823299408, + 0.03214951977133751, + 0.019321314990520477, + 0.010504232719540596, + 0.023153584450483322, + 0.017205309122800827, + 0.016708990558981895, + 0.010216038674116135, + 0.01845630444586277, + 0.023829998448491096, + 0.017681999132037163, + 0.012394590303301811, + 0.011332719586789608, + 0.0050794510170817375, + 0.009972186759114265, + 0.005900983698666096, + 0.010449384339153767, + 0.005684129428118467, + 0.009749230928719044, + 0.006363820284605026, + 0.007261927239596844, + 0.013427689671516418, + 0.006205333396792412, + 0.007338280789554119, + 0.01957656443119049, + 0.009664244018495083, + 0.025107163935899734, + 0.010528693906962872, + 0.01239260844886303, + 0.02203652635216713, + 0.012823873199522495, + 0.017016028985381126, + 0.011234462261199951, + 0.005232910625636578, + 0.008585186675190926, + 0.017359696328639984, + 0.01002426166087389, + 0.010367177426815033, + 0.019938141107559204, + 0.012701169587671757, + 0.005170351359993219, + 0.004043207969516516, + 0.008103026077151299, + 0.008215134032070637, + 0.003823487553745508, + 0.0040315017104148865, + 0.009014664217829704, + 0.0058224196545779705, + 0.004896130878478289, + 0.006608583033084869, + 0.011817123740911484, + 0.01309062261134386, + 0.008188795298337936, + 0.008029082790017128, + 0.014479650184512138, + 0.012609842233359814, + 0.011805690824985504, + 0.016948971897363663, + 0.015361877158284187, + 0.02235809899866581, + 0.02153410017490387, + 0.024715308099985123, + 0.02530735172331333, + 0.0270485058426857, + 0.00877415295690298, + 0.011792677454650402, + 0.0078909732401371, + 0.010375142097473145, + 0.005744061432778835, + 0.010603504255414009, + 0.007285665720701218, + 0.007744550239294767, + 0.005228014662861824, + 0.011360183358192444, + 0.009195991791784763, + 0.008342714048922062, + 0.013994055800139904, + 0.01262539904564619, + 0.018706994131207466, + 0.007874253205955029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.01256160531193018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.07900167256593704 + ], + "opposite_logprob_tokens": [ + -2.5790016651153564 + ], + "logprob_reward": -0.07900167256593704, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900167256593704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11102842539548874, + 0.04856513440608978, + 0.03332109376788139, + 0.18224285542964935, + 0.05315073952078819, + 0.1002621054649353, + 0.2234717309474945, + 0.06692559272050858, + 0.11563801020383835, + 0.06539425998926163 + ], + "credit_map": [ + 0.017791273072361946, + 0.023382501676678658, + 0.011856756173074245, + 0.01081389281898737, + 0.014581987634301186, + 0.011249000206589699, + 0.012035509571433067, + 0.009317511692643166, + 0.007672187872231007, + 0.007717118598520756, + 0.00727513711899519, + 0.0060583059675991535, + 0.0065068891271948814, + 0.005580598954111338, + 0.0036997185088694096, + 0.00405518151819706, + 0.0048683034256100655, + 0.004603092558681965, + 0.005099295638501644, + 0.002689789282158017, + 0.0037213321775197983, + 0.005841100122779608, + 0.003446796676144004, + 0.003051384352147579, + 0.016852255910634995, + 0.027296066284179688, + 0.018750948831439018, + 0.01739528402686119, + 0.0289257001131773, + 0.03164114058017731, + 0.01725195161998272, + 0.02412952296435833, + 0.004306409973651171, + 0.00426456006243825, + 0.004941402468830347, + 0.00915173627436161, + 0.005951296538114548, + 0.0068225073628127575, + 0.010167957283556461, + 0.0075448742136359215, + 0.010301141068339348, + 0.006895368453115225, + 0.016414061188697815, + 0.015040833503007889, + 0.010343040339648724, + 0.008966072462499142, + 0.0205383338034153, + 0.011763266287744045, + 0.014306585304439068, + 0.018636783584952354, + 0.03831182047724724, + 0.036310888826847076, + 0.0309903547167778, + 0.02443733997642994, + 0.03244168311357498, + 0.028036300092935562, + 0.005288559477776289, + 0.007072583306580782, + 0.010059180669486523, + 0.008112654089927673, + 0.007266984321177006, + 0.006882355082780123, + 0.011751865968108177, + 0.010491414926946163, + 0.013972560875117779, + 0.013354971073567867, + 0.016694383695721626, + 0.01620848849415779, + 0.009877538308501244, + 0.018097158521413803, + 0.0126657634973526, + 0.014767156913876534, + 0.004018362611532211, + 0.006950451526790857, + 0.0038730951491743326, + 0.005932742264121771, + 0.012532541528344154, + 0.009676693007349968, + 0.014234019443392754, + 0.008176358416676521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.07900167256593704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.04305243864655495 + ], + "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": [ + 0.07828367501497269, + 0.037817370146512985, + 0.05112089216709137, + 0.044018812477588654, + 0.0723331868648529, + 0.13232778012752533, + 0.32141631841659546, + 0.14775967597961426, + 0.05251067504286766, + 0.06241163611412048 + ], + "credit_map": [ + 0.011454878374934196, + 0.020358720794320107, + 0.007847473956644535, + 0.005823327228426933, + 0.01157684437930584, + 0.009245810098946095, + 0.00656790379434824, + 0.005408713594079018, + 0.005820739548653364, + 0.006747216917574406, + 0.004867859650403261, + 0.004164757672697306, + 0.005553493741899729, + 0.003704804927110672, + 0.0033548439387232065, + 0.003603655379265547, + 0.00991905853152275, + 0.005142934620380402, + 0.012958679348230362, + 0.003397606313228607, + 0.003665967844426632, + 0.005716949701309204, + 0.003539864206686616, + 0.0067798299714922905, + 0.0060803755186498165, + 0.005811050068587065, + 0.006819566711783409, + 0.003998527303338051, + 0.005785183981060982, + 0.005595133174210787, + 0.0040562693029642105, + 0.005872706882655621, + 0.008967378176748753, + 0.00613529235124588, + 0.006573510356247425, + 0.010323201306164265, + 0.006431578658521175, + 0.008479349315166473, + 0.01500531192868948, + 0.010417572222650051, + 0.014277354814112186, + 0.011718169786036015, + 0.023396719247102737, + 0.020059403032064438, + 0.012596656568348408, + 0.011402909643948078, + 0.023041198030114174, + 0.015835367143154144, + 0.02017250843346119, + 0.03383200615644455, + 0.06188470125198364, + 0.044979751110076904, + 0.04260359704494476, + 0.036560963839292526, + 0.043388474732637405, + 0.03799433261156082, + 0.01123796682804823, + 0.016527846455574036, + 0.017004117369651794, + 0.01959623210132122, + 0.017357783392071724, + 0.020798953250050545, + 0.024140886962413788, + 0.02109588496387005, + 0.0065909335389733315, + 0.005965226795524359, + 0.007305517327040434, + 0.006015858147293329, + 0.005460732616484165, + 0.008914710022509098, + 0.004738100804388523, + 0.007519593462347984, + 0.006233043037354946, + 0.011258309707045555, + 0.004362020175904036, + 0.004651928320527077, + 0.010208941996097565, + 0.010446153581142426, + 0.008571858517825603, + 0.006679384969174862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.04305243864655495 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.006746372207999229 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746372207999229, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746372207999229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09567632526159286, + 0.03540617600083351, + 0.04669483006000519, + 0.05135687068104744, + 0.03750674054026604, + 0.030599651858210564, + 0.1784215122461319, + 0.2239214926958084, + 0.15422147512435913, + 0.1461949497461319 + ], + "credit_map": [ + 0.01555674895644188, + 0.024305634200572968, + 0.009449077770113945, + 0.008494761772453785, + 0.014037009328603745, + 0.00901820044964552, + 0.008090510033071041, + 0.006724374834448099, + 0.005292105488479137, + 0.006921723950654268, + 0.004278002772480249, + 0.004553664475679398, + 0.004889707081019878, + 0.003336029127240181, + 0.0031115517485886812, + 0.0030233871657401323, + 0.008181743323802948, + 0.00689675472676754, + 0.0067714485339820385, + 0.004788251128047705, + 0.004320864100009203, + 0.005083282943814993, + 0.003565581049770117, + 0.007086900994181633, + 0.006229852791875601, + 0.008462038822472095, + 0.004776411224156618, + 0.006152456626296043, + 0.007182426284998655, + 0.005948068108409643, + 0.0061008380725979805, + 0.006504775956273079, + 0.003320785006508231, + 0.0034056296572089195, + 0.004272518213838339, + 0.009057453833520412, + 0.003248815890401602, + 0.0035335018765181303, + 0.004873861558735371, + 0.005794172640889883, + 0.004296584986150265, + 0.0025441597681492567, + 0.002820884110406041, + 0.0037702922709286213, + 0.00396052747964859, + 0.0031346518080681562, + 0.005804030224680901, + 0.004268517252057791, + 0.011969424784183502, + 0.018039239570498466, + 0.030981097370386124, + 0.024934343993663788, + 0.023271901533007622, + 0.022632814943790436, + 0.02494177594780922, + 0.021650897338986397, + 0.01979512721300125, + 0.023796087130904198, + 0.032402172684669495, + 0.031478822231292725, + 0.019640738144516945, + 0.022547781467437744, + 0.04275752231478691, + 0.03150321915745735, + 0.018123604357242584, + 0.018048033118247986, + 0.026605753228068352, + 0.029223598539829254, + 0.014434785582125187, + 0.01823527179658413, + 0.014752598479390144, + 0.014797818847000599, + 0.01129444595426321, + 0.026279740035533905, + 0.014351447112858295, + 0.011559850536286831, + 0.023143813014030457, + 0.021074136719107628, + 0.025925979018211365, + 0.01256551407277584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.006746372207999229 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level, medium shot frames a heavy, dark mahogany table in a dimly lit, classic study. The table's polished surface vaguely reflects the warm, concentrated glow of a green-shaded bank lamp, casting long, dancing shadows across the room. A stack of aged, off-white manuscripts rests placidly on the tabletop. Suddenly, an unseen force disrupts the stillness; the papers begin to flutter, then violently lift and swirl above the table, forming a chaotic vortex before dispersing and drifting onto the dark wooden floor. Simultaneously, the table's drawers, adorned with ornate, tarnished brass handles, begin to creak violently. In agitated and convulsive movements, they are flung open and then violently closed in a rapid, arrhythmic sequence, causing the entire sturdy furniture to tremble. The entire scene is imbued with a tense, supernatural atmosphere, the dramatic interplay of focused light and deep shadows enhancing the eerie, inexplicable animation of the inanimate objects.", + "seed": 23079, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_11/index_5512_seed_23079_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": -0.06645787693560123, + "vlm_logprob": -0.06645787693560123, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06645787693560123, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.020110052078962326 + ], + "opposite_logprob_tokens": [ + -5.145110130310059 + ], + "logprob_reward": -0.020110052078962326, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.020110052078962326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10117638111114502, + 0.0350545234978199, + 0.03555045649409294, + 0.018332988023757935, + 0.03891146928071976, + 0.055636800825595856, + 0.10063616931438446, + 0.07126902043819427, + 0.36403173208236694, + 0.1794004887342453 + ], + "credit_map": [ + 0.009590688161551952, + 0.015337317250669003, + 0.00862619187682867, + 0.008773391135036945, + 0.00974975060671568, + 0.019360151141881943, + 0.014567205682396889, + 0.015171692706644535, + 0.004892976023256779, + 0.0020678869914263487, + 0.003587254323065281, + 0.004293395671993494, + 0.0024453308433294296, + 0.006164758466184139, + 0.005960613954812288, + 0.005642307922244072, + 0.004000337794423103, + 0.003087306395173073, + 0.003953324630856514, + 0.004239988047629595, + 0.0030866600573062897, + 0.0055128238163888454, + 0.006487098056823015, + 0.005182918161153793, + 0.0024819078389555216, + 0.002068977104499936, + 0.002480751369148493, + 0.0017365268431603909, + 0.0024847486056387424, + 0.0019654398784041405, + 0.0017537445528432727, + 0.003360891481861472, + 0.004257515538483858, + 0.003539270954206586, + 0.006032062228769064, + 0.006503130309283733, + 0.004166023340076208, + 0.005453519523143768, + 0.004819463938474655, + 0.004140481352806091, + 0.008017700165510178, + 0.0048116594552993774, + 0.013133704662322998, + 0.007947954349219799, + 0.002527192234992981, + 0.0040391054935753345, + 0.007148429285734892, + 0.008011054247617722, + 0.007687137462198734, + 0.00714877899736166, + 0.01308087445795536, + 0.020299702882766724, + 0.005827036686241627, + 0.017636096104979515, + 0.016585156321525574, + 0.012371380813419819, + 0.006498508155345917, + 0.008380689658224583, + 0.009038274176418781, + 0.00873984582722187, + 0.010025468654930592, + 0.012292975559830666, + 0.007562480866909027, + 0.00873077753931284, + 0.017024118453264236, + 0.011530213057994843, + 0.008758914656937122, + 0.008556805551052094, + 0.027580952271819115, + 0.017777882516384125, + 0.14156895875930786, + 0.13123391568660736, + 0.02078654058277607, + 0.02153376117348671, + 0.017580222338438034, + 0.02404581755399704, + 0.01536974310874939, + 0.019916942343115807, + 0.032699067145586014, + 0.027468392625451088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020110052078962326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.006767803803086281 + ], + "opposite_logprob_tokens": [ + -5.006767749786377 + ], + "logprob_reward": -0.006767803803086281, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006767803803086281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09866846352815628, + 0.04423350468277931, + 0.07033082842826843, + 0.13740791380405426, + 0.09720585495233536, + 0.02716379053890705, + 0.0723423957824707, + 0.05455748736858368, + 0.10019324719905853, + 0.2978965640068054 + ], + "credit_map": [ + 0.008755521848797798, + 0.013076391071081161, + 0.008972370065748692, + 0.013182834722101688, + 0.011970236897468567, + 0.0128419678658247, + 0.016104400157928467, + 0.013764740899205208, + 0.006975427735596895, + 0.0030770613811910152, + 0.006685253232717514, + 0.0036266352981328964, + 0.002777600195258856, + 0.006975478492677212, + 0.006729866378009319, + 0.007386178243905306, + 0.008508196100592613, + 0.011953622102737427, + 0.008379514329135418, + 0.006497013382613659, + 0.0063610072247684, + 0.01053524948656559, + 0.008060123771429062, + 0.010036095045506954, + 0.017359070479869843, + 0.015778787434101105, + 0.015475546009838581, + 0.01527484878897667, + 0.01744915544986725, + 0.016907215118408203, + 0.017288073897361755, + 0.021875198930501938, + 0.010657825507223606, + 0.008564963936805725, + 0.014751576818525791, + 0.013176229782402515, + 0.009628823958337307, + 0.01588498055934906, + 0.01228321623057127, + 0.012258230708539486, + 0.004742682911455631, + 0.002102104015648365, + 0.005286959931254387, + 0.0038660874124616385, + 0.002530061174184084, + 0.002403977559879422, + 0.0033549321815371513, + 0.002876984653994441, + 0.0066798655316233635, + 0.004800653085112572, + 0.010025192983448505, + 0.013598877005279064, + 0.004919071216136217, + 0.011145297437906265, + 0.010621542111039162, + 0.010551887564361095, + 0.005583733785897493, + 0.009107613936066628, + 0.00863046757876873, + 0.005795938428491354, + 0.005936249624937773, + 0.005395574029535055, + 0.006456977222114801, + 0.007650929968804121, + 0.007858998142182827, + 0.0059842076152563095, + 0.006485290359705687, + 0.004624924156814814, + 0.01148110069334507, + 0.007252643350511789, + 0.029127372428774834, + 0.027378704398870468, + 0.03169049322605133, + 0.03262906149029732, + 0.04184172302484512, + 0.03682113066315651, + 0.026017852127552032, + 0.024473609402775764, + 0.054782938212156296, + 0.04963972419500351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.006767803803086281 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.07004260271787643 + ], + "opposite_logprob_tokens": [ + -2.695042610168457 + ], + "logprob_reward": -0.07004260271787643, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004260271787643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09256766736507416, + 0.06195157766342163, + 0.023685595020651817, + 0.09843413531780243, + 0.020081352442502975, + 0.03238806873559952, + 0.11555228382349014, + 0.21501564979553223, + 0.24114850163459778, + 0.09917516261339188 + ], + "credit_map": [ + 0.006655243691056967, + 0.012654011137783527, + 0.007582360878586769, + 0.013933642767369747, + 0.010971465148031712, + 0.01328289695084095, + 0.014844595454633236, + 0.012643448077142239, + 0.012685555033385754, + 0.0067421491257846355, + 0.008666268549859524, + 0.005665904842317104, + 0.00459321727976203, + 0.006692508235573769, + 0.007995441555976868, + 0.008910533040761948, + 0.0029286572244018316, + 0.003227179404348135, + 0.002793541643768549, + 0.0025179449003189802, + 0.001663867267780006, + 0.003159038256853819, + 0.004121784586459398, + 0.0032735811546444893, + 0.01606888137757778, + 0.00937391072511673, + 0.012032079510390759, + 0.010102974250912666, + 0.009522812440991402, + 0.011696845293045044, + 0.013856383040547371, + 0.015780258923768997, + 0.003533449023962021, + 0.002125163096934557, + 0.002837595297023654, + 0.0020834323950111866, + 0.0021789621096104383, + 0.0023115258663892746, + 0.0020357302855700254, + 0.002975492738187313, + 0.005281127989292145, + 0.0028855078853666782, + 0.0069960420951247215, + 0.004548023920506239, + 0.0019649574533104897, + 0.0027137065771967173, + 0.004017738159745932, + 0.0039809634909033775, + 0.01117967814207077, + 0.010688506998121738, + 0.015122191980481148, + 0.021888429298996925, + 0.007596852723509073, + 0.015167095698416233, + 0.0170687735080719, + 0.016840750351548195, + 0.0202924907207489, + 0.029973510652780533, + 0.0325126051902771, + 0.020692473277449608, + 0.02508465386927128, + 0.026317287236452103, + 0.029694924131035805, + 0.030447695404291153, + 0.011719170026481152, + 0.008807648904621601, + 0.00753505015745759, + 0.005678486544638872, + 0.018726609647274017, + 0.012291895225644112, + 0.0904267355799675, + 0.08596291393041611, + 0.00958763062953949, + 0.013111880049109459, + 0.008332418277859688, + 0.007402295712381601, + 0.009357607923448086, + 0.010912840254604816, + 0.01957075484097004, + 0.020899733528494835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.07004260271787643 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.22558008134365082 + ], + "opposite_logprob_tokens": [ + -1.600580096244812 + ], + "logprob_reward": -0.22558008134365082, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558008134365082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0542009174823761, + 0.031827330589294434, + 0.01880616508424282, + 0.012908843345940113, + 0.013816135935485363, + 0.02811545506119728, + 0.11878670752048492, + 0.042563311755657196, + 0.08004794269800186, + 0.5989271998405457 + ], + "credit_map": [ + 0.004344517830759287, + 0.007306382991373539, + 0.00487531116232276, + 0.005948219448328018, + 0.007320743519812822, + 0.007514484692364931, + 0.009462736546993256, + 0.007428524550050497, + 0.004273277707397938, + 0.001676619634963572, + 0.0035588322207331657, + 0.005282174330204725, + 0.0024278725031763315, + 0.003553468734025955, + 0.006191374734044075, + 0.004863712936639786, + 0.0022937750909477472, + 0.001595169655047357, + 0.00210908567532897, + 0.002209748374298215, + 0.0012692827731370926, + 0.002429273445159197, + 0.004161295481026173, + 0.0027385340072214603, + 0.0017920748796314, + 0.0010414678836241364, + 0.0013014387805014849, + 0.0014254628913477063, + 0.001491042785346508, + 0.00212930073030293, + 0.0016713523073121905, + 0.002056701807305217, + 0.0011292427079752088, + 0.0016230085166171193, + 0.002056257799267769, + 0.0018488249043002725, + 0.0015790850156918168, + 0.0017494386993348598, + 0.001639802474528551, + 0.002190476283431053, + 0.0028625603299587965, + 0.002749738050624728, + 0.00744624063372612, + 0.004762176889926195, + 0.0018518874421715736, + 0.002083073602989316, + 0.0036387683358043432, + 0.002721009310334921, + 0.011160004884004593, + 0.009146451018750668, + 0.016101829707622528, + 0.02169622667133808, + 0.00966192502528429, + 0.020249556750059128, + 0.01755012758076191, + 0.013220583088696003, + 0.00434389803558588, + 0.0046570501290261745, + 0.006464483682066202, + 0.0064367675222456455, + 0.004981794394552708, + 0.004252797923982143, + 0.006530392915010452, + 0.004896125756204128, + 0.00849855039268732, + 0.011902455240488052, + 0.006116657517850399, + 0.0055190520361065865, + 0.01855914294719696, + 0.010942860506474972, + 0.009081578813493252, + 0.00942764151841402, + 0.056765079498291016, + 0.07173080742359161, + 0.06561024487018585, + 0.08191967755556107, + 0.05495542287826538, + 0.062493424862623215, + 0.10715698450803757, + 0.09829553961753845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.22558008134365082 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, classic study featuring a heavy, dark mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a green-shaded bank lamp visible on the table casting warm light and long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do aged, off-white manuscripts initially rest on the table before being lifted into a chaotic vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the table drawers with ornate, tarnished brass handles open and close violently in a rapid, arrhythmic sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the table tremble and do the papers eventually drift onto the dark wooden 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": [ + -0.009788844734430313 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.009788844734430313, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788844734430313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06348639726638794, + 0.03355737030506134, + 0.021336747333407402, + 0.035662803798913956, + 0.03881257027387619, + 0.031191667541861534, + 0.07880405336618423, + 0.34724506735801697, + 0.28836292028427124, + 0.061540450900793076 + ], + "credit_map": [ + 0.004055338446050882, + 0.007647484540939331, + 0.005926483776420355, + 0.008138887584209442, + 0.007112671155482531, + 0.0070750233717262745, + 0.014571454375982285, + 0.008959047496318817, + 0.006236652843654156, + 0.0025057843886315823, + 0.0052400571294128895, + 0.004042410291731358, + 0.0022365832701325417, + 0.004493110813200474, + 0.003984730690717697, + 0.004818037152290344, + 0.0026770217809826136, + 0.001543925958685577, + 0.002924888161942363, + 0.002407371997833252, + 0.0015824974980205297, + 0.002842388581484556, + 0.004152680281549692, + 0.0032059713266789913, + 0.005935276858508587, + 0.0034012901596724987, + 0.003906548488885164, + 0.0035706334747374058, + 0.0038966466672718525, + 0.00517059862613678, + 0.004112169146537781, + 0.005669634323567152, + 0.004866066854447126, + 0.0035532887559384108, + 0.0060335309244692326, + 0.006212899461388588, + 0.0041878982447087765, + 0.0038935013581067324, + 0.006208849139511585, + 0.0038565329741686583, + 0.003717305837199092, + 0.002841693814843893, + 0.008295765146613121, + 0.004946371540427208, + 0.0015682991361245513, + 0.001726822811178863, + 0.0040651182644069195, + 0.004030289594084024, + 0.008442076854407787, + 0.006317012943327427, + 0.013360988348722458, + 0.010472637601196766, + 0.006045880261808634, + 0.011123371310532093, + 0.008765505626797676, + 0.014276574365794659, + 0.030789054930210114, + 0.05282120779156685, + 0.046447400003671646, + 0.03585583344101906, + 0.05215482413768768, + 0.0391697995364666, + 0.04224824905395508, + 0.04775865375995636, + 0.01388137973845005, + 0.007774501573294401, + 0.00883560162037611, + 0.008844591677188873, + 0.019487958401441574, + 0.011812981218099594, + 0.11262202262878418, + 0.10510385036468506, + 0.006522392854094505, + 0.007207813207060099, + 0.007034431677311659, + 0.005136669613420963, + 0.007705427706241608, + 0.00685442890971899, + 0.011999590322375298, + 0.009079692885279655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.009788844734430313 + } + ], + "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/0000058/group_12/index_tensor([9331, 9331])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_12/index_tensor([9331, 9331])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11faf0e406d3ea771ee11f7601669d687a332052 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_12/index_tensor([9331, 9331])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed, sports documentary-style, low-angle, full-body tracking shot follows several Fremen warriors as they race across an endless expanse of sun-baked sand dunes. One warrior in the foreground is captured in slow motion, clad in a meticulously designed, slightly worn, dark gray desert stillsuit, its tubes and filters clinging to his contours. As he leans forward, muscles tensing, and his right foot pushes back against the soft sand, the camera captures the explosive force of the moment with pinpoint clarity: the dune surface beneath his feet instantly collapses, then explodes backward, sending up a shower of visible sand grains that shimmer like crystal in the sunlight. His iconic eyes, shadowed by his mask, radiate a deep, luminous blue, their gaze focused and unwavering. The harsh sunlight casts sharp edges on their dark desert stillsuits, contrasting sharply with the soft curves of the dunes. Behind him, the figures of other soldiers are slightly blurred in the high-speed movement, together forming a dynamic line crossing the golden desert. The whole picture is full of primitive explosive power and physical tension. Against the warm background, the faint blue becomes the visual focus, perfectly freezing this decisive moment full of power and determination.", + "seed": 11083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_12/index_9331_seed_11083_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": -0.020700574293732642, + "vlm_logprob": -0.020700574293732642, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.020700574293732642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01972079463303089 + ], + "opposite_logprob_tokens": [ + -4.019721031188965 + ], + "logprob_reward": -0.01972079463303089, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01972079463303089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12107059359550476, + 0.123871810734272, + 0.10574574023485184, + 0.10694312304258347, + 0.04182661697268486, + 0.10606569796800613, + 0.08394302427768707, + 0.08255761861801147, + 0.032921336591243744, + 0.19505444169044495 + ], + "credit_map": [ + 0.019489720463752747, + 0.019672363996505737, + 0.018620576709508896, + 0.011467010714113712, + 0.008574867621064186, + 0.016138842329382896, + 0.011684320867061615, + 0.01542289275676012, + 0.014893022365868092, + 0.025527596473693848, + 0.012485211715102196, + 0.017025336623191833, + 0.010130930691957474, + 0.013665549457073212, + 0.016430005431175232, + 0.013714165426790714, + 0.01703045144677162, + 0.016786692664027214, + 0.011635050177574158, + 0.01239009015262127, + 0.009244290180504322, + 0.011961721815168858, + 0.011320759542286396, + 0.015376683324575424, + 0.014796164818108082, + 0.008793381974101067, + 0.014422914013266563, + 0.012963533401489258, + 0.007060837931931019, + 0.017058609053492546, + 0.016448643058538437, + 0.015399038791656494, + 0.004996152129024267, + 0.0043251607567071915, + 0.004842917900532484, + 0.0054047973826527596, + 0.0028649955056607723, + 0.0060549546033144, + 0.0053938296623528, + 0.007943808101117611, + 0.006993195507675409, + 0.007563145365566015, + 0.011493237689137459, + 0.007363761775195599, + 0.027703024446964264, + 0.028232570737600327, + 0.00755105447024107, + 0.009165712632238865, + 0.009509468451142311, + 0.008905134163796902, + 0.0071441433392465115, + 0.007818752899765968, + 0.01168391015380621, + 0.017005911096930504, + 0.008075296878814697, + 0.013800404034554958, + 0.018959561362862587, + 0.013949400745332241, + 0.011114896275103092, + 0.010848776437342167, + 0.006587376352399588, + 0.0064719789661467075, + 0.005628923419862986, + 0.008996705524623394, + 0.002778945490717888, + 0.005993206519633532, + 0.004950822331011295, + 0.004475741647183895, + 0.005924880038946867, + 0.002789765829220414, + 0.0028660628013312817, + 0.003141912864521146, + 0.024724995717406273, + 0.03643210232257843, + 0.033820223063230515, + 0.019815148785710335, + 0.0336742028594017, + 0.01358974538743496, + 0.02027004212141037, + 0.01272797305136919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.01972079463303089 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647167891263962 + ], + "opposite_logprob_tokens": [ + -4.8826470375061035 + ], + "logprob_reward": -0.007647167891263962, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647167891263962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05471029877662659, + 0.03134319558739662, + 0.026492230594158173, + 0.030226267874240875, + 0.11285942047834396, + 0.04971274733543396, + 0.06205476447939873, + 0.02579953335225582, + 0.07802250236272812, + 0.5287790894508362 + ], + "credit_map": [ + 0.007884003221988678, + 0.006791599094867706, + 0.008349517360329628, + 0.0052712406031787395, + 0.0035628224723041058, + 0.00987275317311287, + 0.005626361817121506, + 0.007351993583142757, + 0.0035461774095892906, + 0.006589945871382952, + 0.0043576001189649105, + 0.004043957684189081, + 0.002435357542708516, + 0.003267174120992422, + 0.0038677656557410955, + 0.003235212527215481, + 0.0027594007551670074, + 0.004744091536849737, + 0.0030082259327173233, + 0.003946691285818815, + 0.003265016945078969, + 0.0022155216429382563, + 0.003296536859124899, + 0.003256743075326085, + 0.002789626130834222, + 0.0037576567847281694, + 0.003550264285877347, + 0.003942716866731644, + 0.00229402887634933, + 0.00591188995167613, + 0.004051357973366976, + 0.003928727470338345, + 0.009727288968861103, + 0.018643096089363098, + 0.013414157554507256, + 0.013133998028934002, + 0.009927640669047832, + 0.013360299170017242, + 0.011827305890619755, + 0.02282561920583248, + 0.0030145139899104834, + 0.002724721562117338, + 0.003167566377669573, + 0.002324871253222227, + 0.01142445020377636, + 0.014182058162987232, + 0.0028116684406995773, + 0.010062888264656067, + 0.007059345953166485, + 0.010670101270079613, + 0.0076665873639285564, + 0.007244095206260681, + 0.0076543898321688175, + 0.008901880122721195, + 0.004439069423824549, + 0.008419289253652096, + 0.004214146174490452, + 0.0032818999607115984, + 0.0036484573502093554, + 0.003145481226965785, + 0.002654057927429676, + 0.002665033796802163, + 0.003408062970265746, + 0.0027823897544294596, + 0.006837818305939436, + 0.015385013073682785, + 0.011894243769347668, + 0.010489829815924168, + 0.013607720844447613, + 0.006263704504817724, + 0.006341701373457909, + 0.007202464621514082, + 0.06754186749458313, + 0.09495926648378372, + 0.09336618334054947, + 0.05463048070669174, + 0.0927802175283432, + 0.03600678592920303, + 0.054946087300777435, + 0.03454817086458206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.007647167891263962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306454211473465 + ], + "opposite_logprob_tokens": [ + -3.168064594268799 + ], + "logprob_reward": -0.04306454211473465, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306454211473465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.007227053865790367, + 0.0039251395501196384, + 0.003568417625501752, + 0.006160120479762554, + 0.00775684230029583, + 0.007870921865105629, + 0.014257223345339298, + 0.015717243775725365, + 0.016394607722759247, + 0.917122483253479 + ], + "credit_map": [ + 0.0007663460564799607, + 0.001164889195933938, + 0.001123653957620263, + 0.0008307990501634777, + 0.0005833940231241286, + 0.0011948406463488936, + 0.0005645803175866604, + 0.000998550676740706, + 0.0003570489934645593, + 0.0005438089719973505, + 0.0007334445836022496, + 0.0005318867042660713, + 0.0003278096264693886, + 0.0004541330272331834, + 0.0005270702531561255, + 0.00044993680785410106, + 0.0004674979718402028, + 0.0006137709133327007, + 0.000428078870754689, + 0.0003884659963659942, + 0.00033748988062143326, + 0.00047280709259212017, + 0.000369535613572225, + 0.0004907710826955736, + 0.0004229528713040054, + 0.0006149077089503407, + 0.0008293145219795406, + 0.0009684254182502627, + 0.00047972387983463705, + 0.0012415664969012141, + 0.0007516107289120555, + 0.0008516190573573112, + 0.000888309848960489, + 0.0007243863074108958, + 0.0009697289788164198, + 0.001186674227938056, + 0.0007442058413289487, + 0.0010162473190575838, + 0.0006710425950586796, + 0.001556247123517096, + 0.00046929772361181676, + 0.0005794854951091111, + 0.000762932060752064, + 0.00042728439439088106, + 0.0015091071836650372, + 0.0018872179789468646, + 0.0006588047835975885, + 0.001576792448759079, + 0.0016898417379707098, + 0.0031048175878822803, + 0.0016192018520087004, + 0.0012189737753942609, + 0.0011868588626384735, + 0.0019230501493439078, + 0.000900939863640815, + 0.002613539807498455, + 0.0029197486583143473, + 0.0032628674525767565, + 0.001962312962859869, + 0.001804968691430986, + 0.0014866420533508062, + 0.0014273928245529532, + 0.0012114397250115871, + 0.0016418708255514503, + 0.001362436800263822, + 0.0029677359852939844, + 0.0026968996971845627, + 0.0021962709724903107, + 0.002788779092952609, + 0.0014188650529831648, + 0.0015969923697412014, + 0.0013666287995874882, + 0.11672573536634445, + 0.17080074548721313, + 0.15955738723278046, + 0.09187695384025574, + 0.15830494463443756, + 0.06262387335300446, + 0.09807087481021881, + 0.0591619499027729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.04306454211473465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04398823529481888, + 0.04783613979816437, + 0.03142232075333595, + 0.06046368554234505, + 0.41088372468948364, + 0.1954490691423416, + 0.051557160913944244, + 0.024980755522847176, + 0.039877213537693024, + 0.09354164451360703 + ], + "credit_map": [ + 0.004790840670466423, + 0.007458097767084837, + 0.005793447140604258, + 0.0037973017897456884, + 0.003968124743551016, + 0.007888219319283962, + 0.0038416509050875902, + 0.006450555752962828, + 0.006389311980456114, + 0.007772122509777546, + 0.006473100278526545, + 0.00625517638400197, + 0.004451820161193609, + 0.005032457411289215, + 0.006220256444066763, + 0.00524189742282033, + 0.0031718327663838863, + 0.003140147076919675, + 0.002878345549106598, + 0.003023633500561118, + 0.0028126551769673824, + 0.006561140064150095, + 0.004510034807026386, + 0.005324534606188536, + 0.006842075381428003, + 0.006505797617137432, + 0.006358811166137457, + 0.007967964746057987, + 0.004864048212766647, + 0.00864635594189167, + 0.008408033289015293, + 0.01087060198187828, + 0.025601033121347427, + 0.02803177572786808, + 0.03478380665183067, + 0.06789355725049973, + 0.03645238280296326, + 0.056780848652124405, + 0.03792402893304825, + 0.12341631203889847, + 0.013306600041687489, + 0.01570603996515274, + 0.015296741388738155, + 0.010024757124483585, + 0.038243867456912994, + 0.046625953167676926, + 0.013483926653862, + 0.04276120662689209, + 0.005014451686292887, + 0.010178840719163418, + 0.006416977848857641, + 0.004350030794739723, + 0.004972060676664114, + 0.00691956328228116, + 0.00481780618429184, + 0.008887432515621185, + 0.004492527339607477, + 0.004378193989396095, + 0.0030060280114412308, + 0.002838574815541506, + 0.002169891493394971, + 0.00205686641857028, + 0.0017130166525021195, + 0.0043256571516394615, + 0.0032852923031896353, + 0.006038167513906956, + 0.0051699611358344555, + 0.005429914686828852, + 0.007034484297037125, + 0.004650729708373547, + 0.003382735187187791, + 0.0048859310336411, + 0.01204531081020832, + 0.019667867571115494, + 0.015648173168301582, + 0.00948924757540226, + 0.01307998038828373, + 0.0063516185618937016, + 0.011514210142195225, + 0.005745238624513149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.023292500525712967 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777866303920746 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.009777866303920746, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777866303920746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050562553107738495, + 0.07101540267467499, + 0.04948185384273529, + 0.061258960515260696, + 0.08788219839334488, + 0.07845968753099442, + 0.06387187540531158, + 0.058381613343954086, + 0.04465976357460022, + 0.43442606925964355 + ], + "credit_map": [ + 0.008090057410299778, + 0.007444527465850115, + 0.006288568954914808, + 0.0040720244869589806, + 0.0037890200037509203, + 0.008988991379737854, + 0.004971956368535757, + 0.006917406804859638, + 0.00562745425850153, + 0.013233008794486523, + 0.010491782799363136, + 0.009929136373102665, + 0.007000692654401064, + 0.007473225239664316, + 0.009636685252189636, + 0.007623425684869289, + 0.008469704538583755, + 0.005705362185835838, + 0.006759061943739653, + 0.004638704936951399, + 0.004233375657349825, + 0.006544507574290037, + 0.008167514577507973, + 0.004963623825460672, + 0.010203552432358265, + 0.009630589745938778, + 0.007135071326047182, + 0.0069170319475233555, + 0.0034617853816598654, + 0.007977473549544811, + 0.005795048549771309, + 0.010138406418263912, + 0.013993021100759506, + 0.009812915697693825, + 0.01286072377115488, + 0.010639427229762077, + 0.007990635000169277, + 0.012519894167780876, + 0.00802602432668209, + 0.012039562687277794, + 0.004982351791113615, + 0.0050977193750441074, + 0.009227580390870571, + 0.005205856636166573, + 0.019559776410460472, + 0.02029646560549736, + 0.006166990380734205, + 0.007922950200736523, + 0.008003195747733116, + 0.014154795557260513, + 0.006070490926504135, + 0.00630920147523284, + 0.006281367503106594, + 0.009693547151982784, + 0.004503818694502115, + 0.008855460211634636, + 0.005808286368846893, + 0.006266515702009201, + 0.007822619751095772, + 0.0073544918559491634, + 0.005588543135672808, + 0.005834387615323067, + 0.008659868501126766, + 0.011046898551285267, + 0.004913222510367632, + 0.006710620131343603, + 0.00549934571608901, + 0.007012521848082542, + 0.005761056672781706, + 0.0053289965726435184, + 0.004535971209406853, + 0.004898030776530504, + 0.05103537440299988, + 0.09454303979873657, + 0.07019489258527756, + 0.04284343495965004, + 0.070900559425354, + 0.028896261006593704, + 0.05063363164663315, + 0.025378884747624397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.009777866303920746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, sports documentary-style, low-angle, full-body tracking shot follows several Fremen warriors as they race across an endless expanse of sun-baked sand dunes. One warrior in the foreground is captured in slow motion, clad in a meticulously designed, slightly worn, dark gray desert stillsuit, its tubes and filters clinging to his contours. As he leans forward, muscles tensing, and his right foot pushes back against the soft sand, the camera captures the explosive force of the moment with pinpoint clarity: the dune surface beneath his feet instantly collapses, then explodes backward, sending up a shower of visible sand grains that shimmer like crystal in the sunlight. His iconic eyes, shadowed by his mask, radiate a deep, luminous blue, their gaze focused and unwavering. The harsh sunlight casts sharp edges on their dark desert stillsuits, contrasting sharply with the soft curves of the dunes. Behind him, the figures of other soldiers are slightly blurred in the high-speed movement, together forming a dynamic line crossing the golden desert. The whole picture is full of primitive explosive power and physical tension. Against the warm background, the faint blue becomes the visual focus, perfectly freezing this decisive moment full of power and determination.", + "seed": 11084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_12/index_9331_seed_11084_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": -0.021002269722521304, + "vlm_logprob": -0.021002269722521304, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.021002269722521304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012803022749722004 + ], + "opposite_logprob_tokens": [ + -4.512803077697754 + ], + "logprob_reward": -0.012803022749722004, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012803022749722004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5237114429473877, + 0.15171588957309723, + 0.03242246061563492, + 0.024274427443742752, + 0.06372883170843124, + 0.03578599542379379, + 0.020642120391130447, + 0.028386173769831657, + 0.04841095954179764, + 0.07092174887657166 + ], + "credit_map": [ + 0.06604650616645813, + 0.049376893788576126, + 0.052219208329916, + 0.12462460249662399, + 0.06636153906583786, + 0.04087063670158386, + 0.0322435237467289, + 0.09196852147579193, + 0.023641688749194145, + 0.023070093244314194, + 0.014736717566847801, + 0.01042980421334505, + 0.034752875566482544, + 0.015144865028560162, + 0.02076999843120575, + 0.009169832803308964, + 0.004754772409796715, + 0.004778913222253323, + 0.0036425460129976273, + 0.00429502734914422, + 0.003614806104451418, + 0.004658562131226063, + 0.0038077409844845533, + 0.002870094496756792, + 0.003634769469499588, + 0.003764689201489091, + 0.003261533798649907, + 0.002803673269227147, + 0.0028057212475687265, + 0.003083067014813423, + 0.002368970774114132, + 0.00255200220271945, + 0.01144040934741497, + 0.0076833064667880535, + 0.010003685019910336, + 0.006593781057745218, + 0.009581583552062511, + 0.006934873294085264, + 0.0049743251875042915, + 0.0065168640576303005, + 0.003655374050140381, + 0.005978577304631472, + 0.004064561333507299, + 0.003031917382031679, + 0.005431786645203829, + 0.004377668257802725, + 0.003034789115190506, + 0.006211324129253626, + 0.0024958536960184574, + 0.003043087664991617, + 0.0030934687238186598, + 0.0027346338611096144, + 0.002292198594659567, + 0.0022139614447951317, + 0.0019657479133456945, + 0.0028031696565449238, + 0.002470310777425766, + 0.0040409862995147705, + 0.0044678752310574055, + 0.003602932905778289, + 0.004292528610676527, + 0.00450851721689105, + 0.002162800170481205, + 0.002840223954990506, + 0.007856895215809345, + 0.005236760247498751, + 0.008713730610907078, + 0.004645063541829586, + 0.003274941351264715, + 0.004078688099980354, + 0.00617581233382225, + 0.008429065346717834, + 0.00565218273550272, + 0.014659792184829712, + 0.005936784669756889, + 0.017556888982653618, + 0.007029086351394653, + 0.007925985381007195, + 0.006368841975927353, + 0.005792188923805952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012803022749722004 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608080230653286 + ], + "opposite_logprob_tokens": [ + -4.141080856323242 + ], + "logprob_reward": -0.01608080230653286, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608080230653286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37215477228164673, + 0.07600626349449158, + 0.06669232994318008, + 0.04740706831216812, + 0.11651511490345001, + 0.11332598328590393, + 0.039241302758455276, + 0.06704118102788925, + 0.031738992780447006, + 0.06987691670656204 + ], + "credit_map": [ + 0.04714314639568329, + 0.027945172041654587, + 0.0340307280421257, + 0.08999001234769821, + 0.04001069813966751, + 0.032777633517980576, + 0.019893785938620567, + 0.08036364614963531, + 0.007845567539334297, + 0.018046986311674118, + 0.007917956449091434, + 0.007287176325917244, + 0.010859326459467411, + 0.009785751812160015, + 0.00994520727545023, + 0.0043183001689612865, + 0.007893119007349014, + 0.009499622508883476, + 0.007340441923588514, + 0.008436501957476139, + 0.007717367727309465, + 0.010003575123846531, + 0.007608094252645969, + 0.008193612098693848, + 0.0063203005120158195, + 0.007348509971052408, + 0.005375536624342203, + 0.00618160841986537, + 0.006541997194290161, + 0.0057127512991428375, + 0.005399879068136215, + 0.0045264908112585545, + 0.014740553684532642, + 0.01644202508032322, + 0.02002166584134102, + 0.009148810058832169, + 0.021729186177253723, + 0.015945978462696075, + 0.007245722226798534, + 0.011241180822253227, + 0.011357586830854416, + 0.02005138248205185, + 0.014133105054497719, + 0.0075537036173045635, + 0.02321448363363743, + 0.012763645499944687, + 0.011128658428788185, + 0.013123429380357265, + 0.003986210562288761, + 0.004158909432590008, + 0.007383381482213736, + 0.004659045487642288, + 0.004697870463132858, + 0.0041908505372703075, + 0.005009052809327841, + 0.005155985709279776, + 0.007736233528703451, + 0.012105632573366165, + 0.007951660081744194, + 0.0068745361641049385, + 0.014208241365849972, + 0.007743473630398512, + 0.00499693164601922, + 0.005424484144896269, + 0.003938674461096525, + 0.002412292640656233, + 0.004082844126969576, + 0.0041852849535644054, + 0.0036916565150022507, + 0.003203463042154908, + 0.004480718169361353, + 0.005744061898440123, + 0.0043243165127933025, + 0.019081778824329376, + 0.0042353468015789986, + 0.021926414221525192, + 0.004002729430794716, + 0.0054615638218820095, + 0.006463482044637203, + 0.0043813008815050125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.01608080230653286 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306522756814957 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306522756814957, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306522756814957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2632589638233185, + 0.11871936917304993, + 0.168195903301239, + 0.08060944080352783, + 0.18076658248901367, + 0.07618612796068192, + 0.023166829720139503, + 0.024257387965917587, + 0.046892959624528885, + 0.017946453765034676 + ], + "credit_map": [ + 0.03023296594619751, + 0.030944395810365677, + 0.02792574279010296, + 0.05925664305686951, + 0.025019055232405663, + 0.024688372388482094, + 0.015868330374360085, + 0.0493234246969223, + 0.012324041686952114, + 0.02871674671769142, + 0.010600708425045013, + 0.01116405613720417, + 0.020101457834243774, + 0.013942044228315353, + 0.014716565608978271, + 0.007153743412345648, + 0.022037457674741745, + 0.03865664452314377, + 0.015201697126030922, + 0.013100498355925083, + 0.021046120673418045, + 0.027614766731858253, + 0.014277334325015545, + 0.016261357814073563, + 0.010554133914411068, + 0.014074633829295635, + 0.01006638165563345, + 0.010609887540340424, + 0.010056938044726849, + 0.009315148927271366, + 0.009439494460821152, + 0.006492819637060165, + 0.030897045508027077, + 0.03536943346261978, + 0.020905498415231705, + 0.010631957091391087, + 0.03794817253947258, + 0.019604111090302467, + 0.01118784211575985, + 0.014222506433725357, + 0.007178825326263905, + 0.00998434703797102, + 0.010266182012856007, + 0.006303075700998306, + 0.02129032276570797, + 0.00772952800616622, + 0.00613378593698144, + 0.007300050929188728, + 0.002144145080819726, + 0.004415536765009165, + 0.004343239590525627, + 0.0024478936102241278, + 0.003220666665583849, + 0.002427441533654928, + 0.002245348645374179, + 0.0019225559663027525, + 0.0032465544063597918, + 0.0039865500293672085, + 0.0027703135274350643, + 0.0023542726412415504, + 0.005400523077696562, + 0.0027192782144993544, + 0.0018562192562967539, + 0.001923676230944693, + 0.004896773491054773, + 0.008736513555049896, + 0.003956063650548458, + 0.004520691931247711, + 0.0030245163943618536, + 0.004386804066598415, + 0.004287364427000284, + 0.01308422815054655, + 0.0014609675854444504, + 0.004106359556317329, + 0.001142551307566464, + 0.00550534063950181, + 0.001454864046536386, + 0.001533972448669374, + 0.0011901991674676538, + 0.001552198431454599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.04306522756814957 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286908864974976 + ], + "opposite_logprob_tokens": [ + -3.773286819458008 + ], + "logprob_reward": -0.023286908864974976, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286908864974976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2470996379852295, + 0.10341350734233856, + 0.14063389599323273, + 0.23633408546447754, + 0.06215706840157509, + 0.04238176345825195, + 0.03712204098701477, + 0.030628949403762817, + 0.0436696894466877, + 0.05655931681394577 + ], + "credit_map": [ + 0.025597862899303436, + 0.026544000953435898, + 0.026784410700201988, + 0.05642752721905708, + 0.022437071427702904, + 0.023577850311994553, + 0.016626529395580292, + 0.04910440370440483, + 0.012991991825401783, + 0.011791402474045753, + 0.017870843410491943, + 0.013425178825855255, + 0.017604319378733635, + 0.008942493237555027, + 0.012571322731673717, + 0.008215969428420067, + 0.019381649792194366, + 0.013387409038841724, + 0.021719984710216522, + 0.013806532137095928, + 0.01994480937719345, + 0.018750494346022606, + 0.021203547716140747, + 0.012439468875527382, + 0.030523819848895073, + 0.03887319564819336, + 0.030207784846425056, + 0.0278185922652483, + 0.029084110632538795, + 0.035946935415267944, + 0.026359442621469498, + 0.01752021349966526, + 0.014312836341559887, + 0.008838977664709091, + 0.0061344304122030735, + 0.005234797019511461, + 0.00889277271926403, + 0.009905794635415077, + 0.004913618788123131, + 0.003923846874386072, + 0.004427731037139893, + 0.005279230419546366, + 0.005237965378910303, + 0.003362644463777542, + 0.008485693484544754, + 0.004846043884754181, + 0.0031393016688525677, + 0.007603154983371496, + 0.0037525382358580828, + 0.0067323241382837296, + 0.007974346168339252, + 0.004692226182669401, + 0.004767272155731916, + 0.0036970183718949556, + 0.0029334838036447763, + 0.002572837518528104, + 0.002846880815923214, + 0.006232855375856161, + 0.0037072922568768263, + 0.003325440688058734, + 0.004676621872931719, + 0.004424924496561289, + 0.002260899869725108, + 0.0031540379859507084, + 0.0068405973725020885, + 0.0055757611989974976, + 0.006930241826921701, + 0.004310652147978544, + 0.0035770235117524862, + 0.004224966745823622, + 0.004289514850825071, + 0.00792093388736248, + 0.007324208039790392, + 0.00620416272431612, + 0.007998174987733364, + 0.008787296712398529, + 0.007143534719944, + 0.007582699880003929, + 0.006075665820389986, + 0.005443581845611334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.023286908864974976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977538712322712 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.00977538712322712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977538712322712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5577717423439026, + 0.026170795783400536, + 0.03351664915680885, + 0.028531936928629875, + 0.09439616650342941, + 0.06399320065975189, + 0.04015744850039482, + 0.018071334809064865, + 0.03430217504501343, + 0.10308857262134552 + ], + "credit_map": [ + 0.07022283971309662, + 0.03921016305685043, + 0.048099394887685776, + 0.1333877295255661, + 0.06061064451932907, + 0.04991648346185684, + 0.03039919026196003, + 0.12592530250549316, + 0.003975125961005688, + 0.0024376867804676294, + 0.0032832592260092497, + 0.003020355710759759, + 0.004119789227843285, + 0.0026271985843777657, + 0.0034533189609646797, + 0.003254063194617629, + 0.004094987642019987, + 0.00498091708868742, + 0.004231959115713835, + 0.0029964521527290344, + 0.004322707187384367, + 0.005075066350400448, + 0.00486562866717577, + 0.002948931185528636, + 0.003217063145712018, + 0.004565640352666378, + 0.0038883087690919638, + 0.0041545117273926735, + 0.003885536687448621, + 0.0031352450605481863, + 0.003541467012837529, + 0.002144167199730873, + 0.01267956756055355, + 0.014570011757314205, + 0.016506226733326912, + 0.00796523131430149, + 0.014935726299881935, + 0.01145688071846962, + 0.006385446060448885, + 0.009897072799503803, + 0.005761995911598206, + 0.009823130443692207, + 0.008819353766739368, + 0.005201409105211496, + 0.015335164032876492, + 0.006822370924055576, + 0.005793863907456398, + 0.006435917224735022, + 0.00393932918086648, + 0.008846713230013847, + 0.007390362676233053, + 0.004912182223051786, + 0.005192745942622423, + 0.0034734925720840693, + 0.0035079093649983406, + 0.0028947126120328903, + 0.0020377985201776028, + 0.0020609572529792786, + 0.002199846785515547, + 0.0030529808718711138, + 0.0024080295115709305, + 0.002650287002325058, + 0.0018692829180508852, + 0.0017921526450663805, + 0.004935555625706911, + 0.0034943879581987858, + 0.005598718300461769, + 0.0039057163521647453, + 0.0036950521171092987, + 0.0039239609614014626, + 0.005276129115372896, + 0.0034726541489362717, + 0.008801527321338654, + 0.025490734726190567, + 0.010213413275778294, + 0.02906683087348938, + 0.007771593984216452, + 0.0076451669447124004, + 0.008837244473397732, + 0.005262061953544617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.00977538712322712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, sports documentary-style, low-angle, full-body tracking shot follows several Fremen warriors as they race across an endless expanse of sun-baked sand dunes. One warrior in the foreground is captured in slow motion, clad in a meticulously designed, slightly worn, dark gray desert stillsuit, its tubes and filters clinging to his contours. As he leans forward, muscles tensing, and his right foot pushes back against the soft sand, the camera captures the explosive force of the moment with pinpoint clarity: the dune surface beneath his feet instantly collapses, then explodes backward, sending up a shower of visible sand grains that shimmer like crystal in the sunlight. His iconic eyes, shadowed by his mask, radiate a deep, luminous blue, their gaze focused and unwavering. The harsh sunlight casts sharp edges on their dark desert stillsuits, contrasting sharply with the soft curves of the dunes. Behind him, the figures of other soldiers are slightly blurred in the high-speed movement, together forming a dynamic line crossing the golden desert. The whole picture is full of primitive explosive power and physical tension. Against the warm background, the faint blue becomes the visual focus, perfectly freezing this decisive moment full of power and determination.", + "seed": 11085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_12/index_9331_seed_11085_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": -0.05080962479114533, + "vlm_logprob": -0.05080962479114533, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.05080962479114533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12805478274822235 + ], + "opposite_logprob_tokens": [ + -2.1280548572540283 + ], + "logprob_reward": -0.12805478274822235, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12805478274822235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09221378713846207, + 0.0294214878231287, + 0.06436067074537277, + 0.056235421448946, + 0.0383402556180954, + 0.4839145541191101, + 0.12170173972845078, + 0.0465361513197422, + 0.03978295996785164, + 0.02749297395348549 + ], + "credit_map": [ + 0.011668291874229908, + 0.010754004120826721, + 0.014481926336884499, + 0.010348649695515633, + 0.015497947111725807, + 0.010537412948906422, + 0.01114611979573965, + 0.007779438514262438, + 0.004857154563069344, + 0.002620918909087777, + 0.007239950355142355, + 0.0027343137189745903, + 0.003449981566518545, + 0.0030878812540322542, + 0.0033747393172234297, + 0.0020565458107739687, + 0.017626771703362465, + 0.0048210048116743565, + 0.018149210140109062, + 0.004668349400162697, + 0.005305931903421879, + 0.006292529404163361, + 0.004372072406113148, + 0.0031248016748577356, + 0.006701440550386906, + 0.008455240167677402, + 0.005609817337244749, + 0.015481619164347649, + 0.006021700333803892, + 0.0034671491011977196, + 0.008195793256163597, + 0.0023026582784950733, + 0.004380304832011461, + 0.005114837549626827, + 0.006051739677786827, + 0.003906152443960309, + 0.005096484441310167, + 0.006710481364279985, + 0.00393782090395689, + 0.0031424362678080797, + 0.13867445290088654, + 0.01423037238419056, + 0.017606908455491066, + 0.05722968652844429, + 0.022361597046256065, + 0.01637249067425728, + 0.19240108132362366, + 0.02503797970712185, + 0.02368132770061493, + 0.010679728351533413, + 0.0086505226790905, + 0.03048565238714218, + 0.013805431313812733, + 0.01443468313664198, + 0.011913166381418705, + 0.008051225915551186, + 0.0038640813436359167, + 0.006061793304979801, + 0.007898075506091118, + 0.00499936705455184, + 0.009983586147427559, + 0.004348161164671183, + 0.005344563163816929, + 0.004036521073430777, + 0.0028939805924892426, + 0.006663430016487837, + 0.00615675188601017, + 0.005634496919810772, + 0.006513387430459261, + 0.004584642127156258, + 0.0040772706270217896, + 0.003258998040109873, + 0.003197724698111415, + 0.0034953311551362276, + 0.004576453473418951, + 0.0028478149324655533, + 0.003928991034626961, + 0.003717106766998768, + 0.0028455089777708054, + 0.002884043613448739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.12805478274822235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087606549263 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087606549263, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087606549263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06096372753381729, + 0.09839440882205963, + 0.06299930810928345, + 0.09942147135734558, + 0.05391205847263336, + 0.2435360699892044, + 0.14958561956882477, + 0.12467747181653976, + 0.03552187606692314, + 0.07098797708749771 + ], + "credit_map": [ + 0.00820131879299879, + 0.010059550404548645, + 0.008233177475631237, + 0.008317426778376102, + 0.008837802335619926, + 0.006441925652325153, + 0.006311826407909393, + 0.00456070713698864, + 0.011844677850604057, + 0.016845805570483208, + 0.012943943962454796, + 0.012997765094041824, + 0.018625639379024506, + 0.010437361896038055, + 0.0077944970689713955, + 0.006904731038957834, + 0.018447967246174812, + 0.005352498032152653, + 0.019918564707040787, + 0.003570977132767439, + 0.004797148983925581, + 0.003420355496928096, + 0.004123967606574297, + 0.0033678431063890457, + 0.015071001835167408, + 0.01648196578025818, + 0.008601619862020016, + 0.013128053396940231, + 0.014099127613008022, + 0.0056230477057397366, + 0.02074095420539379, + 0.00567571260035038, + 0.005596194416284561, + 0.009990794584155083, + 0.006735322531312704, + 0.006055631209164858, + 0.010143031366169453, + 0.004638419486582279, + 0.007428357843309641, + 0.0033243109937757254, + 0.07040171325206757, + 0.007511784788221121, + 0.009393700398504734, + 0.02794365957379341, + 0.01180806290358305, + 0.008288967423141003, + 0.0955003947019577, + 0.012687806971371174, + 0.028305230662226677, + 0.013105164282023907, + 0.011699521914124489, + 0.03321469575166702, + 0.01732632890343666, + 0.02026446908712387, + 0.01698918081820011, + 0.008681033737957478, + 0.014662830159068108, + 0.014966151677072048, + 0.01897992566227913, + 0.031882453709840775, + 0.015756748616695404, + 0.008592039346694946, + 0.012808377854526043, + 0.007028956897556782, + 0.002655790885910392, + 0.00398872559890151, + 0.003334277542307973, + 0.003478779923170805, + 0.004559851251542568, + 0.003224889049306512, + 0.008350218646228313, + 0.0059293475933372974, + 0.006090313196182251, + 0.009077969007194042, + 0.009586035273969173, + 0.0074218339286744595, + 0.015200849622488022, + 0.00744105177000165, + 0.009330883622169495, + 0.006839049980044365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.016087606549263 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381123311817646 + ], + "opposite_logprob_tokens": [ + -3.2881124019622803 + ], + "logprob_reward": -0.0381123311817646, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381123311817646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12188940495252609, + 0.06514137238264084, + 0.04541958123445511, + 0.05727057904005051, + 0.07213962823152542, + 0.08772343397140503, + 0.08269717544317245, + 0.12052411586046219, + 0.06539983302354813, + 0.28179487586021423 + ], + "credit_map": [ + 0.020094415172934532, + 0.018382279202342033, + 0.02254190295934677, + 0.013671119697391987, + 0.014394881203770638, + 0.015645844861865044, + 0.008779422380030155, + 0.008379539474844933, + 0.010068641044199467, + 0.007733058650046587, + 0.014741060324013233, + 0.006645109970122576, + 0.011183241382241249, + 0.005591185763478279, + 0.005838475190103054, + 0.0033406016882508993, + 0.007141980342566967, + 0.006151205860078335, + 0.007125276606529951, + 0.009114346466958523, + 0.004652700386941433, + 0.005469852592796087, + 0.0026155724190175533, + 0.00314864469692111, + 0.010201815515756607, + 0.006099051330238581, + 0.005030721426010132, + 0.008046622388064861, + 0.0052655781619250774, + 0.004085490480065346, + 0.01426665484905243, + 0.004274646285921335, + 0.007427733391523361, + 0.009625229053199291, + 0.012259514071047306, + 0.006442577578127384, + 0.012297380715608597, + 0.01039775088429451, + 0.007622091099619865, + 0.006067345384508371, + 0.009378965944051743, + 0.012237164191901684, + 0.011266003362834454, + 0.010482841171324253, + 0.016718672588467598, + 0.009394032880663872, + 0.010939103551208973, + 0.007306646555662155, + 0.013697285205125809, + 0.007289724890142679, + 0.006916142534464598, + 0.02034149132668972, + 0.011529025621712208, + 0.009264604188501835, + 0.007883435115218163, + 0.005775465629994869, + 0.017214924097061157, + 0.010605813935399055, + 0.013600175268948078, + 0.029258541762828827, + 0.02026449516415596, + 0.008789444342255592, + 0.0139304893091321, + 0.0068602366372942924, + 0.004906904883682728, + 0.010258736088871956, + 0.007176140323281288, + 0.009698803536593914, + 0.01517455093562603, + 0.005688308738172054, + 0.006536716595292091, + 0.005959673784673214, + 0.023824794217944145, + 0.05776149779558182, + 0.028886185958981514, + 0.029800964519381523, + 0.06597655266523361, + 0.024603651836514473, + 0.025599466636776924, + 0.025341765955090523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0381123311817646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810602053999901 + ], + "opposite_logprob_tokens": [ + -3.2881059646606445 + ], + "logprob_reward": -0.03810602053999901, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810602053999901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2160302698612213, + 0.046491630375385284, + 0.07763299345970154, + 0.04380746930837631, + 0.06849738955497742, + 0.3567705452442169, + 0.05686822533607483, + 0.04871919006109238, + 0.03505425527691841, + 0.050128087401390076 + ], + "credit_map": [ + 0.03248303383588791, + 0.028421401977539062, + 0.05447946488857269, + 0.026506176218390465, + 0.02311125583946705, + 0.023529119789600372, + 0.015978526324033737, + 0.011521266773343086, + 0.007375810761004686, + 0.005824455060064793, + 0.009209833107888699, + 0.004624724388122559, + 0.005358217284083366, + 0.006066560745239258, + 0.00431149872019887, + 0.0037205228582024574, + 0.021030070260167122, + 0.004817926790565252, + 0.022853218019008636, + 0.012043892405927181, + 0.004911981988698244, + 0.0030996159184724092, + 0.005277344491332769, + 0.0035989328753203154, + 0.008451179601252079, + 0.0042123012244701385, + 0.0034330219496041536, + 0.00902928039431572, + 0.0034606163389980793, + 0.0027905020397156477, + 0.010286816395819187, + 0.0021437471732497215, + 0.006360842380672693, + 0.011797172017395496, + 0.008936750702559948, + 0.007059437222778797, + 0.01377721969038248, + 0.00782504491508007, + 0.007737156003713608, + 0.005003755446523428, + 0.10311014205217361, + 0.012002605944871902, + 0.015238910913467407, + 0.03993107005953789, + 0.016230635344982147, + 0.012723452411592007, + 0.13937370479106903, + 0.01815997064113617, + 0.003984345588833094, + 0.006042975466698408, + 0.006292646285146475, + 0.009473685175180435, + 0.012155132368206978, + 0.005787635687738657, + 0.007049710489809513, + 0.006082088220864534, + 0.004598397295922041, + 0.005342372693121433, + 0.007180590182542801, + 0.006989935878664255, + 0.010916125029325485, + 0.0044861468486487865, + 0.005792555864900351, + 0.003413059050217271, + 0.0025269370526075363, + 0.0046203951351344585, + 0.003839981509372592, + 0.003770420327782631, + 0.0046092309057712555, + 0.003453038865700364, + 0.00743044214323163, + 0.004803802818059921, + 0.005693033803254366, + 0.005899874027818441, + 0.0074582407251000404, + 0.004674120806157589, + 0.0068208035081624985, + 0.00540122902020812, + 0.00792447105050087, + 0.006256306543946266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.03810602053999901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.038480158895254135, + 0.018874220550060272, + 0.03584098070859909, + 0.026726162061095238, + 0.022912854328751564, + 0.6014824509620667, + 0.1554340422153473, + 0.02751207910478115, + 0.03129970654845238, + 0.04143737256526947 + ], + "credit_map": [ + 0.005624007899314165, + 0.004454623907804489, + 0.005502008367329836, + 0.0058047156780958176, + 0.006148944143205881, + 0.004329093731939793, + 0.004264492075890303, + 0.002352276351302862, + 0.0023514670319855213, + 0.0028045980725437403, + 0.0035795196890830994, + 0.001850207569077611, + 0.002021385356783867, + 0.002694808878004551, + 0.0019424170022830367, + 0.0016298152040690184, + 0.004423619247972965, + 0.005983687937259674, + 0.0037908190861344337, + 0.004654797725379467, + 0.0053548142313957214, + 0.005291457287967205, + 0.0033049210906028748, + 0.0030368627049028873, + 0.0024039107374846935, + 0.003966995049268007, + 0.0032802764326334, + 0.006098933517932892, + 0.003724161069840193, + 0.0027101633604615927, + 0.002588320057839155, + 0.00195340090431273, + 0.0027316163759678602, + 0.002993799978867173, + 0.004217179026454687, + 0.0023773969151079655, + 0.0026505575515329838, + 0.0032538350205868483, + 0.00213042926043272, + 0.0025580390356481075, + 0.17048902809619904, + 0.01734493114054203, + 0.0220771636813879, + 0.07197453826665878, + 0.028374234214425087, + 0.019764261320233345, + 0.2408958226442337, + 0.03056248649954796, + 0.02875504456460476, + 0.013052060268819332, + 0.010752582922577858, + 0.03715013340115547, + 0.020743487402796745, + 0.019842954352498055, + 0.015563476830720901, + 0.009574306197464466, + 0.002653365023434162, + 0.003914467990398407, + 0.00405013607814908, + 0.003352794563397765, + 0.005734363570809364, + 0.0024016988463699818, + 0.0033720226492732763, + 0.0020332306157797575, + 0.0027370559982955456, + 0.004673831630498171, + 0.0033147423528134823, + 0.003175858175382018, + 0.005412037018686533, + 0.004160057287663221, + 0.0039571416564285755, + 0.003868982195854187, + 0.005092420149594545, + 0.0044495961628854275, + 0.006092391908168793, + 0.004275225102901459, + 0.005298740696161985, + 0.005637243390083313, + 0.005704123992472887, + 0.004887633491307497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.03368738293647766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed, sports documentary-style, low-angle, full-body tracking shot follows several Fremen warriors as they race across an endless expanse of sun-baked sand dunes. One warrior in the foreground is captured in slow motion, clad in a meticulously designed, slightly worn, dark gray desert stillsuit, its tubes and filters clinging to his contours. As he leans forward, muscles tensing, and his right foot pushes back against the soft sand, the camera captures the explosive force of the moment with pinpoint clarity: the dune surface beneath his feet instantly collapses, then explodes backward, sending up a shower of visible sand grains that shimmer like crystal in the sunlight. His iconic eyes, shadowed by his mask, radiate a deep, luminous blue, their gaze focused and unwavering. The harsh sunlight casts sharp edges on their dark desert stillsuits, contrasting sharply with the soft curves of the dunes. Behind him, the figures of other soldiers are slightly blurred in the high-speed movement, together forming a dynamic line crossing the golden desert. The whole picture is full of primitive explosive power and physical tension. Against the warm background, the faint blue becomes the visual focus, perfectly freezing this decisive moment full of power and determination.", + "seed": 11086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_12/index_9331_seed_11086_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": -0.023620307072997095, + "vlm_logprob": -0.023620307072997095, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.023620307072997095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016991399228572845 + ], + "opposite_logprob_tokens": [ + -4.14199161529541 + ], + "logprob_reward": -0.016991399228572845, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016991399228572845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03389034420251846, + 0.03744380176067352, + 0.7526910901069641, + 0.009204406291246414, + 0.02069854736328125, + 0.04545506834983826, + 0.02431531809270382, + 0.033874742686748505, + 0.0101943239569664, + 0.032232362776994705 + ], + "credit_map": [ + 0.007014711387455463, + 0.008266759105026722, + 0.002991229761391878, + 0.002834995975717902, + 0.002777399029582739, + 0.0033864036668092012, + 0.003951892256736755, + 0.0026669513899832964, + 0.003330487059429288, + 0.0034912668634206057, + 0.004452803172171116, + 0.007378014270216227, + 0.004488922189921141, + 0.004831154365092516, + 0.005749714560806751, + 0.003721441375091672, + 0.050167713314294815, + 0.051792290061712265, + 0.22660617530345917, + 0.0941319465637207, + 0.047074127942323685, + 0.11685733497142792, + 0.035560328513383865, + 0.13050121068954468, + 0.0008608520147390664, + 0.0009374175569973886, + 0.001762972678989172, + 0.0010386018548160791, + 0.0011784830130636692, + 0.0013193940976634622, + 0.001100576831959188, + 0.0010061083594337106, + 0.0015897918492555618, + 0.003530851099640131, + 0.001774287549778819, + 0.0020962506532669067, + 0.002654117066413164, + 0.0037192623130977154, + 0.002728243824094534, + 0.0026057418435811996, + 0.0025574560277163982, + 0.003486979752779007, + 0.004778658505529165, + 0.003017190145328641, + 0.003629853716120124, + 0.013028503395617008, + 0.0031997314654290676, + 0.011756694875657558, + 0.003880186704918742, + 0.004448998253792524, + 0.002773541258648038, + 0.002659417688846588, + 0.0034010123927146196, + 0.0024634588044136763, + 0.0019976762123405933, + 0.0026910286396741867, + 0.0030069893691688776, + 0.0034131729044020176, + 0.003887865459546447, + 0.0047751846723258495, + 0.004750589840114117, + 0.0029669562354683876, + 0.00880235806107521, + 0.002271626377478242, + 0.0006721769459545612, + 0.002040307503193617, + 0.001543037942610681, + 0.0016453700372949243, + 0.0013847326627001166, + 0.0009235433535650373, + 0.0010029422119259834, + 0.000982212950475514, + 0.0025885358918458223, + 0.004288889933377504, + 0.002278269501402974, + 0.0042067705653607845, + 0.005873001646250486, + 0.0036259242333471775, + 0.0052641029469668865, + 0.004106868989765644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.016991399228572845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089714244008064 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089714244008064, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089714244008064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031872548162937164, + 0.007470407988876104, + 0.7566023468971252, + 0.005489064380526543, + 0.023072272539138794, + 0.03106009028851986, + 0.02207980304956436, + 0.027815623208880424, + 0.020847856998443604, + 0.07368995249271393 + ], + "credit_map": [ + 0.006619871128350496, + 0.006808911450207233, + 0.002872904995456338, + 0.0038881879299879074, + 0.0033605333883315325, + 0.0025014218408614397, + 0.00353516754694283, + 0.0022855496499687433, + 0.0006804061704315245, + 0.0006241821683943272, + 0.0010123791871592402, + 0.0011511652264744043, + 0.0007040103664621711, + 0.000937154982239008, + 0.0012252549640834332, + 0.0011358547490090132, + 0.0504494346678257, + 0.0515148900449276, + 0.2275836169719696, + 0.09524790197610855, + 0.0477396622300148, + 0.11588915437459946, + 0.03660479933023453, + 0.13157284259796143, + 0.0004290266369935125, + 0.0007478292682208121, + 0.0006509175291284919, + 0.0006357594393193722, + 0.0006876827101223171, + 0.0009906438644975424, + 0.0007213575881905854, + 0.0006258474313654006, + 0.0018448299961164594, + 0.0017300719628110528, + 0.003953325562179089, + 0.003842225531116128, + 0.0019882984925061464, + 0.003355384571477771, + 0.0032546729780733585, + 0.0031034641433507204, + 0.0016153487376868725, + 0.0024582359474152327, + 0.002509667072445154, + 0.002457187743857503, + 0.0026075660716742277, + 0.008174663409590721, + 0.002854553982615471, + 0.008382867090404034, + 0.001988160889595747, + 0.0034923262428492308, + 0.0034968883264809847, + 0.0025927089154720306, + 0.0023731105029582977, + 0.0028643335681408644, + 0.0022164250258356333, + 0.003055848879739642, + 0.003860717872157693, + 0.0029511163011193275, + 0.00241953507065773, + 0.0023524854332208633, + 0.0038992525078356266, + 0.0032035796903073788, + 0.004118766635656357, + 0.005010169465094805, + 0.0016273159999400377, + 0.0038256782572716475, + 0.003119301749393344, + 0.003286516759544611, + 0.0026662787422537804, + 0.002133136847987771, + 0.0021582278423011303, + 0.0020314014982432127, + 0.0039612045511603355, + 0.005693153478205204, + 0.024489544332027435, + 0.008333351463079453, + 0.008864413015544415, + 0.004462955053895712, + 0.007180389948189259, + 0.010704945772886276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.011089714244008064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062058594077825546 + ], + "opposite_logprob_tokens": [ + -2.812058687210083 + ], + "logprob_reward": -0.062058594077825546, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062058594077825546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12625637650489807, + 0.03257513791322708, + 0.2744940519332886, + 0.01251237839460373, + 0.05230871960520744, + 0.046668827533721924, + 0.08365482091903687, + 0.10515085607767105, + 0.02981721982359886, + 0.2365616112947464 + ], + "credit_map": [ + 0.032693278044462204, + 0.03411765396595001, + 0.011365704238414764, + 0.01128911692649126, + 0.01469760574400425, + 0.005762291140854359, + 0.010030020959675312, + 0.006300703156739473, + 0.0032052581664174795, + 0.0023906163405627012, + 0.004536931402981281, + 0.007012161426246166, + 0.002970244502648711, + 0.0032619533594697714, + 0.005702473688870668, + 0.0034954999573528767, + 0.018478998914361, + 0.019531525671482086, + 0.08085611462593079, + 0.033848077058792114, + 0.018661992624402046, + 0.04189489781856537, + 0.013977080583572388, + 0.047245386987924576, + 0.0010537431808188558, + 0.0015812594210729003, + 0.0014766243984922767, + 0.0015086757484823465, + 0.0015147841768339276, + 0.001677037333138287, + 0.0019614670891314745, + 0.0017387868138030171, + 0.0040555368177592754, + 0.005474803037941456, + 0.009172129444777966, + 0.00932699628174305, + 0.00478605180978775, + 0.006387701723724604, + 0.006262218579649925, + 0.00684328144416213, + 0.00366646284237504, + 0.005599116440862417, + 0.006445427890866995, + 0.0050807069055736065, + 0.004849912133067846, + 0.01003275066614151, + 0.005370356608182192, + 0.005624090787023306, + 0.007647930644452572, + 0.013523264788091183, + 0.012332068756222725, + 0.010578124783933163, + 0.006930043455213308, + 0.010938710533082485, + 0.007807333488017321, + 0.013897343538701534, + 0.009130693040788174, + 0.01295529492199421, + 0.009462114423513412, + 0.012558437883853912, + 0.011636101640760899, + 0.013701735995709896, + 0.02685711719095707, + 0.008849361911416054, + 0.002445589518174529, + 0.0054092612117528915, + 0.004578708205372095, + 0.004738223273307085, + 0.005165220703929663, + 0.0020770817063748837, + 0.003220238257199526, + 0.002182896714657545, + 0.014573333784937859, + 0.053245656192302704, + 0.019628584384918213, + 0.03220658004283905, + 0.042138174176216125, + 0.030732326209545135, + 0.029715126380324364, + 0.014321841299533844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.062058594077825546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02869221195578575, + 0.010837670415639877, + 0.7541221380233765, + 0.007641814183443785, + 0.04858489707112312, + 0.023309053853154182, + 0.034601133316755295, + 0.08208810538053513, + 0.005857095587998629, + 0.004265904892235994 + ], + "credit_map": [ + 0.003509253729134798, + 0.0033682321663945913, + 0.003973217215389013, + 0.0039039740804582834, + 0.0031242321711033583, + 0.0029996091034263372, + 0.004793904721736908, + 0.003019790630787611, + 0.0011633806861937046, + 0.0009554109419696033, + 0.0016415133140981197, + 0.0019483922515064478, + 0.0011102412827312946, + 0.0014879351947456598, + 0.0013301344588398933, + 0.0012006626930087805, + 0.05028907582163811, + 0.05186423659324646, + 0.22663581371307373, + 0.09355004876852036, + 0.04797730594873428, + 0.11650355905294418, + 0.036335039883852005, + 0.13096706569194794, + 0.0006100424216128886, + 0.0009447849006392062, + 0.0008065874571911991, + 0.0010716876713559031, + 0.0011067645391449332, + 0.0010005055228248239, + 0.0010979629587382078, + 0.0010034784208983183, + 0.002834325423464179, + 0.007391350343823433, + 0.0061101121827960014, + 0.00637165829539299, + 0.004462974611669779, + 0.009472725912928581, + 0.005983579903841019, + 0.005958169233053923, + 0.0016074038576334715, + 0.0016177144134417176, + 0.002062140265479684, + 0.0017045466229319572, + 0.0024810403119772673, + 0.006796344183385372, + 0.0018954691477119923, + 0.005144395399838686, + 0.003106201533228159, + 0.00431658374145627, + 0.0043065957725048065, + 0.0036027831956744194, + 0.0022877876181155443, + 0.0063621229492127895, + 0.003526755841448903, + 0.007092304527759552, + 0.005844625178724527, + 0.007840577512979507, + 0.006052451208233833, + 0.011215481907129288, + 0.005501766689121723, + 0.010140690952539444, + 0.030297258868813515, + 0.00519525445997715, + 0.0005641906755045056, + 0.0009855274111032486, + 0.0008456665091216564, + 0.000986461527645588, + 0.0006597484461963177, + 0.0006303515401668847, + 0.0006263463874347508, + 0.0005588029162026942, + 0.0003279585507698357, + 0.0006542698829434812, + 0.0002585119800642133, + 0.0005733583820983768, + 0.0006264643161557615, + 0.0005809836438857019, + 0.0007940753712318838, + 0.00045028282329440117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 several Fremen warriors visible racing across sun-baked sand dunes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the foreground warrior wearing a dark gray desert stillsuit with visible tubes and filters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foreground warrior's eyes glow with a deep luminous blue color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sand beneath the foreground warrior's feet collapse and explode backward into visible grains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle full-body tracking shot following the warriors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177704885601997 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177704885601997, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018177704885601997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10600273311138153, + 0.03446437045931816, + 0.5306158065795898, + 0.013557050377130508, + 0.05329789221286774, + 0.07043249905109406, + 0.025478648021817207, + 0.09843908250331879, + 0.029962431639432907, + 0.03774941340088844 + ], + "credit_map": [ + 0.021268142387270927, + 0.02146470919251442, + 0.011180698871612549, + 0.010504846461117268, + 0.013493369333446026, + 0.009773772209882736, + 0.010794512927532196, + 0.007522680796682835, + 0.0038566971197724342, + 0.0032024963293224573, + 0.004162849858403206, + 0.00684282137081027, + 0.004234360530972481, + 0.0036310816649347544, + 0.005482639651745558, + 0.003051426261663437, + 0.03519066423177719, + 0.03724512830376625, + 0.15645046532154083, + 0.0675482228398323, + 0.03446279466152191, + 0.08215951919555664, + 0.026192734017968178, + 0.09136630594730377, + 0.0009932679822668433, + 0.0015310465823858976, + 0.0016683181747794151, + 0.0016769604990258813, + 0.0016368224751204252, + 0.002051259158179164, + 0.002343122148886323, + 0.0016562535893172026, + 0.004455083981156349, + 0.005438225343823433, + 0.0075474395416677, + 0.009132611565291882, + 0.005359336733818054, + 0.007470625918358564, + 0.0070738596841692924, + 0.006820707116276026, + 0.004684934392571449, + 0.007490041200071573, + 0.007057476323097944, + 0.006472393870353699, + 0.005831546615809202, + 0.018187623471021652, + 0.005493721459060907, + 0.015214767307043076, + 0.0028731166385114193, + 0.0047760638408362865, + 0.0033628277014940977, + 0.0030774103943258524, + 0.002842285903170705, + 0.002768182195723057, + 0.0023953961208462715, + 0.003383365925401449, + 0.010413238778710365, + 0.00933059025555849, + 0.005992690101265907, + 0.012126271612942219, + 0.009336387738585472, + 0.011844906024634838, + 0.032330434769392014, + 0.007064566481858492, + 0.002756484318524599, + 0.004403269849717617, + 0.004177478142082691, + 0.005911398213356733, + 0.0030679514165967703, + 0.002951172413304448, + 0.003609397914260626, + 0.0030852779746055603, + 0.002145748119801283, + 0.00355561263859272, + 0.009410025551915169, + 0.0035200491547584534, + 0.004698717035353184, + 0.0031333514489233494, + 0.0048041073605418205, + 0.006481802556663752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.018177704885601997 + } + ], + "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/0000058/group_13/index_tensor([966, 966])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_tensor([966, 966])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b940e39a5b208a9259793e2f3a03d1e9c7acf283 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_tensor([966, 966])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a low angle, just above the turbulent sea, the camera moving parallel to the subject, captures a massive tank, clad in deep gunmetal armor, tearing through the deep blue, almost black, water with unstoppable force. The tank's bow-like front armor, etched with battle-worn scratches, rivets, and weathered rust, appears impenetrable. The camera's focus is pinpointed on the massive splash of water erupting from the tank's side at high speed. This white wall of water, nearly engulfing the camera, is composed of countless shimmering droplets and roiling foam. The slow-motion effect allows every splash to be clearly discerned, like a momentarily frozen liquid sculpture. The extremely shallow depth of field completely blurs the distant background, leaving only the faint outline of a gray, gloomy sky. The entire scene, illuminated by a cold, diffuse light against a gloomy sky, features a dominant palette of icy metallic gray, deep ocean blue, and the dazzling white of the waves, creating an atmosphere imbued with raw power, tension, and epic grandeur.", + "seed": 18338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_966_seed_18338_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": -0.034192612394690514, + "vlm_logprob": -0.034192612394690514, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.034192612394690514, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.11240918189287186 + ], + "opposite_logprob_tokens": [ + -2.3624091148376465 + ], + "logprob_reward": -0.11240918189287186, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11240918189287186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24846318364143372, + 0.11432361602783203, + 0.06253987550735474, + 0.06019360572099686, + 0.1570902168750763, + 0.078022800385952, + 0.02885689213871956, + 0.03136357665061951, + 0.04586051031947136, + 0.17328575253486633 + ], + "credit_map": [ + 0.03939682990312576, + 0.025070710107684135, + 0.017468875274062157, + 0.05414586141705513, + 0.020919684320688248, + 0.04102374240756035, + 0.040337689220905304, + 0.01009980496019125, + 0.017999641597270966, + 0.004719393793493509, + 0.004707609303295612, + 0.010591230355203152, + 0.007064168807119131, + 0.03181164339184761, + 0.032545216381549835, + 0.0048847203142941, + 0.0069159697741270065, + 0.007472989149391651, + 0.009423600509762764, + 0.014695518650114536, + 0.009524345397949219, + 0.005399653688073158, + 0.0053030867129564285, + 0.0038047139532864094, + 0.005133116617798805, + 0.004819204099476337, + 0.013680201023817062, + 0.012547492049634457, + 0.005732876714318991, + 0.004511731676757336, + 0.00687656132504344, + 0.006892418023198843, + 0.005566374398767948, + 0.007195597514510155, + 0.059979721903800964, + 0.055305227637290955, + 0.007419251371175051, + 0.006304854992777109, + 0.007387144025415182, + 0.007932057604193687, + 0.006986855994910002, + 0.011517221108078957, + 0.00871733296662569, + 0.009080621413886547, + 0.01225134078413248, + 0.010663039982318878, + 0.010835015214979649, + 0.00797137338668108, + 0.0032937433570623398, + 0.003387295873835683, + 0.0047340416349470615, + 0.003508641617372632, + 0.0038797936867922544, + 0.0033465002197772264, + 0.0035396176390349865, + 0.003167257411405444, + 0.00542610976845026, + 0.004277653526514769, + 0.0035134952049702406, + 0.0038670115172863007, + 0.004080906976014376, + 0.002973568392917514, + 0.004114761017262936, + 0.0031100697815418243, + 0.005211903713643551, + 0.004604939837008715, + 0.0075926873832941055, + 0.007317518815398216, + 0.005554468836635351, + 0.0035974469501525164, + 0.00430698785930872, + 0.007674556691199541, + 0.007077808957546949, + 0.008986632339656353, + 0.010182813741266727, + 0.06142739579081535, + 0.008278313092887402, + 0.0063211871311068535, + 0.061566177755594254, + 0.009445426985621452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.11240918189287186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.018217036500573158 + ], + "opposite_logprob_tokens": [ + -4.018217086791992 + ], + "logprob_reward": -0.018217036500573158, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018217036500573158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09596633911132812, + 0.059476882219314575, + 0.04931780695915222, + 0.04326193407177925, + 0.4344772696495056, + 0.08714189380407333, + 0.0290359016507864, + 0.07529214024543762, + 0.028052033856511116, + 0.09797781705856323 + ], + "credit_map": [ + 0.013003885746002197, + 0.006018173880875111, + 0.0061389384791255, + 0.028001898899674416, + 0.007421937771141529, + 0.01109582744538784, + 0.02042696252465248, + 0.0038587129674851894, + 0.01195470429956913, + 0.005101555027067661, + 0.003122817724943161, + 0.003451831405982375, + 0.00789992231875658, + 0.018952354788780212, + 0.0046502756886184216, + 0.004343417007476091, + 0.0060283769853413105, + 0.008677816018462181, + 0.004457632079720497, + 0.00848610233515501, + 0.007654184941202402, + 0.005059666931629181, + 0.004431014880537987, + 0.004523011855781078, + 0.0041945879347622395, + 0.0032557060476392508, + 0.01064409501850605, + 0.008857550099492073, + 0.0023267040960490704, + 0.003528146306052804, + 0.003738571424037218, + 0.006716570351272821, + 0.014651435427367687, + 0.012543804943561554, + 0.17477135360240936, + 0.17674760520458221, + 0.007755513768643141, + 0.016406474635004997, + 0.015889020636677742, + 0.015712032094597816, + 0.0070465863682329655, + 0.015943821519613266, + 0.010662379674613476, + 0.009591879323124886, + 0.010901677422225475, + 0.01078001782298088, + 0.013516596518456936, + 0.008698933757841587, + 0.00380254490301013, + 0.003326855134218931, + 0.004497527610510588, + 0.0031626136042177677, + 0.0038978683296591043, + 0.003025050973519683, + 0.004315838683396578, + 0.003007601946592331, + 0.011558984406292439, + 0.007774772122502327, + 0.008001482114195824, + 0.008774696849286556, + 0.008928422816097736, + 0.009491278789937496, + 0.011599344201385975, + 0.009163164533674717, + 0.003299650503322482, + 0.0030626016668975353, + 0.004683898761868477, + 0.005021442659199238, + 0.0034643688704818487, + 0.0021265579853206873, + 0.0030175175052136183, + 0.0033759959042072296, + 0.0037036484573036432, + 0.005147371906787157, + 0.006280218716710806, + 0.03609013184905052, + 0.004580753855407238, + 0.002904749708250165, + 0.03444632887840271, + 0.004824609961360693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.018217036500573158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.011091836728155613 + ], + "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": [ + 0.076743483543396, + 0.21062904596328735, + 0.0886274129152298, + 0.026683811098337173, + 0.25617435574531555, + 0.08359520882368088, + 0.028261646628379822, + 0.05759529769420624, + 0.08133973181247711, + 0.0903499647974968 + ], + "credit_map": [ + 0.010851235128939152, + 0.010388336144387722, + 0.00729094073176384, + 0.009866733103990555, + 0.008160806261003017, + 0.011650605127215385, + 0.012478543445467949, + 0.006056283600628376, + 0.03317501023411751, + 0.009216994047164917, + 0.006565391086041927, + 0.01741795986890793, + 0.009986728429794312, + 0.06322664767503738, + 0.06106256693601608, + 0.009977763518691063, + 0.008479008451104164, + 0.01210244745016098, + 0.00849965214729309, + 0.019869346171617508, + 0.01445250678807497, + 0.007900135591626167, + 0.009591856971383095, + 0.007732452359050512, + 0.0021091080270707607, + 0.0028994770254939795, + 0.005914757959544659, + 0.0037619811482727528, + 0.002355520147830248, + 0.002603142987936735, + 0.0029403662774711847, + 0.004099457990378141, + 0.008364797569811344, + 0.00770904216915369, + 0.09876403957605362, + 0.10442380607128143, + 0.006863774731755257, + 0.010550527833402157, + 0.010068054310977459, + 0.009430318139493465, + 0.007333024870604277, + 0.012033108621835709, + 0.010308746248483658, + 0.00878150388598442, + 0.010687286034226418, + 0.011382636614143848, + 0.013883078470826149, + 0.009185823611915112, + 0.004302590154111385, + 0.0030161673203110695, + 0.004128939472138882, + 0.002930102404206991, + 0.004548666533082724, + 0.002742954296991229, + 0.003354676766321063, + 0.0032375494483858347, + 0.009777655825018883, + 0.006891180761158466, + 0.005239177495241165, + 0.006385978776961565, + 0.007560133934020996, + 0.008921069093048573, + 0.007301795296370983, + 0.005518308840692043, + 0.010367373935878277, + 0.006566493771970272, + 0.01027971412986517, + 0.01724042370915413, + 0.005957635585218668, + 0.006291334517300129, + 0.010230830870568752, + 0.0144059332087636, + 0.0038222214207053185, + 0.004977227188646793, + 0.004817655310034752, + 0.033644817769527435, + 0.0041483561508357525, + 0.0030305301770567894, + 0.030047312378883362, + 0.005861846264451742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.011091836728155613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.02329145185649395 + ], + "opposite_logprob_tokens": [ + -3.7732913494110107 + ], + "logprob_reward": -0.02329145185649395, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329145185649395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09768655151128769, + 0.09254765510559082, + 0.09273369610309601, + 0.037221383303403854, + 0.23787979781627655, + 0.041800323873758316, + 0.0836232602596283, + 0.02749214880168438, + 0.17427267134189606, + 0.11474258452653885 + ], + "credit_map": [ + 0.020114164799451828, + 0.007405892945826054, + 0.004897480830550194, + 0.015739696100354195, + 0.009838512167334557, + 0.023007340729236603, + 0.011805514805018902, + 0.004877944011241198, + 0.010701779276132584, + 0.005111214239150286, + 0.0051093799993395805, + 0.008404804393649101, + 0.004796856082975864, + 0.020250748842954636, + 0.02942175231873989, + 0.008751117624342442, + 0.00800663884729147, + 0.008391179144382477, + 0.013888047076761723, + 0.026163455098867416, + 0.010197319090366364, + 0.00541803240776062, + 0.010890765115618706, + 0.009778250940144062, + 0.0034530097618699074, + 0.0034818171989172697, + 0.008144153282046318, + 0.008578783832490444, + 0.0032989357132464647, + 0.0027390688192099333, + 0.0039476691745221615, + 0.003577941097319126, + 0.012396953999996185, + 0.010612750425934792, + 0.08909939974546432, + 0.07830782979726791, + 0.009079262614250183, + 0.009543120861053467, + 0.01245921291410923, + 0.016381245106458664, + 0.00393690587952733, + 0.005685311742126942, + 0.006053323857486248, + 0.005099967587739229, + 0.005027692299336195, + 0.0048390221782028675, + 0.006715771276503801, + 0.004442327190190554, + 0.009385401383042336, + 0.009247911162674427, + 0.010285918600857258, + 0.009288795292377472, + 0.013309179805219173, + 0.008004853501915932, + 0.014552476815879345, + 0.009548718109726906, + 0.0034047740045934916, + 0.003376121399924159, + 0.0029443211387842894, + 0.0050683519802987576, + 0.0027535075787454844, + 0.003423541784286499, + 0.003829936496913433, + 0.0026915937196463346, + 0.021524621173739433, + 0.013275775127112865, + 0.022840863093733788, + 0.03556313365697861, + 0.01281046960502863, + 0.013212122954428196, + 0.021448984742164612, + 0.03359668329358101, + 0.006173812318593264, + 0.00701781315729022, + 0.008235326036810875, + 0.03861398994922638, + 0.0061837839893996716, + 0.00497774500399828, + 0.036205314099788666, + 0.0073347920551896095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.02329145185649395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.00595355499535799 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595355499535799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595355499535799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23598738014698029, + 0.062049221247434616, + 0.053779009729623795, + 0.019505232572555542, + 0.2870044410228729, + 0.05841647833585739, + 0.05791441351175308, + 0.0916406661272049, + 0.10716830939054489, + 0.026534872129559517 + ], + "credit_map": [ + 0.03251175209879875, + 0.01179363764822483, + 0.009307081811130047, + 0.08618024736642838, + 0.010485509410500526, + 0.02167576178908348, + 0.057008758187294006, + 0.007024628110229969, + 0.008990154601633549, + 0.0028189048171043396, + 0.0039683240465819836, + 0.005939100868999958, + 0.003931874874979258, + 0.014737516641616821, + 0.017230840399861336, + 0.00443250359967351, + 0.004263132810592651, + 0.005131315439939499, + 0.006849965080618858, + 0.015023271553218365, + 0.006020262371748686, + 0.0037156345788389444, + 0.008280303329229355, + 0.004495122469961643, + 0.0022698589600622654, + 0.002124706283211708, + 0.0026763142086565495, + 0.003898999188095331, + 0.0020358357578516006, + 0.0025347613263875246, + 0.002397914882749319, + 0.001566841616295278, + 0.009427200071513653, + 0.009190689772367477, + 0.11040793359279633, + 0.11887890100479126, + 0.006839316338300705, + 0.012249783612787724, + 0.011352539993822575, + 0.008658082224428654, + 0.005135118030011654, + 0.011087593622505665, + 0.0053305416367948055, + 0.00542272487655282, + 0.012981845997273922, + 0.009035026654601097, + 0.005344918463379145, + 0.004078705795109272, + 0.005526547320187092, + 0.005606398917734623, + 0.010833670385181904, + 0.0063675688579678535, + 0.0071372357197105885, + 0.006072188261896372, + 0.009256244637072086, + 0.007114555686712265, + 0.014825538732111454, + 0.01301497034728527, + 0.008358349092304707, + 0.012432943098247051, + 0.011872133240103722, + 0.010258623398840427, + 0.011183579452335835, + 0.00969452690333128, + 0.013763480819761753, + 0.008271520957350731, + 0.013186218217015266, + 0.019489848986268044, + 0.006533314939588308, + 0.009740890935063362, + 0.013608391396701336, + 0.02257462777197361, + 0.0026717407163232565, + 0.0041725547052919865, + 0.0036265263333916664, + 0.004050143528729677, + 0.0025182352401316166, + 0.0033105977345257998, + 0.0027163957711309195, + 0.003468675073236227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.00595355499535799 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, just above the turbulent sea, the camera moving parallel to the subject, captures a massive tank, clad in deep gunmetal armor, tearing through the deep blue, almost black, water with unstoppable force. The tank's bow-like front armor, etched with battle-worn scratches, rivets, and weathered rust, appears impenetrable. The camera's focus is pinpointed on the massive splash of water erupting from the tank's side at high speed. This white wall of water, nearly engulfing the camera, is composed of countless shimmering droplets and roiling foam. The slow-motion effect allows every splash to be clearly discerned, like a momentarily frozen liquid sculpture. The extremely shallow depth of field completely blurs the distant background, leaving only the faint outline of a gray, gloomy sky. The entire scene, illuminated by a cold, diffuse light against a gloomy sky, features a dominant palette of icy metallic gray, deep ocean blue, and the dazzling white of the waves, creating an atmosphere imbued with raw power, tension, and epic grandeur.", + "seed": 18339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_966_seed_18339_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": -0.10102718705311417, + "vlm_logprob": -0.10102718705311417, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.10102718705311417, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.43773531913757324 + ], + "opposite_logprob_tokens": [ + -1.0627353191375732 + ], + "logprob_reward": -0.43773531913757324, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43773531913757324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6551552414894104, + 0.08381395787000656, + 0.012400440871715546, + 0.023507457226514816, + 0.017817534506320953, + 0.028579363599419594, + 0.10329612344503403, + 0.021486161276698112, + 0.0220783744007349, + 0.03186535835266113 + ], + "credit_map": [ + 0.15081356465816498, + 0.06581871956586838, + 0.036888398230075836, + 0.052459072321653366, + 0.11820819973945618, + 0.0971316397190094, + 0.07192587852478027, + 0.06190976873040199, + 0.009016383439302444, + 0.00752806942909956, + 0.010612308979034424, + 0.012867839075624943, + 0.010656802915036678, + 0.008119593374431133, + 0.014955643564462662, + 0.01005730964243412, + 0.0018716375343501568, + 0.0021034367382526398, + 0.0015226405812427402, + 0.0014487146399915218, + 0.0013850473333150148, + 0.001844792510382831, + 0.0010790519881993532, + 0.001145119545981288, + 0.003992229234427214, + 0.0037415572442114353, + 0.002785280579701066, + 0.004448489751666784, + 0.0016954863676801324, + 0.0014464681735262275, + 0.0030129393562674522, + 0.0023850074503570795, + 0.002393700648099184, + 0.0017598317936062813, + 0.0018256035400554538, + 0.0033329068683087826, + 0.00224664737470448, + 0.0013278292026370764, + 0.0021316036581993103, + 0.002799412701278925, + 0.003829857101663947, + 0.0027392099145799875, + 0.004226382356137037, + 0.0022111660800874233, + 0.0038594899233430624, + 0.003728312673047185, + 0.0028919419273734093, + 0.005093004088848829, + 0.013350654393434525, + 0.013282151892781258, + 0.010400722734630108, + 0.009179924614727497, + 0.01979631558060646, + 0.013687832280993462, + 0.013883709907531738, + 0.009714805521070957, + 0.006079382728785276, + 0.0037235429044812918, + 0.0012556500732898712, + 0.001664364943280816, + 0.0030196337029337883, + 0.0021570471581071615, + 0.001357730943709612, + 0.0022288088221102953, + 0.003461801912635565, + 0.003027025144547224, + 0.0016250362386927009, + 0.0017023510299623013, + 0.001735364436171949, + 0.005089347716420889, + 0.0016067519318312407, + 0.0038306957576423883, + 0.0046789986081421375, + 0.004354873206466436, + 0.0032403108198195696, + 0.0025117441546171904, + 0.003880877513438463, + 0.00747530534863472, + 0.0021764130797237158, + 0.0035468358546495438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.43773531913757324 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.02983291819691658 + ], + "opposite_logprob_tokens": [ + -3.5298328399658203 + ], + "logprob_reward": -0.02983291819691658, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983291819691658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31183895468711853, + 0.15115824341773987, + 0.07187087088823318, + 0.12676933407783508, + 0.023306583985686302, + 0.07318074256181717, + 0.14171601831912994, + 0.037107206881046295, + 0.024549782276153564, + 0.03850221261382103 + ], + "credit_map": [ + 0.06427495181560516, + 0.032732121646404266, + 0.028349077329039574, + 0.027233965694904327, + 0.05256209895014763, + 0.044797834008932114, + 0.03770424798130989, + 0.02418466843664646, + 0.01724589802324772, + 0.01628454588353634, + 0.01881207711994648, + 0.023343615233898163, + 0.015647729858756065, + 0.017301594838500023, + 0.023614997044205666, + 0.018907777965068817, + 0.011985616758465767, + 0.009055865928530693, + 0.013957922346889973, + 0.014821301214396954, + 0.006102582439780235, + 0.005282415077090263, + 0.005344794597476721, + 0.00532037066295743, + 0.01893441192805767, + 0.020378591492772102, + 0.017508508637547493, + 0.019621476531028748, + 0.008728706277906895, + 0.007879958488047123, + 0.020679449662566185, + 0.01303823571652174, + 0.004922146443277597, + 0.002766759367659688, + 0.0027726050466299057, + 0.0032248031347990036, + 0.0024338052608072758, + 0.002168823964893818, + 0.0029313783161342144, + 0.0020862629171460867, + 0.010052463971078396, + 0.008302077651023865, + 0.010035095736384392, + 0.005947495810687542, + 0.012225447222590446, + 0.01044727023690939, + 0.007107864134013653, + 0.009063024073839188, + 0.01679452322423458, + 0.012364196591079235, + 0.016245117411017418, + 0.01318288967013359, + 0.026274971663951874, + 0.02423926256597042, + 0.0149103794246912, + 0.0177046749740839, + 0.007069061975926161, + 0.00485189538449049, + 0.0038006326649338007, + 0.0039753089658916, + 0.005930005572736263, + 0.003951772581785917, + 0.0028033999260514975, + 0.004725131206214428, + 0.003478326601907611, + 0.003193615935742855, + 0.00262279249727726, + 0.0030429293401539326, + 0.002953090937808156, + 0.0035732975229620934, + 0.0031383575405925512, + 0.0025473705027252436, + 0.008577629923820496, + 0.0069895642809569836, + 0.005088150501251221, + 0.004363960586488247, + 0.0033547861967235804, + 0.004019982647150755, + 0.003258329350501299, + 0.0028498084284365177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02983291819691658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.005962086841464043 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005962086841464043, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962086841464043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2384018450975418, + 0.3893842101097107, + 0.07361364364624023, + 0.0555022656917572, + 0.014896158128976822, + 0.05173840746283531, + 0.04226349666714668, + 0.0857599526643753, + 0.02288183569908142, + 0.025558164343237877 + ], + "credit_map": [ + 0.05912606790661812, + 0.023206356912851334, + 0.01904386468231678, + 0.017741985619068146, + 0.0370071716606617, + 0.03958020731806755, + 0.02602950483560562, + 0.016666682437062263, + 0.04643033817410469, + 0.06600430607795715, + 0.04471507668495178, + 0.05329638719558716, + 0.05159253254532814, + 0.03128199651837349, + 0.05088958889245987, + 0.045173969119787216, + 0.011171439662575722, + 0.006952679716050625, + 0.01248778123408556, + 0.018430836498737335, + 0.005122428759932518, + 0.005627486854791641, + 0.008088482543826103, + 0.005732506513595581, + 0.008345904760062695, + 0.008100147359073162, + 0.007621543016284704, + 0.009636711329221725, + 0.005841950420290232, + 0.003915790002793074, + 0.005373470950871706, + 0.006666748784482479, + 0.0024831232149153948, + 0.0017415694892406464, + 0.002059937221929431, + 0.002209103899076581, + 0.0014559203991666436, + 0.0015742286341264844, + 0.0016525534447282553, + 0.0017197211273014545, + 0.006506709381937981, + 0.00647690100595355, + 0.006697963923215866, + 0.005849544890224934, + 0.007503123953938484, + 0.006829519756138325, + 0.004454555455595255, + 0.007420089095830917, + 0.00474697258323431, + 0.004064397420734167, + 0.005335671827197075, + 0.003635929198935628, + 0.00769241014495492, + 0.0077381678856909275, + 0.0045145004987716675, + 0.00453544408082962, + 0.013936936855316162, + 0.009652524255216122, + 0.007605860009789467, + 0.010655999183654785, + 0.016871875151991844, + 0.013232075609266758, + 0.006561386864632368, + 0.0072433012537658215, + 0.004205987323075533, + 0.0032299526501446962, + 0.0019659993704408407, + 0.00220663589425385, + 0.0020735962316393852, + 0.002331308787688613, + 0.002844056813046336, + 0.0040242974646389484, + 0.004995900206267834, + 0.0036760373041033745, + 0.003479793667793274, + 0.002952110255137086, + 0.002064424566924572, + 0.0035702597815543413, + 0.0024893779773265123, + 0.002330259419977665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.005962086841464043 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.026349563151597977 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349563151597977, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349563151597977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3128218948841095, + 0.0937679260969162, + 0.10835796594619751, + 0.0840512365102768, + 0.03989488631486893, + 0.10670742392539978, + 0.11403826624155045, + 0.04706001654267311, + 0.030110588297247887, + 0.06318981945514679 + ], + "credit_map": [ + 0.06665780395269394, + 0.032141540199518204, + 0.018801884725689888, + 0.026623858138918877, + 0.053702984005212784, + 0.04393516480922699, + 0.03930146247148514, + 0.03165718540549278, + 0.01097610779106617, + 0.011163371615111828, + 0.009868060238659382, + 0.013644064776599407, + 0.012614245526492596, + 0.009824128821492195, + 0.014028353616595268, + 0.011649583466351032, + 0.01714419201016426, + 0.012030425481498241, + 0.014457207173109055, + 0.023503845557570457, + 0.007165813352912664, + 0.009335741400718689, + 0.013352037407457829, + 0.0113687040284276, + 0.012163328006863594, + 0.010953527875244617, + 0.01171022653579712, + 0.017117641866207123, + 0.006674566306173801, + 0.0056442865170538425, + 0.010530448518693447, + 0.009257207624614239, + 0.006600675173103809, + 0.00440952368080616, + 0.004468474071472883, + 0.0073090242221951485, + 0.004757318180054426, + 0.003397989086806774, + 0.004276935942471027, + 0.004674948286265135, + 0.013890310190618038, + 0.013309849426150322, + 0.013758864253759384, + 0.012609051540493965, + 0.013738200068473816, + 0.01155257411301136, + 0.010456356219947338, + 0.01739221066236496, + 0.013610131107270718, + 0.010876576416194439, + 0.013566712848842144, + 0.00979875773191452, + 0.023644456639885902, + 0.019566353410482407, + 0.0117423040792346, + 0.011232971213757992, + 0.008190817199647427, + 0.005429302342236042, + 0.004636578727513552, + 0.00475339824333787, + 0.008710552006959915, + 0.008200871758162975, + 0.003427986754104495, + 0.003710508346557617, + 0.004188308026641607, + 0.004264517687261105, + 0.002356687793508172, + 0.002971396315842867, + 0.002328957198187709, + 0.004084970336407423, + 0.004584700334817171, + 0.005331047810614109, + 0.011054180562496185, + 0.010112281888723373, + 0.005656978115439415, + 0.00648067332804203, + 0.006191978696733713, + 0.012624737806618214, + 0.004580496810376644, + 0.006488483399152756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.026349563151597977 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22069260478019714, + 0.14701461791992188, + 0.12410040944814682, + 0.09931610524654388, + 0.05960828810930252, + 0.07766997814178467, + 0.05001014471054077, + 0.10365515202283859, + 0.0666612759232521, + 0.05127144232392311 + ], + "credit_map": [ + 0.04107634350657463, + 0.02358436957001686, + 0.022245310246944427, + 0.022463863715529442, + 0.025898225605487823, + 0.025940630584955215, + 0.03497632220387459, + 0.024507543072104454, + 0.018942272290587425, + 0.0295155830681324, + 0.014816203154623508, + 0.016477221623063087, + 0.018042998388409615, + 0.018548768013715744, + 0.01624760590493679, + 0.014423967339098454, + 0.01784215308725834, + 0.015599440783262253, + 0.016555076465010643, + 0.026155592873692513, + 0.007898034527897835, + 0.012239477597177029, + 0.014364358969032764, + 0.013446283526718616, + 0.016646355390548706, + 0.013781712390482426, + 0.014054358936846256, + 0.02144397422671318, + 0.006172118242830038, + 0.007137718144804239, + 0.009962973184883595, + 0.010116903111338615, + 0.009164713323116302, + 0.006969410926103592, + 0.006596916820853949, + 0.012204893864691257, + 0.006453682668507099, + 0.00606435677036643, + 0.006143439561128616, + 0.006010873708873987, + 0.012113007716834545, + 0.006811115425080061, + 0.01568555273115635, + 0.010561810806393623, + 0.008358811028301716, + 0.008318829350173473, + 0.008282202295958996, + 0.007538652513176203, + 0.006667510140687227, + 0.009404938668012619, + 0.006989413406699896, + 0.004688374698162079, + 0.0051614707335829735, + 0.005281607620418072, + 0.004846907686442137, + 0.006969920825213194, + 0.023784907534718513, + 0.015362801030278206, + 0.00933659914880991, + 0.013182162307202816, + 0.016956055536866188, + 0.011071251705288887, + 0.0066429306752979755, + 0.007318446878343821, + 0.01234352681785822, + 0.009932282380759716, + 0.0034833166282624006, + 0.0055452133528888226, + 0.005065679084509611, + 0.00813282560557127, + 0.007854357361793518, + 0.014304074458777905, + 0.008704924955964088, + 0.007108317222446203, + 0.006567648611962795, + 0.008435525000095367, + 0.0033992729149758816, + 0.0036470876075327396, + 0.0071679591201245785, + 0.0062407092191278934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005256047938019037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, just above the turbulent sea, the camera moving parallel to the subject, captures a massive tank, clad in deep gunmetal armor, tearing through the deep blue, almost black, water with unstoppable force. The tank's bow-like front armor, etched with battle-worn scratches, rivets, and weathered rust, appears impenetrable. The camera's focus is pinpointed on the massive splash of water erupting from the tank's side at high speed. This white wall of water, nearly engulfing the camera, is composed of countless shimmering droplets and roiling foam. The slow-motion effect allows every splash to be clearly discerned, like a momentarily frozen liquid sculpture. The extremely shallow depth of field completely blurs the distant background, leaving only the faint outline of a gray, gloomy sky. The entire scene, illuminated by a cold, diffuse light against a gloomy sky, features a dominant palette of icy metallic gray, deep ocean blue, and the dazzling white of the waves, creating an atmosphere imbued with raw power, tension, and epic grandeur.", + "seed": 18340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_966_seed_18340_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": -0.12286551157012582, + "vlm_logprob": -0.12286551157012582, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.12286551157012582, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.5328981876373291 + ], + "opposite_logprob_tokens": [ + -0.9078981876373291 + ], + "logprob_reward": -0.5328981876373291, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5328981876373291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10402518510818481, + 0.055489104241132736, + 0.03941880166530609, + 0.203830286860466, + 0.04358692467212677, + 0.08755793422460556, + 0.02330789342522621, + 0.11742518842220306, + 0.14123740792274475, + 0.1841212511062622 + ], + "credit_map": [ + 0.013410849496722221, + 0.011482296511530876, + 0.012242862954735756, + 0.014045956544578075, + 0.015962747856974602, + 0.012205606326460838, + 0.013991892337799072, + 0.010682981461286545, + 0.006670177914202213, + 0.007041990756988525, + 0.005583964753895998, + 0.007861848920583725, + 0.010032921098172665, + 0.007684365380555391, + 0.006003514397889376, + 0.004610327538102865, + 0.004238520283252001, + 0.007246547844260931, + 0.005170932039618492, + 0.006749880034476519, + 0.004169239196926355, + 0.0033946058247238398, + 0.004175872076302767, + 0.004273205995559692, + 0.02973080612719059, + 0.02405267022550106, + 0.025721488520503044, + 0.026826700195670128, + 0.020329564809799194, + 0.021671736612915993, + 0.023667961359024048, + 0.031829364597797394, + 0.006051034666597843, + 0.010571816936135292, + 0.003748214803636074, + 0.0034261655528098345, + 0.006860022433102131, + 0.0038786670193076134, + 0.004454968497157097, + 0.0045960391871631145, + 0.012231203727424145, + 0.013364030048251152, + 0.013462685979902744, + 0.014721269719302654, + 0.007867388427257538, + 0.005340301897376776, + 0.009692923165857792, + 0.010878142900764942, + 0.002077312208712101, + 0.0030157389119267464, + 0.0033095763064920902, + 0.002445422811433673, + 0.003440115600824356, + 0.004572184756398201, + 0.0022376272827386856, + 0.0022099167108535767, + 0.01588410884141922, + 0.01434248499572277, + 0.015441026538610458, + 0.008976323530077934, + 0.024636877700686455, + 0.018381135538220406, + 0.009514819830656052, + 0.010248417966067791, + 0.01912076584994793, + 0.015173695050179958, + 0.019794495776295662, + 0.01743231900036335, + 0.01876075565814972, + 0.01534690149128437, + 0.024147238582372665, + 0.011461252346634865, + 0.03738049417734146, + 0.02388552576303482, + 0.015416155569255352, + 0.02054475247859955, + 0.034303221851587296, + 0.016558941453695297, + 0.021404841914772987, + 0.014627331867814064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.5328981876373291 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.033733829855918884 + ], + "opposite_logprob_tokens": [ + -3.40873384475708 + ], + "logprob_reward": -0.033733829855918884, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033733829855918884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10828372091054916, + 0.046832144260406494, + 0.0590628981590271, + 0.1434299349784851, + 0.09949050098657608, + 0.08041810989379883, + 0.1009795144200325, + 0.11855092644691467, + 0.13222192227840424, + 0.11073033511638641 + ], + "credit_map": [ + 0.017892595380544662, + 0.013142935931682587, + 0.014309643767774105, + 0.017236851155757904, + 0.012368935160338879, + 0.009462643414735794, + 0.014017317444086075, + 0.009852795861661434, + 0.005242472048848867, + 0.00722914282232523, + 0.0057401033118367195, + 0.007526037283241749, + 0.00603485107421875, + 0.005119498819112778, + 0.005876805633306503, + 0.004063230007886887, + 0.008477823808789253, + 0.008192075416445732, + 0.007700103800743818, + 0.006700560450553894, + 0.006034021265804768, + 0.005664447788149118, + 0.007379361428320408, + 0.008914508856832981, + 0.021832594648003578, + 0.015389137901365757, + 0.017896439880132675, + 0.019740108400583267, + 0.015361951664090157, + 0.016987144947052002, + 0.016679566353559494, + 0.019542984664440155, + 0.012842778116464615, + 0.026395626366138458, + 0.01194731704890728, + 0.007784528192132711, + 0.013508336618542671, + 0.005309140309691429, + 0.009686891920864582, + 0.012015885673463345, + 0.010280423797667027, + 0.00931249838322401, + 0.012776804156601429, + 0.015388552099466324, + 0.007318020332604647, + 0.005579646211117506, + 0.008931566029787064, + 0.010830603539943695, + 0.01003851555287838, + 0.016376588493585587, + 0.015555215999484062, + 0.013133920729160309, + 0.01457353588193655, + 0.01185815129429102, + 0.010052611120045185, + 0.009390966035425663, + 0.013532159850001335, + 0.014442809857428074, + 0.02202373743057251, + 0.00878607016056776, + 0.021729737520217896, + 0.0166942086070776, + 0.00943875964730978, + 0.011903446167707443, + 0.015597803518176079, + 0.014633688144385815, + 0.021071292459964752, + 0.017293687909841537, + 0.014661461114883423, + 0.013154048472642899, + 0.02096988633275032, + 0.01484005805104971, + 0.02287445031106472, + 0.015248841606080532, + 0.009250533767044544, + 0.012095990590751171, + 0.021068844944238663, + 0.012742741033434868, + 0.009248976595699787, + 0.008199949748814106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.033733829855918884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.016083383932709694 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.016083383932709694, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083383932709694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17296823859214783, + 0.15478835999965668, + 0.04457155615091324, + 0.08878006041049957, + 0.06119578331708908, + 0.11674559116363525, + 0.04422133415937424, + 0.09013834595680237, + 0.09385044872760773, + 0.13274028897285461 + ], + "credit_map": [ + 0.02064821496605873, + 0.024698346853256226, + 0.017940737307071686, + 0.028612222522497177, + 0.028128115460276604, + 0.01760496199131012, + 0.018556807190179825, + 0.01677883416414261, + 0.016933100298047066, + 0.024411138147115707, + 0.01807563193142414, + 0.020289616659283638, + 0.02690173126757145, + 0.019429825246334076, + 0.01698172278702259, + 0.011765571311116219, + 0.0041244300082325935, + 0.005897626746445894, + 0.004633708856999874, + 0.007388618309050798, + 0.00523549085482955, + 0.0047544632107019424, + 0.006867776624858379, + 0.005669434554874897, + 0.010839465074241161, + 0.012988626025617123, + 0.009351343847811222, + 0.014679666608572006, + 0.011657461524009705, + 0.008571043610572815, + 0.01058097556233406, + 0.010111470706760883, + 0.01293741725385189, + 0.009156245738267899, + 0.005408016499131918, + 0.00760924257338047, + 0.007912902161478996, + 0.00499259028583765, + 0.00650012819096446, + 0.006679238751530647, + 0.011286452412605286, + 0.01107403077185154, + 0.011611121706664562, + 0.024489477276802063, + 0.01327124610543251, + 0.012791165150702, + 0.013644164428114891, + 0.018577920272946358, + 0.0035691875964403152, + 0.0036726160906255245, + 0.004095280077308416, + 0.004478965885937214, + 0.006455270107835531, + 0.011740664951503277, + 0.004309249110519886, + 0.0059000966139137745, + 0.01108228787779808, + 0.010723032057285309, + 0.013733577914536, + 0.008463233709335327, + 0.01598370261490345, + 0.013638722710311413, + 0.007648245897144079, + 0.00886552780866623, + 0.010805013589560986, + 0.012112759053707123, + 0.012298821471631527, + 0.009680586867034435, + 0.013048683293163776, + 0.016311166808009148, + 0.011638862080872059, + 0.007954546250402927, + 0.028575267642736435, + 0.018466593697667122, + 0.010766281746327877, + 0.013555178418755531, + 0.02322927676141262, + 0.016925891861319542, + 0.012923934496939182, + 0.008297848515212536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.016083383932709694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.026353973895311356 + ], + "opposite_logprob_tokens": [ + -3.6513540744781494 + ], + "logprob_reward": -0.026353973895311356, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353973895311356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1039128378033638, + 0.049985434859991074, + 0.03636107221245766, + 0.3167697489261627, + 0.06537597626447678, + 0.11231642961502075, + 0.028420474380254745, + 0.09926643967628479, + 0.10167309641838074, + 0.08591844141483307 + ], + "credit_map": [ + 0.011082700453698635, + 0.01352451741695404, + 0.008924764581024647, + 0.01752023957669735, + 0.010152732953429222, + 0.010938943363726139, + 0.01644929312169552, + 0.0153196444734931, + 0.00533713074401021, + 0.005702128633856773, + 0.005500501021742821, + 0.006289719603955746, + 0.00892226118594408, + 0.007174413185566664, + 0.00634727394208312, + 0.0047120084054768085, + 0.003916425164788961, + 0.004532968159765005, + 0.004136682488024235, + 0.0045111472718417645, + 0.003914745524525642, + 0.003876629052683711, + 0.00533161498606205, + 0.006140860728919506, + 0.0328042171895504, + 0.03891604393720627, + 0.03689928725361824, + 0.04119694605469704, + 0.043800465762615204, + 0.0674670860171318, + 0.026267338544130325, + 0.029418379068374634, + 0.013670763932168484, + 0.00910227745771408, + 0.006966507527977228, + 0.006871648598462343, + 0.009641999378800392, + 0.005200974643230438, + 0.006962741259485483, + 0.00695905601605773, + 0.017020130529999733, + 0.01820327155292034, + 0.016293048858642578, + 0.015969546511769295, + 0.011593377217650414, + 0.008990244008600712, + 0.011742765083909035, + 0.012504053302109241, + 0.002537545282393694, + 0.002040462801232934, + 0.0031819608993828297, + 0.0038765515200793743, + 0.004486286547034979, + 0.004656643141061068, + 0.0032159544061869383, + 0.004425069782882929, + 0.011910652741789818, + 0.013119558803737164, + 0.0253433920443058, + 0.007736138068139553, + 0.01796969212591648, + 0.00942038744688034, + 0.007135041058063507, + 0.006631579715758562, + 0.014982610940933228, + 0.011532846838235855, + 0.012651827186346054, + 0.014121475629508495, + 0.01384186651557684, + 0.013110661879181862, + 0.013493525795638561, + 0.007938277907669544, + 0.01519804447889328, + 0.01044628955423832, + 0.006883375346660614, + 0.01003422960639, + 0.013127204030752182, + 0.013408258557319641, + 0.010147140361368656, + 0.006673905998468399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.026353973895311356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.005258182529360056 + ], + "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": [ + 0.16631564497947693, + 0.06028028205037117, + 0.05037502944469452, + 0.20434802770614624, + 0.0922003835439682, + 0.07030616700649261, + 0.03029298596084118, + 0.15244419872760773, + 0.10706139355897903, + 0.06637585163116455 + ], + "credit_map": [ + 0.03307732939720154, + 0.038358960300683975, + 0.025629006326198578, + 0.018581582233309746, + 0.016682203859090805, + 0.012054771184921265, + 0.011300989426672459, + 0.010630796663463116, + 0.005974686238914728, + 0.01122566033154726, + 0.007645928300917149, + 0.007830443792045116, + 0.009617672301828861, + 0.0069311633706092834, + 0.006445007398724556, + 0.004609719850122929, + 0.0074627031572163105, + 0.007756564766168594, + 0.005861580837517977, + 0.0054398695938289165, + 0.004078006371855736, + 0.005082516465336084, + 0.008779969066381454, + 0.005913821514695883, + 0.027766691520810127, + 0.024122994393110275, + 0.026277434080839157, + 0.02443741075694561, + 0.024460388347506523, + 0.035267069935798645, + 0.017085697501897812, + 0.024930350482463837, + 0.018066875636577606, + 0.013970807194709778, + 0.009257912635803223, + 0.009835917502641678, + 0.010949729941785336, + 0.008785546757280827, + 0.008944996632635593, + 0.01238859724253416, + 0.006444175727665424, + 0.006176217924803495, + 0.009927946142852306, + 0.01760709472000599, + 0.00608466612175107, + 0.007429266348481178, + 0.007672728039324284, + 0.008964069187641144, + 0.003023749915882945, + 0.0032390074338763952, + 0.003146397415548563, + 0.0034829971846193075, + 0.004180503077805042, + 0.006290210876613855, + 0.003300959011539817, + 0.0036291605792939663, + 0.021939191967248917, + 0.0196998193860054, + 0.018083680421113968, + 0.009744795970618725, + 0.03403472527861595, + 0.018755044788122177, + 0.011692282743752003, + 0.018494660034775734, + 0.024845916777849197, + 0.016168002039194107, + 0.012250143103301525, + 0.011044997721910477, + 0.018567277118563652, + 0.007193562109023333, + 0.009748999960720539, + 0.0072424872778356075, + 0.010297928005456924, + 0.008910578675568104, + 0.008047686889767647, + 0.00993435550481081, + 0.008720037527382374, + 0.006665035150945187, + 0.007388981059193611, + 0.006411249749362469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.005258182529360056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, just above the turbulent sea, the camera moving parallel to the subject, captures a massive tank, clad in deep gunmetal armor, tearing through the deep blue, almost black, water with unstoppable force. The tank's bow-like front armor, etched with battle-worn scratches, rivets, and weathered rust, appears impenetrable. The camera's focus is pinpointed on the massive splash of water erupting from the tank's side at high speed. This white wall of water, nearly engulfing the camera, is composed of countless shimmering droplets and roiling foam. The slow-motion effect allows every splash to be clearly discerned, like a momentarily frozen liquid sculpture. The extremely shallow depth of field completely blurs the distant background, leaving only the faint outline of a gray, gloomy sky. The entire scene, illuminated by a cold, diffuse light against a gloomy sky, features a dominant palette of icy metallic gray, deep ocean blue, and the dazzling white of the waves, creating an atmosphere imbued with raw power, tension, and epic grandeur.", + "seed": 18341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_13/index_966_seed_18341_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": -0.06954798530787229, + "vlm_logprob": -0.06954798530787229, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.06954798530787229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.2574019134044647 + ], + "opposite_logprob_tokens": [ + -1.507401943206787 + ], + "logprob_reward": -0.2574019134044647, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2574019134044647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0628380998969078, + 0.10196762531995773, + 0.03841938450932503, + 0.06636174023151398, + 0.3074645400047302, + 0.044942766427993774, + 0.05019740015268326, + 0.160352423787117, + 0.07598567754030228, + 0.0914703756570816 + ], + "credit_map": [ + 0.007487111259251833, + 0.006083240732550621, + 0.01049085520207882, + 0.007708048913627863, + 0.008958176709711552, + 0.009079535491764545, + 0.007530653849244118, + 0.005500475410372019, + 0.006832565646618605, + 0.016819963231682777, + 0.01211480237543583, + 0.005730990320444107, + 0.028017345815896988, + 0.018856724724173546, + 0.007755957078188658, + 0.005839263554662466, + 0.0030213866848498583, + 0.007707509212195873, + 0.006434067152440548, + 0.003973451908677816, + 0.004869470372796059, + 0.0046981582418084145, + 0.004188939929008484, + 0.003526397980749607, + 0.00713525153696537, + 0.008429442532360554, + 0.008773431181907654, + 0.007938423193991184, + 0.010002337396144867, + 0.007466437760740519, + 0.010310771875083447, + 0.006305640563368797, + 0.028363557532429695, + 0.06766177713871002, + 0.023078376427292824, + 0.028440706431865692, + 0.0788252055644989, + 0.03190075606107712, + 0.028590595349669456, + 0.020603537559509277, + 0.003650915576145053, + 0.0057084765285253525, + 0.00337282195687294, + 0.008952909149229527, + 0.005551508627831936, + 0.006038806866854429, + 0.00687399972230196, + 0.004793322179466486, + 0.003210377413779497, + 0.004949220921844244, + 0.011255367659032345, + 0.0055335331708192825, + 0.007860380224883556, + 0.008069599978625774, + 0.004598160274326801, + 0.004720755852758884, + 0.015561842359602451, + 0.019612913951277733, + 0.019150225445628166, + 0.02319534681737423, + 0.021207211539149284, + 0.016470396891236305, + 0.021244486793875694, + 0.023909972980618477, + 0.006431883201003075, + 0.005440861452370882, + 0.0047379108145833015, + 0.006844502408057451, + 0.006718288641422987, + 0.005821786820888519, + 0.0202142596244812, + 0.019776174798607826, + 0.008155351504683495, + 0.006095404736697674, + 0.0052045597694814205, + 0.025996169075369835, + 0.03177085146307945, + 0.00480742147192359, + 0.006643624510616064, + 0.002796987071633339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.2574019134044647 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.03812805563211441 + ], + "opposite_logprob_tokens": [ + -3.288128137588501 + ], + "logprob_reward": -0.03812805563211441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812805563211441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046979546546936035, + 0.09501790255308151, + 0.08447961509227753, + 0.050742220133543015, + 0.23410053551197052, + 0.03745744004845619, + 0.034742433577775955, + 0.039084650576114655, + 0.03865979611873627, + 0.3387358784675598 + ], + "credit_map": [ + 0.004883467685431242, + 0.004743778612464666, + 0.006470296531915665, + 0.0062159523367881775, + 0.006902971770614386, + 0.006060025654733181, + 0.007291099987924099, + 0.004411954432725906, + 0.00994920078665018, + 0.01631259173154831, + 0.011892409063875675, + 0.009097251109778881, + 0.01704416424036026, + 0.012659984640777111, + 0.008173872716724873, + 0.00988843198865652, + 0.005947668105363846, + 0.01112012006342411, + 0.012076787650585175, + 0.01385818887501955, + 0.011921566911041737, + 0.015547956340014935, + 0.00927904061973095, + 0.004728282336145639, + 0.005517578683793545, + 0.00738814240321517, + 0.006302351132035255, + 0.005832988768815994, + 0.00839407928287983, + 0.005307684652507305, + 0.007096074987202883, + 0.004903318360447884, + 0.024068240076303482, + 0.0556490533053875, + 0.01743750087916851, + 0.020914889872074127, + 0.05712401866912842, + 0.023762434720993042, + 0.01949695125222206, + 0.015647446736693382, + 0.004056824371218681, + 0.005516733042895794, + 0.004093424417078495, + 0.006380052771419287, + 0.004444090649485588, + 0.0036988090723752975, + 0.005842949263751507, + 0.003424556227400899, + 0.002462314674630761, + 0.0026314235292375088, + 0.003998007625341415, + 0.008026557974517345, + 0.0032071403693407774, + 0.0024382220581173897, + 0.007666822988539934, + 0.004311941098421812, + 0.004083107691258192, + 0.00535194156691432, + 0.005263797007501125, + 0.0053530908189713955, + 0.004530859179794788, + 0.0042150202207267284, + 0.004983185790479183, + 0.00530364690348506, + 0.0035674406681209803, + 0.003134749596938491, + 0.0045783487148582935, + 0.0035569407045841217, + 0.003050965489819646, + 0.00282183475792408, + 0.007329181302338839, + 0.010620336048305035, + 0.02894945628941059, + 0.024747109040617943, + 0.01761915534734726, + 0.07445752620697021, + 0.12478289008140564, + 0.01953141577541828, + 0.02282108925282955, + 0.02582724392414093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.03812805563211441 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.012553717941045761 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553717941045761, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553717941045761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13185198605060577, + 0.13050319254398346, + 0.05794538930058479, + 0.0823785737156868, + 0.17639921605587006, + 0.07707984000444412, + 0.09578169882297516, + 0.04031582176685333, + 0.02380090020596981, + 0.18394339084625244 + ], + "credit_map": [ + 0.012960701249539852, + 0.016175856813788414, + 0.026369890198111534, + 0.01356843113899231, + 0.017436543479561806, + 0.015517733991146088, + 0.018138207495212555, + 0.011684633791446686, + 0.00854086596518755, + 0.013490879908204079, + 0.012645446695387363, + 0.006681491620838642, + 0.04158056899905205, + 0.02889544516801834, + 0.009692572057247162, + 0.008975929580628872, + 0.004583757370710373, + 0.011552185751497746, + 0.008974963799118996, + 0.006309207528829575, + 0.008450317196547985, + 0.006952799391001463, + 0.006108502857387066, + 0.005013657733798027, + 0.010289223864674568, + 0.014496097341179848, + 0.010618012398481369, + 0.007693267427384853, + 0.01271446980535984, + 0.0076889097690582275, + 0.010867124423384666, + 0.00801147148013115, + 0.018079649657011032, + 0.03605801239609718, + 0.013376167975366116, + 0.018788794055581093, + 0.039089612662792206, + 0.019033314660191536, + 0.017359895631670952, + 0.014613796956837177, + 0.006926670204848051, + 0.013230744749307632, + 0.00882718712091446, + 0.013824278488755226, + 0.01093631237745285, + 0.007283387705683708, + 0.009821722283959389, + 0.00622953474521637, + 0.008281766436994076, + 0.009365803562104702, + 0.021407999098300934, + 0.009560958482325077, + 0.010590272955596447, + 0.008893259800970554, + 0.014625374227762222, + 0.013056260533630848, + 0.0034023320768028498, + 0.004139956086874008, + 0.004233011044561863, + 0.008035961538553238, + 0.004352965857833624, + 0.0037304398138076067, + 0.007009680382907391, + 0.005411470774561167, + 0.0034932196140289307, + 0.002685128478333354, + 0.003766599576920271, + 0.002469196217134595, + 0.002863061847165227, + 0.0026997423265129328, + 0.0024685789830982685, + 0.00335537432692945, + 0.016477808356285095, + 0.011459380388259888, + 0.01110004261136055, + 0.05390535667538643, + 0.060011643916368484, + 0.00870173703879118, + 0.015108391642570496, + 0.0071790385991334915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.012553717941045761 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.03370363265275955 + ], + "opposite_logprob_tokens": [ + -3.408703565597534 + ], + "logprob_reward": -0.03370363265275955, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370363265275955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0949801579117775, + 0.14156031608581543, + 0.06441020965576172, + 0.07008043676614761, + 0.3908616006374359, + 0.048880401998758316, + 0.03544781357049942, + 0.07354496419429779, + 0.039889004081487656, + 0.04034510254859924 + ], + "credit_map": [ + 0.008057564496994019, + 0.010485517792403698, + 0.01670514978468418, + 0.018584761768579483, + 0.007775587495416403, + 0.011139577254652977, + 0.014772196300327778, + 0.007459803018718958, + 0.0139021510258317, + 0.012461393140256405, + 0.02490733191370964, + 0.012195003218948841, + 0.03152250871062279, + 0.019793836399912834, + 0.012050029821693897, + 0.014728058129549026, + 0.005383474286645651, + 0.010545964352786541, + 0.01025307085365057, + 0.006712498143315315, + 0.012260802090168, + 0.007418186869472265, + 0.006644070148468018, + 0.005192140582948923, + 0.005038293078541756, + 0.012231063097715378, + 0.008836698718369007, + 0.007772150915116072, + 0.009095621295273304, + 0.009093309752643108, + 0.009532696567475796, + 0.008480602875351906, + 0.0368381142616272, + 0.08727242797613144, + 0.028591793030500412, + 0.03619498759508133, + 0.09875018894672394, + 0.04083385691046715, + 0.03454696014523506, + 0.027833260595798492, + 0.004585422109812498, + 0.00976977963000536, + 0.005864844191819429, + 0.0068389372900128365, + 0.00677716126665473, + 0.006822545081377029, + 0.004672968294471502, + 0.0035487462300807238, + 0.003423571353778243, + 0.003189040580764413, + 0.004586219787597656, + 0.004164134152233601, + 0.005204115994274616, + 0.0060771917924284935, + 0.0036235095467418432, + 0.005180030595511198, + 0.007780494634062052, + 0.011011602357029915, + 0.008018060587346554, + 0.009780251421034336, + 0.009509969502687454, + 0.009523703716695309, + 0.010292310267686844, + 0.007628577761352062, + 0.004172143526375294, + 0.004395927302539349, + 0.005413397215306759, + 0.005424707196652889, + 0.004237364511936903, + 0.003785970387980342, + 0.005454381927847862, + 0.007005113642662764, + 0.004283400718122721, + 0.0029989355243742466, + 0.0035994648933410645, + 0.011421358212828636, + 0.006571473088115454, + 0.0033437060192227364, + 0.005261713173240423, + 0.0028650483582168818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.03370363265275955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 tank clad in deep gunmetal armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank moving through deep blue, almost black water with unstoppable force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned in a close-up shot from a low angle just above the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera moving parallel to the subject while focusing on a massive splash erupting from the tank's side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene rendered with a slow-motion effect and extremely shallow depth of field that blurs the distant 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": [ + -0.005952606908977032 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952606908977032, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952606908977032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11556557565927505, + 0.05787919461727142, + 0.05995273217558861, + 0.0671730637550354, + 0.3204840421676636, + 0.06889551877975464, + 0.05538451671600342, + 0.06874512135982513, + 0.05353708937764168, + 0.13238315284252167 + ], + "credit_map": [ + 0.011918531730771065, + 0.014057927764952183, + 0.014670961536467075, + 0.013632260262966156, + 0.020571354776620865, + 0.013979601673781872, + 0.017596855759620667, + 0.009138076566159725, + 0.003455479396507144, + 0.009459230117499828, + 0.008234762586653233, + 0.005010680295526981, + 0.010747908614575863, + 0.0074525391682982445, + 0.005542757920920849, + 0.007975834421813488, + 0.004158827941864729, + 0.01248515397310257, + 0.009505274705588818, + 0.006993100978434086, + 0.008657079190015793, + 0.008368992246687412, + 0.006181331817060709, + 0.0036029675975441933, + 0.008037883788347244, + 0.010360434651374817, + 0.0074590714648365974, + 0.006104602944105864, + 0.01118335872888565, + 0.005858088843524456, + 0.008658112026751041, + 0.00951151642948389, + 0.03236577659845352, + 0.06930030137300491, + 0.023886816576123238, + 0.03264661878347397, + 0.07524125277996063, + 0.032020922750234604, + 0.03085414320230484, + 0.024168211966753006, + 0.006285212002694607, + 0.01019293162971735, + 0.008301760070025921, + 0.011450966820120811, + 0.00845288299024105, + 0.006963553838431835, + 0.010739458724856377, + 0.00650875736027956, + 0.004507048055529594, + 0.005153896287083626, + 0.012480027042329311, + 0.005313858389854431, + 0.006694104056805372, + 0.007354832254350185, + 0.006328815594315529, + 0.007551931776106358, + 0.006539718247950077, + 0.008654825389385223, + 0.008416250348091125, + 0.009668388403952122, + 0.006904008332639933, + 0.007219719234853983, + 0.010691286996006966, + 0.010650926269590855, + 0.004689360968768597, + 0.007974033243954182, + 0.005581664852797985, + 0.006608147174119949, + 0.0064293998293578625, + 0.007559924852102995, + 0.008842568844556808, + 0.005851986818015575, + 0.012356522493064404, + 0.007727393880486488, + 0.008334096521139145, + 0.03786223381757736, + 0.03928466513752937, + 0.007447392214089632, + 0.012559745460748672, + 0.0068111056461930275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.005952606908977032 + } + ], + "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/0000058/group_14/index_tensor([5757, 5757])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_14/index_tensor([5757, 5757])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11e6dc3a046e96e096ffccffda083177c2837f34 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_14/index_tensor([5757, 5757])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This handheld, slightly shaky close-up begins by focusing on the back of a small child's head, where fine, light brown hair is visible. The child's face is completely buried in the thick, dark gray canvas of their parent's jacket, their tiny body pressed tightly against their parent's chest for safety. The camera slowly and steadily pulls back to a medium shot, revealing the parent's arms tightly and protectively wrapped around the child, their tiny hands clenched into fists, desperately clinging to the jacket's fabric. They are engulfed by a surging crowd at a sunny outdoor festival. The background is vibrant, out-of-focus, with movement and color, filled with cheering people raising their hands, waving colorful flags, and throwing confetti that glitters in the bright afternoon sun. The density and flow of the crowd creates a palpable sense of pressure, contrasting sharply with the intimate, protected space created by the parents, establishing a poignant atmosphere of childhood fragility amidst the overwhelming public cheering.", + "seed": 15634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_14/index_5757_seed_15634_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": -0.09406309127807617, + "vlm_logprob": -0.09406309127807617, + "vlm_yes_no_margin": 2.424999976158142, + "vlm_token_credit": -0.09406309127807617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05628272145986557 + ], + "opposite_logprob_tokens": [ + -3.0562827587127686 + ], + "logprob_reward": -0.05628272145986557, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05628272145986557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16234521567821503, + 0.1465391367673874, + 0.12497072666883469, + 0.05395295470952988, + 0.13592922687530518, + 0.08181798458099365, + 0.04104236513376236, + 0.1050148755311966, + 0.11552486568689346, + 0.03286257013678551 + ], + "credit_map": [ + 0.018987154588103294, + 0.0121082728728652, + 0.02676619589328766, + 0.031097421422600746, + 0.018616247922182083, + 0.018381882458925247, + 0.01934795081615448, + 0.017040105536580086, + 0.008333778008818626, + 0.00911131501197815, + 0.029229136183857918, + 0.03126191347837448, + 0.01783405989408493, + 0.01590871252119541, + 0.02262168377637863, + 0.012238546274602413, + 0.012187127955257893, + 0.017167672514915466, + 0.013447891920804977, + 0.013563142158091068, + 0.014560450799763203, + 0.02112331986427307, + 0.018730295822024345, + 0.014190834946930408, + 0.004139020573347807, + 0.005600352771580219, + 0.006329361815005541, + 0.008946675807237625, + 0.007810727693140507, + 0.00916994083672762, + 0.006586012896150351, + 0.005370865575969219, + 0.01593557931482792, + 0.013753770850598812, + 0.013392738066613674, + 0.022922907024621964, + 0.01634305529296398, + 0.012368504889309406, + 0.020396940410137177, + 0.020815759897232056, + 0.01041876059025526, + 0.008240348659455776, + 0.007908615283668041, + 0.010169761255383492, + 0.013552688993513584, + 0.009603344835340977, + 0.010501601733267307, + 0.011422866955399513, + 0.006014181766659021, + 0.004469512030482292, + 0.004370883107185364, + 0.0032509975135326385, + 0.0033214439172297716, + 0.004839648026973009, + 0.009150981903076172, + 0.00562472315505147, + 0.012744364328682423, + 0.007258475758135319, + 0.013768995180726051, + 0.013044564984738827, + 0.006482755299657583, + 0.007674291729927063, + 0.023902028799057007, + 0.020139411091804504, + 0.013865339569747448, + 0.010287283919751644, + 0.0148725351318717, + 0.015180990099906921, + 0.011781467124819756, + 0.010758369229733944, + 0.014370663091540337, + 0.02440822497010231, + 0.008513572625815868, + 0.004802469164133072, + 0.005378975532948971, + 0.003307651961222291, + 0.0025069250259548426, + 0.0027505001053214073, + 0.002881197491660714, + 0.002721284283325076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05628272145986557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909071981906891 + ], + "opposite_logprob_tokens": [ + -2.4640908241271973 + ], + "logprob_reward": -0.08909071981906891, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909071981906891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23092174530029297, + 0.09520364552736282, + 0.154587984085083, + 0.044093333184719086, + 0.05357575789093971, + 0.017128318548202515, + 0.27260634303092957, + 0.05817051976919174, + 0.055876147001981735, + 0.017836149781942368 + ], + "credit_map": [ + 0.02642040327191353, + 0.016937287524342537, + 0.030640048906207085, + 0.04010350629687309, + 0.031791362911462784, + 0.02277381718158722, + 0.024512432515621185, + 0.037742890417575836, + 0.0059916190803050995, + 0.008594444952905178, + 0.020149020478129387, + 0.02024860307574272, + 0.008159545250236988, + 0.009296132251620293, + 0.017069624736905098, + 0.005694665480405092, + 0.014561350457370281, + 0.0208608265966177, + 0.016148945316672325, + 0.02283485420048237, + 0.019149499014019966, + 0.02206501178443432, + 0.02122524194419384, + 0.01774226315319538, + 0.0037378196138888597, + 0.0055253212340176105, + 0.0036937736440449953, + 0.011515550315380096, + 0.004711560904979706, + 0.004981874022632837, + 0.005097752436995506, + 0.004829687997698784, + 0.007134970277547836, + 0.005154937971383333, + 0.006124953273683786, + 0.0078077903017401695, + 0.005809274036437273, + 0.0044868760742247105, + 0.007718478795140982, + 0.009338479489088058, + 0.0019730860367417336, + 0.0014607558259740472, + 0.00194288301281631, + 0.002240458969026804, + 0.0022779370192438364, + 0.002051158342510462, + 0.002952871611341834, + 0.002229167614132166, + 0.05929410830140114, + 0.01353405136615038, + 0.010479448363184929, + 0.01491917297244072, + 0.021062199026346207, + 0.005686653312295675, + 0.11286680400371552, + 0.03476391360163689, + 0.010828160680830479, + 0.003474382683634758, + 0.0035887404810637236, + 0.006317699793726206, + 0.00418971898034215, + 0.0034353723749518394, + 0.01917761005461216, + 0.007158835884183645, + 0.008042072877287865, + 0.006332018878310919, + 0.006485610269010067, + 0.006633691024035215, + 0.006204281933605671, + 0.005475662648677826, + 0.006936678197234869, + 0.009766136296093464, + 0.002675557741895318, + 0.001938707660883665, + 0.0028607193380594254, + 0.0017189724603667855, + 0.0016653859056532383, + 0.0019023417262360454, + 0.0028711522463709116, + 0.002203314332291484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.08909071981906891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054964564740657806 + ], + "opposite_logprob_tokens": [ + -2.929964542388916 + ], + "logprob_reward": -0.054964564740657806, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054964564740657806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09873542934656143, + 0.12689289450645447, + 0.05016142502427101, + 0.10370106250047684, + 0.11590920388698578, + 0.0878826454281807, + 0.2896149158477783, + 0.05468067526817322, + 0.05492284893989563, + 0.01749885082244873 + ], + "credit_map": [ + 0.008820733055472374, + 0.008475765585899353, + 0.015513203106820583, + 0.014914163388311863, + 0.01056603342294693, + 0.0067633152939379215, + 0.02385067567229271, + 0.009831538423895836, + 0.009301110170781612, + 0.008383607491850853, + 0.02129715494811535, + 0.02535415254533291, + 0.015420154668390751, + 0.015825312584638596, + 0.018879378214478493, + 0.012432023882865906, + 0.0043180170468986034, + 0.006656038109213114, + 0.007467474322766066, + 0.009028573520481586, + 0.005268588662147522, + 0.004532909952104092, + 0.00820345152169466, + 0.004686372354626656, + 0.009158414788544178, + 0.01219308190047741, + 0.009290964342653751, + 0.021986350417137146, + 0.01506535429507494, + 0.014750014990568161, + 0.011468496173620224, + 0.009788389317691326, + 0.012575703673064709, + 0.011440236121416092, + 0.013918597251176834, + 0.020816778764128685, + 0.010599526576697826, + 0.011868185363709927, + 0.017746174708008766, + 0.01694399118423462, + 0.01029256172478199, + 0.00885689351707697, + 0.00831796508282423, + 0.008519052527844906, + 0.015364347957074642, + 0.010508112609386444, + 0.012528308667242527, + 0.013495405204594135, + 0.06419267505407333, + 0.017135420814156532, + 0.011884289793670177, + 0.01606164500117302, + 0.021186018362641335, + 0.010182433761656284, + 0.11267771571874619, + 0.036294713616371155, + 0.008099337108433247, + 0.004483045544475317, + 0.005329919047653675, + 0.006918397266417742, + 0.003702344372868538, + 0.0030197766609489918, + 0.016519518569111824, + 0.00660833390429616, + 0.011288614012300968, + 0.008061311207711697, + 0.006853974889963865, + 0.004849896300584078, + 0.0062658460810780525, + 0.005570828914642334, + 0.006370236165821552, + 0.005662138573825359, + 0.002378120319917798, + 0.0017208169447258115, + 0.003098217537626624, + 0.0014849139843136072, + 0.0018466124311089516, + 0.0019373672548681498, + 0.0024269355926662683, + 0.0026058671064674854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.054964564740657806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12712553143501282 + ], + "opposite_logprob_tokens": [ + -2.1271255016326904 + ], + "logprob_reward": -0.12712553143501282, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12712553143501282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19090111553668976, + 0.060268647968769073, + 0.035049863159656525, + 0.08217426389455795, + 0.1486538201570511, + 0.033444635570049286, + 0.08027469366788864, + 0.09061029553413391, + 0.22007347643375397, + 0.058549072593450546 + ], + "credit_map": [ + 0.012066849507391453, + 0.018709696829319, + 0.033435311168432236, + 0.03825218230485916, + 0.018047694116830826, + 0.010772430337965488, + 0.04514269158244133, + 0.014474262483417988, + 0.004648827016353607, + 0.003350273473188281, + 0.0074208262376487255, + 0.009108891710639, + 0.006307321134954691, + 0.007904646918177605, + 0.009242712520062923, + 0.012285150587558746, + 0.0030801298562437296, + 0.003181174863129854, + 0.005326476413756609, + 0.005095472559332848, + 0.0027153827250003815, + 0.003752810414880514, + 0.007059505674988031, + 0.004838914610445499, + 0.0077843922190368176, + 0.00934574194252491, + 0.008453276008367538, + 0.021874794736504555, + 0.009092877618968487, + 0.008152562193572521, + 0.009598380886018276, + 0.007872235961258411, + 0.016628816723823547, + 0.013883103616535664, + 0.014953583478927612, + 0.025319693610072136, + 0.01869601011276245, + 0.012415498495101929, + 0.02274821698665619, + 0.024008899927139282, + 0.0033191151451319456, + 0.003605083329603076, + 0.003436116501688957, + 0.003907974809408188, + 0.004925318527966738, + 0.004705141298472881, + 0.004990999586880207, + 0.00455488683655858, + 0.01054114568978548, + 0.005936411675065756, + 0.0069711944088339806, + 0.007637197617441416, + 0.005925972014665604, + 0.010869971476495266, + 0.021659981459379196, + 0.010732823982834816, + 0.01119458582252264, + 0.006991443689912558, + 0.009661165066063404, + 0.009184560738503933, + 0.00606779707595706, + 0.014982731081545353, + 0.021311132237315178, + 0.011216888204216957, + 0.03601592034101486, + 0.02426908351480961, + 0.02771541103720665, + 0.035130973905324936, + 0.015363459475338459, + 0.014407128095626831, + 0.03497700393199921, + 0.032194506376981735, + 0.012718887068331242, + 0.006936215329915285, + 0.011539113707840443, + 0.007366655394434929, + 0.0037772194482386112, + 0.004982712212949991, + 0.005651318933814764, + 0.005576954688876867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.12712553143501282 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285191893577576 + ], + "opposite_logprob_tokens": [ + -2.0178518295288086 + ], + "logprob_reward": -0.14285191893577576, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14285191893577576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14136965572834015, + 0.10878856480121613, + 0.346190869808197, + 0.04062632471323013, + 0.04751748591661453, + 0.027637816965579987, + 0.12686528265476227, + 0.04163835942745209, + 0.08879740536212921, + 0.030568188056349754 + ], + "credit_map": [ + 0.012906990945339203, + 0.010155018419027328, + 0.020099926739931107, + 0.023984085768461227, + 0.017563847824931145, + 0.01483945269137621, + 0.02097517065703869, + 0.020845161750912666, + 0.007175609935075045, + 0.006645009852945805, + 0.01573832333087921, + 0.021524589508771896, + 0.014613288454711437, + 0.012152720242738724, + 0.021426865831017494, + 0.009512156248092651, + 0.03578135371208191, + 0.049383822828531265, + 0.04391803592443466, + 0.049432069063186646, + 0.04000003635883331, + 0.04187664017081261, + 0.057188816368579865, + 0.028610073029994965, + 0.0035102295223623514, + 0.0060325623489916325, + 0.0034350829664617777, + 0.009700700640678406, + 0.004372593015432358, + 0.0057784379459917545, + 0.0036602728068828583, + 0.004136444069445133, + 0.0047586578875780106, + 0.003876020200550556, + 0.00526333786547184, + 0.007741680834442377, + 0.007530967239290476, + 0.0034779291599988937, + 0.008348957635462284, + 0.006519936490803957, + 0.003409444121643901, + 0.0028266054578125477, + 0.0035081387031823397, + 0.00413726270198822, + 0.0029696710407733917, + 0.003164118155837059, + 0.003957832232117653, + 0.003664744086563587, + 0.02602461725473404, + 0.006606339477002621, + 0.005740870721638203, + 0.007247242610901594, + 0.009887651540338993, + 0.003972776234149933, + 0.051786500960588455, + 0.015599288046360016, + 0.005585963372141123, + 0.003998260945081711, + 0.004433291032910347, + 0.00503518944606185, + 0.004147294908761978, + 0.004492032341659069, + 0.008928653784096241, + 0.005017674528062344, + 0.015262131579220295, + 0.010943478904664516, + 0.010946832597255707, + 0.010681450366973877, + 0.010980424471199512, + 0.007601507008075714, + 0.011496461927890778, + 0.010885112918913364, + 0.00565841281786561, + 0.0032257011625915766, + 0.005481189116835594, + 0.0035337237641215324, + 0.0025614190381020308, + 0.0026104880962520838, + 0.004166091326624155, + 0.0033311648294329643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.14285191893577576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This handheld, slightly shaky close-up begins by focusing on the back of a small child's head, where fine, light brown hair is visible. The child's face is completely buried in the thick, dark gray canvas of their parent's jacket, their tiny body pressed tightly against their parent's chest for safety. The camera slowly and steadily pulls back to a medium shot, revealing the parent's arms tightly and protectively wrapped around the child, their tiny hands clenched into fists, desperately clinging to the jacket's fabric. They are engulfed by a surging crowd at a sunny outdoor festival. The background is vibrant, out-of-focus, with movement and color, filled with cheering people raising their hands, waving colorful flags, and throwing confetti that glitters in the bright afternoon sun. The density and flow of the crowd creates a palpable sense of pressure, contrasting sharply with the intimate, protected space created by the parents, establishing a poignant atmosphere of childhood fragility amidst the overwhelming public cheering.", + "seed": 15635, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_14/index_5757_seed_15635_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": -0.07409666366875171, + "vlm_logprob": -0.07409666366875171, + "vlm_yes_no_margin": 3.1249999523162844, + "vlm_token_credit": -0.07409666366875171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022880129516124725 + ], + "opposite_logprob_tokens": [ + -4.2728800773620605 + ], + "logprob_reward": -0.022880129516124725, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.022880129516124725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07061011344194412, + 0.1504868119955063, + 0.057280637323856354, + 0.026111148297786713, + 0.033021796494722366, + 0.016991306096315384, + 0.03408509120345116, + 0.5419957041740417, + 0.028059469535946846, + 0.04135792702436447 + ], + "credit_map": [ + 0.009544102475047112, + 0.008215103298425674, + 0.01003691554069519, + 0.008934291079640388, + 0.007565902546048164, + 0.011749057099223137, + 0.0064364769496023655, + 0.008128263056278229, + 0.03830759599804878, + 0.023375190794467926, + 0.050915222615003586, + 0.009502478875219822, + 0.007938391529023647, + 0.0086891520768404, + 0.006649543531239033, + 0.005109224934130907, + 0.005689659621566534, + 0.010001366958022118, + 0.01071025338023901, + 0.005506688263267279, + 0.006442114245146513, + 0.00740445964038372, + 0.005618513096123934, + 0.005907576065510511, + 0.002453887602314353, + 0.005172476638108492, + 0.0034809503704309464, + 0.003241444006562233, + 0.0020809045527130365, + 0.0031640869565308094, + 0.00420603109523654, + 0.0023113645147532225, + 0.0036814024206250906, + 0.005133288446813822, + 0.0034228714648634195, + 0.0041849794797599316, + 0.004363056272268295, + 0.0025515540037304163, + 0.006184604950249195, + 0.003500034799799323, + 0.0013723056763410568, + 0.0032825847156345844, + 0.0016492835711687803, + 0.0021424151491373777, + 0.002410830929875374, + 0.0023800369817763567, + 0.001685542636550963, + 0.002068305853754282, + 0.005150881130248308, + 0.006777779199182987, + 0.004957042168825865, + 0.004704966209828854, + 0.0031513595022261143, + 0.0038299602456390858, + 0.002732860157266259, + 0.002780238864943385, + 0.0638120248913765, + 0.029243402183055878, + 0.17995519936084747, + 0.11120285093784332, + 0.029178211465477943, + 0.07560938596725464, + 0.02272162213921547, + 0.030272984877228737, + 0.0032493637409061193, + 0.0023047584109008312, + 0.007181232329457998, + 0.005989319644868374, + 0.0020910524763166904, + 0.0027773177716881037, + 0.0021860492415726185, + 0.002280374988913536, + 0.004025309346616268, + 0.0033588295336812735, + 0.007076343987137079, + 0.007371147163212299, + 0.004418882541358471, + 0.0051729194819927216, + 0.007975825108587742, + 0.001958664273843169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.022880129516124725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007049769163132 + ], + "opposite_logprob_tokens": [ + -2.695070505142212 + ], + "logprob_reward": -0.07007049769163132, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007049769163132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040736857801675797, + 0.11144866794347763, + 0.029726147651672363, + 0.031460121273994446, + 0.03607359156012535, + 0.0775502473115921, + 0.021421965211629868, + 0.5173517465591431, + 0.07890094071626663, + 0.05532971769571304 + ], + "credit_map": [ + 0.004036729224026203, + 0.005836790893226862, + 0.006298661231994629, + 0.004726053681224585, + 0.00538589945062995, + 0.006235902663320303, + 0.004374356474727392, + 0.0038424613885581493, + 0.026059094816446304, + 0.019630836322903633, + 0.03978824242949486, + 0.007865565828979015, + 0.005367387551814318, + 0.004085937049239874, + 0.005189902614802122, + 0.003461697371676564, + 0.0025477632880210876, + 0.006608838681131601, + 0.00638532405719161, + 0.003334819572046399, + 0.003249370027333498, + 0.002792536513879895, + 0.002587872324511409, + 0.002219622489064932, + 0.0028525309171527624, + 0.004455297254025936, + 0.0050541795790195465, + 0.002971722511574626, + 0.004043521825224161, + 0.004679747857153416, + 0.003481137566268444, + 0.003921987023204565, + 0.004732797853648663, + 0.0031962732318788767, + 0.00391432037577033, + 0.0037555804010480642, + 0.0034456588327884674, + 0.0044164410792291164, + 0.00288555514998734, + 0.009726964868605137, + 0.004319352563470602, + 0.009040088392794132, + 0.006875938735902309, + 0.009249445982277393, + 0.01673603057861328, + 0.010494103655219078, + 0.011674540117383003, + 0.009160742163658142, + 0.003111396450549364, + 0.00441565178334713, + 0.0031309763435274363, + 0.0035536703653633595, + 0.0021137716248631477, + 0.002121282508596778, + 0.001518893288448453, + 0.0014563219156116247, + 0.058855462819337845, + 0.029323527589440346, + 0.16625533998012543, + 0.10724048316478729, + 0.03209740296006203, + 0.06957677751779556, + 0.024777116253972054, + 0.029225636273622513, + 0.00869041122496128, + 0.005559853278100491, + 0.017696725204586983, + 0.0144202196970582, + 0.005888768937438726, + 0.009949189610779285, + 0.006932786665856838, + 0.009762982837855816, + 0.004741695709526539, + 0.003931659273803234, + 0.010467085056006908, + 0.010735360905528069, + 0.00536107225343585, + 0.0065543027594685555, + 0.010057731531560421, + 0.0034808057826012373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.07007049769163132 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041612543165683746, + 0.024500494822859764, + 0.03300895169377327, + 0.032009925693273544, + 0.03737715631723404, + 0.03658360242843628, + 0.027036290615797043, + 0.6013163924217224, + 0.1429748684167862, + 0.02357974462211132 + ], + "credit_map": [ + 0.005194589961320162, + 0.004176601767539978, + 0.005564131308346987, + 0.006127409171313047, + 0.004348218906670809, + 0.0066342963837087154, + 0.004542103968560696, + 0.005025193560868502, + 0.004426281899213791, + 0.003845471888780594, + 0.0025348819326609373, + 0.002800515852868557, + 0.003790206043049693, + 0.002939782105386257, + 0.0024420705158263445, + 0.0017212845850735903, + 0.003327983431518078, + 0.005212479270994663, + 0.006192913744598627, + 0.002179030328989029, + 0.004602327942848206, + 0.005232162307947874, + 0.003125088522210717, + 0.0031369663774967194, + 0.0032973396591842175, + 0.005940019153058529, + 0.004693065769970417, + 0.0037787954788655043, + 0.003074584063142538, + 0.0040823244489729404, + 0.004458287730813026, + 0.00268550799228251, + 0.0037720906548202038, + 0.00517653115093708, + 0.003793406067416072, + 0.004319086670875549, + 0.005321559030562639, + 0.003822327358648181, + 0.006858557928353548, + 0.004313596058636904, + 0.0045357076451182365, + 0.0043709129095077515, + 0.0032454864121973515, + 0.0042808325961232185, + 0.005463483743369579, + 0.0040439278818666935, + 0.005650725215673447, + 0.004992523696273565, + 0.004886578768491745, + 0.004881765693426132, + 0.0038194211665540934, + 0.0038894216995686293, + 0.0024640916381031275, + 0.0024389298632740974, + 0.0023211941588670015, + 0.0023348890244960785, + 0.07098227739334106, + 0.032935477793216705, + 0.1987205594778061, + 0.12222743034362793, + 0.03288675472140312, + 0.08477354049682617, + 0.02517867274582386, + 0.03361168131232262, + 0.018385615199804306, + 0.011692574247717857, + 0.0359581783413887, + 0.02993263490498066, + 0.010276565328240395, + 0.014468688517808914, + 0.011279691942036152, + 0.010980922728776932, + 0.0025047592353075743, + 0.0024706567637622356, + 0.0035866720136255026, + 0.004004740621894598, + 0.0026654284447431564, + 0.0030403181444853544, + 0.003573467954993248, + 0.001733700861223042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.01819819025695324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371607884764671 + ], + "opposite_logprob_tokens": [ + -3.4087159633636475 + ], + "logprob_reward": -0.03371607884764671, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371607884764671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.086274154484272, + 0.06917823106050491, + 0.028508353978395462, + 0.0320039838552475, + 0.030809899792075157, + 0.01648213155567646, + 0.03623885288834572, + 0.6550105214118958, + 0.02193302847445011, + 0.023560836911201477 + ], + "credit_map": [ + 0.008452478796243668, + 0.009035577066242695, + 0.011021416634321213, + 0.007395477965474129, + 0.007821956649422646, + 0.024965953081846237, + 0.008358589373528957, + 0.009222703985869884, + 0.015449749305844307, + 0.012423429638147354, + 0.024824799969792366, + 0.005014712456613779, + 0.0021219535265117884, + 0.003935743123292923, + 0.0031238056253641844, + 0.0022840360179543495, + 0.0033981511369347572, + 0.004382895305752754, + 0.006015710067003965, + 0.002476951340213418, + 0.003573805559426546, + 0.0036796710919588804, + 0.0025557323824614286, + 0.0024254375603049994, + 0.003061629831790924, + 0.0038248123601078987, + 0.006385565735399723, + 0.003079163609072566, + 0.0028831122908741236, + 0.005219253711402416, + 0.0033708561677485704, + 0.0041795894503593445, + 0.0031330138444900513, + 0.00394184235483408, + 0.0028799953870475292, + 0.0035966045688837767, + 0.0056565674021840096, + 0.0042383321560919285, + 0.0029178031254559755, + 0.004445743281394243, + 0.0014510826440528035, + 0.0018156343139708042, + 0.0013899861369282007, + 0.002289118943735957, + 0.0029594902880489826, + 0.0017269740346819162, + 0.0025437786243855953, + 0.0023060678504407406, + 0.004384285770356655, + 0.008245927281677723, + 0.005624047480523586, + 0.005849487613886595, + 0.0034148606937378645, + 0.0034458329901099205, + 0.002990201348438859, + 0.0022842115722596645, + 0.07883842289447784, + 0.035923343151807785, + 0.21560138463974, + 0.1282792091369629, + 0.03503933921456337, + 0.09608780592679977, + 0.027409544214606285, + 0.037831466645002365, + 0.0038228966295719147, + 0.002486548852175474, + 0.0033548707142472267, + 0.0025431241374462843, + 0.002533551538363099, + 0.002733821514993906, + 0.0023957465309649706, + 0.0020624694880098104, + 0.002310000592842698, + 0.0031273120548576117, + 0.0037777707912027836, + 0.003542081918567419, + 0.002581620356068015, + 0.0031247236765921116, + 0.003548253793269396, + 0.001549074426293373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.03371607884764671 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256184220314026 + ], + "opposite_logprob_tokens": [ + -1.6006183624267578 + ], + "logprob_reward": -0.2256184220314026, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256184220314026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03882492333650589, + 0.03634703531861305, + 0.04347255453467369, + 0.032075632363557816, + 0.07247620075941086, + 0.015101950615644455, + 0.07122969627380371, + 0.61396723985672, + 0.04068870097398758, + 0.03581607714295387 + ], + "credit_map": [ + 0.0035741685424000025, + 0.005021731834858656, + 0.00508939940482378, + 0.004923591855913401, + 0.00420983973890543, + 0.008887451142072678, + 0.0034120657946914434, + 0.0037066719960421324, + 0.008592427708208561, + 0.00768303032964468, + 0.00346267013810575, + 0.0038284757174551487, + 0.003955557011067867, + 0.004054514691233635, + 0.002192190382629633, + 0.002578167477622628, + 0.005273689515888691, + 0.005608297418802977, + 0.005688315723091364, + 0.005850057117640972, + 0.0066342889331281185, + 0.006796125788241625, + 0.004005113150924444, + 0.003616667352616787, + 0.0027829930186271667, + 0.004352568183094263, + 0.005551384296268225, + 0.004016638267785311, + 0.0028347333427518606, + 0.004324125126004219, + 0.005137268919497728, + 0.003075922606512904, + 0.0063744657672941685, + 0.011609907262027264, + 0.007437034044414759, + 0.008149816654622555, + 0.01387549377977848, + 0.008556079119443893, + 0.009600546211004257, + 0.006872858852148056, + 0.0011171288788318634, + 0.0023009879514575005, + 0.0014386645052582026, + 0.001582760363817215, + 0.002812405116856098, + 0.0022701777052134275, + 0.0021203202195465565, + 0.001459504826925695, + 0.007617144845426083, + 0.00947728380560875, + 0.012325146235525608, + 0.011108418926596642, + 0.005700049456208944, + 0.0061106435023248196, + 0.012148559093475342, + 0.006742450408637524, + 0.0733460932970047, + 0.03373969346284866, + 0.20356422662734985, + 0.1220674142241478, + 0.03206699714064598, + 0.08866608887910843, + 0.02542720176279545, + 0.03508955240249634, + 0.0044165621511638165, + 0.003195843892171979, + 0.011838843114674091, + 0.00944305956363678, + 0.002805474679917097, + 0.003114804159849882, + 0.0035501234233379364, + 0.0023239918518811464, + 0.004326322581619024, + 0.0036080866120755672, + 0.006699991412460804, + 0.006216061767190695, + 0.003366093849763274, + 0.004068037960678339, + 0.0055854786187410355, + 0.0019460031762719154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.2256184220314026 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This handheld, slightly shaky close-up begins by focusing on the back of a small child's head, where fine, light brown hair is visible. The child's face is completely buried in the thick, dark gray canvas of their parent's jacket, their tiny body pressed tightly against their parent's chest for safety. The camera slowly and steadily pulls back to a medium shot, revealing the parent's arms tightly and protectively wrapped around the child, their tiny hands clenched into fists, desperately clinging to the jacket's fabric. They are engulfed by a surging crowd at a sunny outdoor festival. The background is vibrant, out-of-focus, with movement and color, filled with cheering people raising their hands, waving colorful flags, and throwing confetti that glitters in the bright afternoon sun. The density and flow of the crowd creates a palpable sense of pressure, contrasting sharply with the intimate, protected space created by the parents, establishing a poignant atmosphere of childhood fragility amidst the overwhelming public cheering.", + "seed": 15636, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_14/index_5757_seed_15636_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": -0.1546495661139488, + "vlm_logprob": -0.1546495661139488, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.1546495661139488, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03889887407422066 + ], + "opposite_logprob_tokens": [ + -3.6638989448547363 + ], + "logprob_reward": -0.03889887407422066, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03889887407422066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14766481518745422, + 0.04486745595932007, + 0.07758283615112305, + 0.059096306562423706, + 0.11071278899908066, + 0.04381503909826279, + 0.054742418229579926, + 0.1438700258731842, + 0.10599339753389359, + 0.21165494620800018 + ], + "credit_map": [ + 0.011033724993467331, + 0.015703776851296425, + 0.011814604513347149, + 0.018774814903736115, + 0.018013514578342438, + 0.021489417180418968, + 0.028757305815815926, + 0.022077644243836403, + 0.006598377600312233, + 0.004357227589935064, + 0.009897605516016483, + 0.004605923313647509, + 0.005250305403023958, + 0.003410556586459279, + 0.006131890695542097, + 0.0046155680902302265, + 0.00854572094976902, + 0.010246725752949715, + 0.0067278617061674595, + 0.00932947639375925, + 0.007261588238179684, + 0.015069282613694668, + 0.01007006410509348, + 0.010332112200558186, + 0.0057120849378407, + 0.00774340657517314, + 0.006125723943114281, + 0.008678817190229893, + 0.007583956699818373, + 0.004587145522236824, + 0.010480125434696674, + 0.008185039274394512, + 0.015586872585117817, + 0.009358924813568592, + 0.011759686283767223, + 0.012176639400422573, + 0.0148593969643116, + 0.010905160568654537, + 0.02001965418457985, + 0.01604645326733589, + 0.0052076405845582485, + 0.002572255441918969, + 0.004163482692092657, + 0.006290890276432037, + 0.006199141964316368, + 0.006655295845121145, + 0.006841810420155525, + 0.005884519312530756, + 0.004065064713358879, + 0.005682150833308697, + 0.005071851424872875, + 0.007650434970855713, + 0.00941462256014347, + 0.007252073846757412, + 0.007163181900978088, + 0.008443032391369343, + 0.02014056220650673, + 0.01735713705420494, + 0.012226372957229614, + 0.014137553051114082, + 0.014110698364675045, + 0.017025545239448547, + 0.027651509270071983, + 0.021220635622739792, + 0.014944586902856827, + 0.007169422693550587, + 0.007684663403779268, + 0.007521925028413534, + 0.012088043615221977, + 0.02336614392697811, + 0.016980526968836784, + 0.016238076612353325, + 0.029229873791337013, + 0.015187292359769344, + 0.028814900666475296, + 0.03682280704379082, + 0.012583841569721699, + 0.017059484496712685, + 0.03078838810324669, + 0.04116833955049515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.03889887407422066 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.055017709732055664 + ], + "opposite_logprob_tokens": [ + -2.9300177097320557 + ], + "logprob_reward": -0.055017709732055664, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.055017709732055664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12905962765216827, + 0.03164944797754288, + 0.033421315252780914, + 0.022861475124955177, + 0.07716775685548782, + 0.06736113131046295, + 0.12636667490005493, + 0.08383642137050629, + 0.07400302588939667, + 0.3542731702327728 + ], + "credit_map": [ + 0.005983111914247274, + 0.007441265042871237, + 0.00815548375248909, + 0.012056863866746426, + 0.013618893921375275, + 0.03063272312283516, + 0.018274877220392227, + 0.03289640694856644, + 0.004455932881683111, + 0.003947811666876078, + 0.0029652833472937346, + 0.0035387095995247364, + 0.004999012686312199, + 0.0024554647970944643, + 0.005688442382961512, + 0.0035987875889986753, + 0.004069819115102291, + 0.00426240311935544, + 0.0025647147558629513, + 0.004531493876129389, + 0.002948987064883113, + 0.005492976866662502, + 0.0034299588296562433, + 0.006120959296822548, + 0.0019714629743248224, + 0.003106636693701148, + 0.0023166786413639784, + 0.003970594145357609, + 0.002250051125884056, + 0.0018371801124885678, + 0.004405055195093155, + 0.003003813559189439, + 0.006610995624214411, + 0.007539345882833004, + 0.008450635708868504, + 0.010328453965485096, + 0.01232466846704483, + 0.0077313887886703014, + 0.013429472222924232, + 0.010752790607511997, + 0.0029355832375586033, + 0.0027347556315362453, + 0.004774298518896103, + 0.00864403322339058, + 0.009226315654814243, + 0.005899026989936829, + 0.022100767120718956, + 0.011046339757740498, + 0.009254946373403072, + 0.01337160263210535, + 0.0052762883715331554, + 0.013997793197631836, + 0.024624379351735115, + 0.017504144459962845, + 0.015300959348678589, + 0.027036553248763084, + 0.008389957249164581, + 0.006624598056077957, + 0.007396692410111427, + 0.00742363790050149, + 0.008025050163269043, + 0.01679951511323452, + 0.010668487288057804, + 0.01850847713649273, + 0.004372529219835997, + 0.0035125375725328922, + 0.00252229324541986, + 0.00472228741273284, + 0.007300690282136202, + 0.018814288079738617, + 0.02293800562620163, + 0.00982038676738739, + 0.04972049593925476, + 0.02620389126241207, + 0.049962565302848816, + 0.06142094358801842, + 0.020898332819342613, + 0.02860405668616295, + 0.0504499152302742, + 0.06701293587684631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.055017709732055664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048663634806871414 + ], + "opposite_logprob_tokens": [ + -3.04866361618042 + ], + "logprob_reward": -0.048663634806871414, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048663634806871414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1763485223054886, + 0.038397740572690964, + 0.045166973024606705, + 0.03079182282090187, + 0.10689662396907806, + 0.06893559545278549, + 0.03799784556031227, + 0.10870768874883652, + 0.15878762304782867, + 0.22796954214572906 + ], + "credit_map": [ + 0.010555663146078587, + 0.008171840570867062, + 0.009013287723064423, + 0.01393898669630289, + 0.014721536077558994, + 0.0442063994705677, + 0.02568718232214451, + 0.05005362629890442, + 0.004413057584315538, + 0.0036326644476503134, + 0.004982417449355125, + 0.004916464909911156, + 0.004837237298488617, + 0.0024915053509175777, + 0.007794786244630814, + 0.005329609382897615, + 0.003566439263522625, + 0.004797716625034809, + 0.0056365118362009525, + 0.006838309112936258, + 0.004448986612260342, + 0.006258600857108831, + 0.0064633688889443874, + 0.007157035171985626, + 0.004501183982938528, + 0.0033462794963270426, + 0.0029701138846576214, + 0.0051278406754136086, + 0.0034424399491399527, + 0.0028861933387815952, + 0.005810198839753866, + 0.0027075738180428743, + 0.013147639110684395, + 0.014425251632928848, + 0.012722521089017391, + 0.012446170672774315, + 0.013944612815976143, + 0.010751842521131039, + 0.012953434139490128, + 0.016505151987075806, + 0.004065144807100296, + 0.003432368626818061, + 0.005560565739870071, + 0.009553905576467514, + 0.008186643943190575, + 0.006359869614243507, + 0.020238541066646576, + 0.011538558639585972, + 0.0024087170604616404, + 0.0037046377547085285, + 0.004552318248897791, + 0.006293055601418018, + 0.006002197507768869, + 0.0030659381300210953, + 0.005567724816501141, + 0.006403257139027119, + 0.006816088687628508, + 0.006517915520817041, + 0.005966156255453825, + 0.011064467951655388, + 0.00966082327067852, + 0.019833462312817574, + 0.02469627559185028, + 0.024152498692274094, + 0.011235685087740421, + 0.008244500495493412, + 0.013174375519156456, + 0.013698849827051163, + 0.01684982143342495, + 0.052617285400629044, + 0.023803487420082092, + 0.019163629040122032, + 0.03149130940437317, + 0.017406398430466652, + 0.03352966159582138, + 0.037330497056245804, + 0.015120811760425568, + 0.01844361424446106, + 0.031201768666505814, + 0.04344549402594566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.048663634806871414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4290461242198944 + ], + "opposite_logprob_tokens": [ + -1.0540461540222168 + ], + "logprob_reward": -0.4290461242198944, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4290461242198944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15167583525180817, + 0.08111841231584549, + 0.12440115213394165, + 0.03584759682416916, + 0.02997220866382122, + 0.05240873619914055, + 0.026447778567671776, + 0.029380371794104576, + 0.020842773839831352, + 0.44790515303611755 + ], + "credit_map": [ + 0.007827195338904858, + 0.009676839224994183, + 0.0067537399008870125, + 0.008785076439380646, + 0.008324355818331242, + 0.048622455447912216, + 0.011964257806539536, + 0.049721911549568176, + 0.012348375283181667, + 0.007238649297505617, + 0.01014612801373005, + 0.007280401419848204, + 0.011618203483521938, + 0.005522458348423243, + 0.014817755669355392, + 0.012146431021392345, + 0.014044891111552715, + 0.014398290775716305, + 0.009977276436984539, + 0.01767200604081154, + 0.008861365728080273, + 0.019278796389698982, + 0.017548942938447, + 0.022619573399424553, + 0.00387517549097538, + 0.003839793149381876, + 0.003233203897252679, + 0.00641069607809186, + 0.0032845265232026577, + 0.0043177432380616665, + 0.006480828393250704, + 0.004405627958476543, + 0.004103647544980049, + 0.003895302303135395, + 0.003415771061554551, + 0.0025942593347281218, + 0.0034764637239277363, + 0.00246011302806437, + 0.0044985683634877205, + 0.0055280826054513454, + 0.0033598686568439007, + 0.0026232171803712845, + 0.004171327222138643, + 0.005757555831223726, + 0.007162721827626228, + 0.005068416241556406, + 0.015749631449580193, + 0.008515994995832443, + 0.0018766241846606135, + 0.0034898295998573303, + 0.0028297638054937124, + 0.003479346167296171, + 0.004024998284876347, + 0.0040915533900260925, + 0.003124806797131896, + 0.003530853660777211, + 0.0031284643337130547, + 0.002599189756438136, + 0.001947052893228829, + 0.0028804969042539597, + 0.0029158026445657015, + 0.004127283114939928, + 0.006247167009860277, + 0.005534912925213575, + 0.001947120064869523, + 0.001156667130999267, + 0.0016596713103353977, + 0.0023960843682289124, + 0.0024260450154542923, + 0.0061250352300703526, + 0.002162439748644829, + 0.0029697101563215256, + 0.06498467177152634, + 0.03178286552429199, + 0.06230693310499191, + 0.07152906805276871, + 0.026033082976937294, + 0.0372542142868042, + 0.06222240999341011, + 0.09179188311100006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.4290461242198944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20162148773670197 + ], + "opposite_logprob_tokens": [ + -1.7016215324401855 + ], + "logprob_reward": -0.20162148773670197, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20162148773670197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049014683812856674, + 0.03346531093120575, + 0.03577597066760063, + 0.05485824868083, + 0.09180770814418793, + 0.16384626924991608, + 0.29531219601631165, + 0.14068365097045898, + 0.10600815713405609, + 0.02922777086496353 + ], + "credit_map": [ + 0.00560813257470727, + 0.004477175418287516, + 0.00507782306522131, + 0.005820286925882101, + 0.006551283877342939, + 0.008685018867254257, + 0.005082275718450546, + 0.0077126845717430115, + 0.004715129267424345, + 0.0033506175968796015, + 0.007472870405763388, + 0.0034478420857340097, + 0.0038425286766141653, + 0.0025104943197220564, + 0.004461147356778383, + 0.003664680989459157, + 0.004358807113021612, + 0.004016505554318428, + 0.0032591382041573524, + 0.005219744052737951, + 0.002351097296923399, + 0.005142757669091225, + 0.0048933788202703, + 0.0065345438197255135, + 0.006112814415246248, + 0.006007745396345854, + 0.0060610901564359665, + 0.009445163421332836, + 0.005992498714476824, + 0.007124615833163261, + 0.00883842445909977, + 0.00527589675039053, + 0.00903640128672123, + 0.0077012814581394196, + 0.010786965489387512, + 0.00883349310606718, + 0.011652943678200245, + 0.007414469961076975, + 0.01948854885995388, + 0.016893604770302773, + 0.011996257118880749, + 0.009993095882236958, + 0.013571607880294323, + 0.019071800634264946, + 0.021126585081219673, + 0.014104317873716354, + 0.047503963112831116, + 0.02647865004837513, + 0.021343853324651718, + 0.030826741829514503, + 0.020033936947584152, + 0.034930892288684845, + 0.06247734650969505, + 0.03558298572897911, + 0.03648439794778824, + 0.053632043302059174, + 0.008532323874533176, + 0.008799736388027668, + 0.010418198071420193, + 0.018531883135437965, + 0.016132080927491188, + 0.01961684413254261, + 0.035769954323768616, + 0.022882619872689247, + 0.010102909989655018, + 0.0067178248427808285, + 0.007919455878436565, + 0.007609555032104254, + 0.009548831731081009, + 0.03797697275876999, + 0.013681837357580662, + 0.012450768612325191, + 0.004025908187031746, + 0.0033050156198441982, + 0.003497398691251874, + 0.003860064083710313, + 0.0020851274020969868, + 0.0026222316082566977, + 0.004020686727017164, + 0.00581133970990777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.20162148773670197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This handheld, slightly shaky close-up begins by focusing on the back of a small child's head, where fine, light brown hair is visible. The child's face is completely buried in the thick, dark gray canvas of their parent's jacket, their tiny body pressed tightly against their parent's chest for safety. The camera slowly and steadily pulls back to a medium shot, revealing the parent's arms tightly and protectively wrapped around the child, their tiny hands clenched into fists, desperately clinging to the jacket's fabric. They are engulfed by a surging crowd at a sunny outdoor festival. The background is vibrant, out-of-focus, with movement and color, filled with cheering people raising their hands, waving colorful flags, and throwing confetti that glitters in the bright afternoon sun. The density and flow of the crowd creates a palpable sense of pressure, contrasting sharply with the intimate, protected space created by the parents, establishing a poignant atmosphere of childhood fragility amidst the overwhelming public cheering.", + "seed": 15637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_14/index_5757_seed_15637_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": -0.048990164324641226, + "vlm_logprob": -0.048990164324641226, + "vlm_yes_no_margin": 3.299999976158142, + "vlm_token_credit": -0.048990164324641226, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02691531926393509 + ], + "opposite_logprob_tokens": [ + -3.7769153118133545 + ], + "logprob_reward": -0.02691531926393509, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02691531926393509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20395196974277496, + 0.15238387882709503, + 0.1182650476694107, + 0.06911738961935043, + 0.08197521418333054, + 0.05793684348464012, + 0.07246354967355728, + 0.12465545535087585, + 0.063709557056427, + 0.055541008710861206 + ], + "credit_map": [ + 0.02754213660955429, + 0.019092846661806107, + 0.016800416633486748, + 0.021815165877342224, + 0.03195349872112274, + 0.0359770841896534, + 0.02905995398759842, + 0.021710863336920738, + 0.010834166780114174, + 0.01338147185742855, + 0.018686771392822266, + 0.020627835765480995, + 0.01684408262372017, + 0.02104463055729866, + 0.023133112117648125, + 0.02783181518316269, + 0.010526069439947605, + 0.014249813742935658, + 0.012226371094584465, + 0.014854864217340946, + 0.014282851479947567, + 0.015609227120876312, + 0.020466001704335213, + 0.016049858182668686, + 0.006148721557110548, + 0.005035972688347101, + 0.012938751839101315, + 0.010891688987612724, + 0.011398600414395332, + 0.005385134369134903, + 0.009163282811641693, + 0.008155234158039093, + 0.007068055681884289, + 0.01261521689593792, + 0.00902283564209938, + 0.012376014143228531, + 0.014669488184154034, + 0.007703340612351894, + 0.011186813935637474, + 0.0073334514163434505, + 0.0044067553244531155, + 0.005768346134573221, + 0.007915748283267021, + 0.007329233456403017, + 0.006814343389123678, + 0.007622631266713142, + 0.010211714543402195, + 0.007868069224059582, + 0.0045942217111587524, + 0.012233086861670017, + 0.013994703069329262, + 0.015363339334726334, + 0.005212884396314621, + 0.00573186157271266, + 0.009128362871706486, + 0.0062050893902778625, + 0.01316357683390379, + 0.012620948255062103, + 0.02174857258796692, + 0.019433248788118362, + 0.012286314740777016, + 0.014355900697410107, + 0.01909390464425087, + 0.011952991597354412, + 0.008498912677168846, + 0.011877906508743763, + 0.0056053707376122475, + 0.008188390173017979, + 0.006168350111693144, + 0.005053012631833553, + 0.011875083670020103, + 0.006442532874643803, + 0.004888738039880991, + 0.008735460229218006, + 0.0071694809012115, + 0.008817721158266068, + 0.005336326081305742, + 0.008365864865481853, + 0.006727440282702446, + 0.005499978084117174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02691531926393509 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308783635497093 + ], + "opposite_logprob_tokens": [ + -3.1680877208709717 + ], + "logprob_reward": -0.04308783635497093, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308783635497093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11726363003253937, + 0.08106102794408798, + 0.10279781371355057, + 0.05140533298254013, + 0.03684130683541298, + 0.15158499777317047, + 0.20743586122989655, + 0.14706136286258698, + 0.0672314241528511, + 0.03731723874807358 + ], + "credit_map": [ + 0.00943558756262064, + 0.010383340530097485, + 0.009663055650889874, + 0.011584221385419369, + 0.015829799696803093, + 0.031138526275753975, + 0.014656253159046173, + 0.014572849497199059, + 0.0049756970256567, + 0.006794579327106476, + 0.006162961944937706, + 0.00883894506841898, + 0.009230314753949642, + 0.021721292287111282, + 0.016675781458616257, + 0.00666144909337163, + 0.009656527079641819, + 0.014378335326910019, + 0.009548313915729523, + 0.01524185948073864, + 0.012073311023414135, + 0.011994154192507267, + 0.018814604729413986, + 0.011090713553130627, + 0.005351379979401827, + 0.003997658845037222, + 0.007494387682527304, + 0.008880598470568657, + 0.004355207551270723, + 0.003908904734998941, + 0.009867536835372448, + 0.007549657486379147, + 0.00574647169560194, + 0.0034229187294840813, + 0.004149974789470434, + 0.006712137255817652, + 0.006251324899494648, + 0.0037632363382726908, + 0.004104141145944595, + 0.002691099653020501, + 0.011537089943885803, + 0.014384991489350796, + 0.025206290185451508, + 0.015050929971039295, + 0.03560623154044151, + 0.016196899116039276, + 0.015322444960474968, + 0.018280113115906715, + 0.009823529049754143, + 0.04043809697031975, + 0.03253054991364479, + 0.031063752248883247, + 0.027823522686958313, + 0.025688784196972847, + 0.01984342560172081, + 0.02022419683635235, + 0.01091773621737957, + 0.015618209727108479, + 0.020506009459495544, + 0.022398998960852623, + 0.016893373802304268, + 0.01301212515681982, + 0.03372076153755188, + 0.013994147069752216, + 0.009354778565466404, + 0.011161970905959606, + 0.006987927947193384, + 0.00944070890545845, + 0.005912095308303833, + 0.005787384696304798, + 0.011476997286081314, + 0.007109561003744602, + 0.0027134937699884176, + 0.004029619973152876, + 0.004682573955506086, + 0.005667753983289003, + 0.004765957593917847, + 0.004915035795420408, + 0.005020967684686184, + 0.005521835293620825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.04308783635497093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369786962866783 + ], + "opposite_logprob_tokens": [ + -3.4086978435516357 + ], + "logprob_reward": -0.03369786962866783, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369786962866783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1955145001411438, + 0.07575642317533493, + 0.07134860008955002, + 0.07071630656719208, + 0.09862492978572845, + 0.09920769184827805, + 0.05752949044108391, + 0.128187358379364, + 0.10395549982786179, + 0.09915924817323685 + ], + "credit_map": [ + 0.024724163115024567, + 0.02169267274439335, + 0.020927991718053818, + 0.027506699785590172, + 0.025104163214564323, + 0.03054339624941349, + 0.02388482727110386, + 0.021130593493580818, + 0.007316539064049721, + 0.007019533310085535, + 0.011623960919678211, + 0.009642078541219234, + 0.010602381080389023, + 0.010318445973098278, + 0.01068052276968956, + 0.008552955463528633, + 0.007354772184044123, + 0.010168022476136684, + 0.007310514338314533, + 0.008213095366954803, + 0.009372614324092865, + 0.01007345225661993, + 0.008070504292845726, + 0.010785626247525215, + 0.006430717185139656, + 0.007317952811717987, + 0.01052404660731554, + 0.011524123139679432, + 0.008134661242365837, + 0.005521948914974928, + 0.010398592799901962, + 0.010864264331758022, + 0.011456932872533798, + 0.01239017117768526, + 0.009139299392700195, + 0.0157681442797184, + 0.024132348597049713, + 0.008303516544401646, + 0.01151182223111391, + 0.0059226928278803825, + 0.007215011864900589, + 0.009037184529006481, + 0.017446616664528847, + 0.012199911288917065, + 0.018648959696292877, + 0.011836637742817402, + 0.012016735970973969, + 0.010806629434227943, + 0.003824815386906266, + 0.010664169676601887, + 0.010083035565912724, + 0.011900185607373714, + 0.004380474332720041, + 0.005725384224206209, + 0.006336541846394539, + 0.004614883568137884, + 0.010339722037315369, + 0.011923153884708881, + 0.018695440143346786, + 0.020339323207736015, + 0.009841225109994411, + 0.022244105115532875, + 0.023923324421048164, + 0.010881063528358936, + 0.012096026912331581, + 0.012686098925769329, + 0.008110346272587776, + 0.01684551313519478, + 0.009234447963535786, + 0.010649815201759338, + 0.021900469437241554, + 0.012432773597538471, + 0.007237880490720272, + 0.018097171559929848, + 0.009917664341628551, + 0.013536007143557072, + 0.012067046947777271, + 0.011291217990219593, + 0.014378342777490616, + 0.012633919715881348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.03369786962866783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1661497801542282, + 0.19508154690265656, + 0.1203935295343399, + 0.043119948357343674, + 0.09763671457767487, + 0.07572602480649948, + 0.13494957983493805, + 0.0947541669011116, + 0.048829883337020874, + 0.0233589056879282 + ], + "credit_map": [ + 0.014842594973742962, + 0.017010945826768875, + 0.021036464720964432, + 0.0228382870554924, + 0.020580235868692398, + 0.025811590254306793, + 0.024039428681135178, + 0.019990216940641403, + 0.008807620964944363, + 0.009034291841089725, + 0.010188593529164791, + 0.01644587330520153, + 0.013783495873212814, + 0.05915859714150429, + 0.061421871185302734, + 0.01624119095504284, + 0.01267259381711483, + 0.01411332655698061, + 0.011226477101445198, + 0.0153815234079957, + 0.014252773486077785, + 0.019645802676677704, + 0.014498508535325527, + 0.01860252022743225, + 0.0036463746801018715, + 0.0035086676944047213, + 0.006250590085983276, + 0.006261295638978481, + 0.005661287810653448, + 0.004384632688015699, + 0.0070854974910616875, + 0.0063216001726686954, + 0.008897250518202782, + 0.008367562666535378, + 0.006646862719208002, + 0.024162424728274345, + 0.032905351370573044, + 0.006727274041622877, + 0.004539490211755037, + 0.005390492733567953, + 0.006695002317428589, + 0.0069774058647453785, + 0.010953554883599281, + 0.009295320138335228, + 0.00977298617362976, + 0.012596980668604374, + 0.011352484114468098, + 0.008082282729446888, + 0.009218357503414154, + 0.03359123319387436, + 0.021946407854557037, + 0.020516902208328247, + 0.023164110258221626, + 0.009192111901938915, + 0.011644583195447922, + 0.0056758602149784565, + 0.01120325829833746, + 0.009833152405917645, + 0.01430784072726965, + 0.010767592117190361, + 0.009539177641272545, + 0.01629611663520336, + 0.015023170970380306, + 0.007783856242895126, + 0.006688001565635204, + 0.005084224510937929, + 0.005274411756545305, + 0.006543201860040426, + 0.0034156078472733498, + 0.005044698249548674, + 0.0064148674719035625, + 0.010364865884184837, + 0.0022514716256409883, + 0.0034784770105034113, + 0.0027037288527935743, + 0.004489116370677948, + 0.001836944604292512, + 0.003039018949493766, + 0.0033915555104613304, + 0.0021685895044356585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "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 shot a handheld, slightly shaky close-up at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the child's face buried in a thick, dark gray canvas jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera slowly and steadily pull back to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant, out-of-focus sunny outdoor festival with cheering people, waving flags, and glittering confetti?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the parent's arms tightly wrapped around the child with the child's hands clenched into fists?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270502656698227 + ], + "opposite_logprob_tokens": [ + -2.1270501613616943 + ], + "logprob_reward": -0.1270502656698227, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270502656698227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0718177929520607, + 0.10647188872098923, + 0.07072089612483978, + 0.0688064694404602, + 0.16051988303661346, + 0.052195191383361816, + 0.2960107624530792, + 0.07879020273685455, + 0.03616558387875557, + 0.05850137025117874 + ], + "credit_map": [ + 0.009419215843081474, + 0.0086907297372818, + 0.005770666524767876, + 0.008492905646562576, + 0.008570880629122257, + 0.011578414589166641, + 0.010740499943494797, + 0.008554480038583279, + 0.005131135229021311, + 0.00669469079002738, + 0.006318175233900547, + 0.009914305061101913, + 0.007547176908701658, + 0.030955517664551735, + 0.029567286372184753, + 0.01034359447658062, + 0.007874851115047932, + 0.009040962904691696, + 0.006228941027075052, + 0.010098613798618317, + 0.008314155042171478, + 0.010371786542236805, + 0.009352128952741623, + 0.009439458139240742, + 0.0052926745265722275, + 0.005385438911616802, + 0.013372458517551422, + 0.011179396882653236, + 0.01071768905967474, + 0.004513339605182409, + 0.0083864014595747, + 0.009959066286683083, + 0.01493666972965002, + 0.010457322932779789, + 0.010122598148882389, + 0.03956684470176697, + 0.05368828400969505, + 0.012196207419037819, + 0.011369036510586739, + 0.00818291213363409, + 0.0031822959426790476, + 0.005131955724209547, + 0.010277504101395607, + 0.006135918200016022, + 0.006088676396757364, + 0.006559506990015507, + 0.00813593901693821, + 0.006683390587568283, + 0.019526243209838867, + 0.06334381550550461, + 0.0539807491004467, + 0.042876288294792175, + 0.06393324583768845, + 0.018272217363119125, + 0.02272418700158596, + 0.011354003101587296, + 0.007819203659892082, + 0.009888295084238052, + 0.012427851557731628, + 0.010627285577356815, + 0.009035387076437473, + 0.010440920479595661, + 0.01204413827508688, + 0.006507117301225662, + 0.005301389377564192, + 0.005959353409707546, + 0.0025815730914473534, + 0.005592208821326494, + 0.0036704763770103455, + 0.0026783093344420195, + 0.0050966255366802216, + 0.005285643041133881, + 0.004063740372657776, + 0.008426161482930183, + 0.006131017580628395, + 0.00884836632758379, + 0.007775177247822285, + 0.006368664558976889, + 0.008542140945792198, + 0.008346102200448513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.1270502656698227 + } + ], + "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/0000058/group_15/index_tensor([1449, 1449])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_15/index_tensor([1449, 1449])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d5ffc6e44203c8eb074ea31c23cf4007e68a0002 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_15/index_tensor([1449, 1449])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "camera, starting from a low angle and slowly tilting upward, sweeps across the interior of an ancient temple hall. At the center of the frame lies a massive vermilion wooden column, its paint peeling slightly, revealing dark wood grain and the marks of age. A massive bronze dragon sculpture coils along this pillar, its muscular form imbued with power, its sharp claws digging deeply into the shaft. In the dim, flickering light of several eternal lamps burning overhead, each meticulously carved dragon scale reveals a mottled patina etched by age. As the camera ascends, the light streams across its winding surfaces, giving them a cold, mysterious metallic sheen before fading into deep shadow. The hall is empty and silent, the smooth black stone floor reflecting the vague outlines of the column and dragon. Wisps of smoke from incense burners in the distance drift slowly through the stagnant air, creating a solemn and mysterious atmosphere steeped in history.", + "seed": 23325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_15/index_1449_seed_23325_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": -0.15419153654947876, + "vlm_logprob": -0.15419153654947876, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.15419153654947876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.5367838144302368 + ], + "opposite_logprob_tokens": [ + -0.9117838144302368 + ], + "logprob_reward": -0.5367838144302368, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5367838144302368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18050502240657806, + 0.0950705036520958, + 0.05152896046638489, + 0.03558127582073212, + 0.05147913470864296, + 0.04370564594864845, + 0.14918994903564453, + 0.07347386330366135, + 0.2727438807487488, + 0.04672180116176605 + ], + "credit_map": [ + 0.012644831091165543, + 0.012587065808475018, + 0.04786372929811478, + 0.013979642651975155, + 0.014715920202434063, + 0.011103993281722069, + 0.014836172573268414, + 0.05277367681264877, + 0.00753316655755043, + 0.016646644100546837, + 0.012542512267827988, + 0.010218298994004726, + 0.014214707538485527, + 0.009250730276107788, + 0.01415222603827715, + 0.01051221415400505, + 0.004862379282712936, + 0.007700752466917038, + 0.0060637155547738075, + 0.007243277970701456, + 0.005372249521315098, + 0.007618407253175974, + 0.006539107766002417, + 0.006129070650786161, + 0.0028918155003339052, + 0.003999102860689163, + 0.004116337280720472, + 0.00805732887238264, + 0.003978485241532326, + 0.0022850539535284042, + 0.005808589048683643, + 0.004444561433047056, + 0.004152447916567326, + 0.007457822095602751, + 0.006497981026768684, + 0.0054581621661782265, + 0.008235985413193703, + 0.004989479668438435, + 0.005878381431102753, + 0.008808874525129795, + 0.0022360514849424362, + 0.003972696140408516, + 0.010071385651826859, + 0.004637306090444326, + 0.006063947919756174, + 0.0032021463848650455, + 0.005783969070762396, + 0.007738142739981413, + 0.010261076502501965, + 0.015614914707839489, + 0.01754923164844513, + 0.028551826253533363, + 0.025599978864192963, + 0.011227022856473923, + 0.02442816272377968, + 0.015957744792103767, + 0.005481756757944822, + 0.009207382798194885, + 0.007117609027773142, + 0.012367094866931438, + 0.00929905753582716, + 0.007019113749265671, + 0.012736291624605656, + 0.010245561599731445, + 0.01698293723165989, + 0.040002044290304184, + 0.0569174662232399, + 0.039545733481645584, + 0.053719114512205124, + 0.02128450945019722, + 0.02229662984609604, + 0.02199545130133629, + 0.003263528225943446, + 0.006687390152364969, + 0.005567663349211216, + 0.007403508760035038, + 0.00561344251036644, + 0.006379233673214912, + 0.00542079471051693, + 0.006386242341250181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.5367838144302368 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.0052775111980736256 + ], + "opposite_logprob_tokens": [ + -5.255277633666992 + ], + "logprob_reward": -0.0052775111980736256, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052775111980736256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09863001853227615, + 0.3379136621952057, + 0.04311821609735489, + 0.06456855684518814, + 0.056701019406318665, + 0.10953773558139801, + 0.07013757526874542, + 0.05854003503918648, + 0.10819166898727417, + 0.052661553025245667 + ], + "credit_map": [ + 0.007964893244206905, + 0.008976878598332405, + 0.02072764001786709, + 0.008131581358611584, + 0.008957316167652607, + 0.007509078364819288, + 0.008615691214799881, + 0.027746934443712234, + 0.02847159467637539, + 0.060003116726875305, + 0.03806626424193382, + 0.03606974333524704, + 0.05808529630303383, + 0.03056235983967781, + 0.047720201313495636, + 0.038935039192438126, + 0.003989548422396183, + 0.008469333872199059, + 0.005081031937152147, + 0.0048745740205049515, + 0.005686142016202211, + 0.004109592642635107, + 0.005586700979620218, + 0.005321290343999863, + 0.004809945356100798, + 0.007792152464389801, + 0.00699042109772563, + 0.010686121881008148, + 0.009912037290632725, + 0.006000743247568607, + 0.0093785859644413, + 0.008998543955385685, + 0.004421229008585215, + 0.008608988486230373, + 0.005702242720872164, + 0.007314939517527819, + 0.008631053380668163, + 0.004534658510237932, + 0.008611341938376427, + 0.008876560255885124, + 0.005371103063225746, + 0.014399153180420399, + 0.027957677841186523, + 0.010553148575127125, + 0.010211562737822533, + 0.005162118002772331, + 0.014090270735323429, + 0.021792693063616753, + 0.004491725470870733, + 0.009893840178847313, + 0.010756795294582844, + 0.01028191763907671, + 0.011572669260203838, + 0.0045280419290065765, + 0.011063169687986374, + 0.007549405097961426, + 0.0042007253505289555, + 0.014041857793927193, + 0.00816484447568655, + 0.008093743585050106, + 0.007243738509714603, + 0.003896031528711319, + 0.006055728066712618, + 0.006843360606580973, + 0.004360607825219631, + 0.006809930317103863, + 0.033852506428956985, + 0.010585766285657883, + 0.03135659918189049, + 0.008056024089455605, + 0.005261782091110945, + 0.007908439263701439, + 0.0038652396760880947, + 0.005472110118716955, + 0.00544722331687808, + 0.012262187898159027, + 0.005177758168429136, + 0.00791286863386631, + 0.007153625600039959, + 0.0053705377504229546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.0052775111980736256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.016079746186733246 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079746186733246, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079746186733246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13582924008369446, + 0.15680226683616638, + 0.04573485255241394, + 0.08747954666614532, + 0.12665234506130219, + 0.12984882295131683, + 0.05297716334462166, + 0.11109065264463425, + 0.08022589981555939, + 0.07335915416479111 + ], + "credit_map": [ + 0.01323129516094923, + 0.020142316818237305, + 0.022700661793351173, + 0.015193099156022072, + 0.01759050227701664, + 0.009984786622226238, + 0.015851611271500587, + 0.021134983748197556, + 0.013080895878374577, + 0.023335246369242668, + 0.022944074124097824, + 0.01826315186917782, + 0.030207278206944466, + 0.016034746542572975, + 0.01821867562830448, + 0.014718214981257915, + 0.0037040484603494406, + 0.00536656379699707, + 0.005703054368495941, + 0.006165423896163702, + 0.007718096021562815, + 0.0047954656183719635, + 0.006340598221868277, + 0.0059416089206933975, + 0.006705606821924448, + 0.009621471166610718, + 0.009797279722988605, + 0.013485444709658623, + 0.013498171232640743, + 0.007936983369290829, + 0.014113141223788261, + 0.012321460992097855, + 0.007685018703341484, + 0.01749240607023239, + 0.015752475708723068, + 0.016644582152366638, + 0.019268972799181938, + 0.013145294971764088, + 0.015427564270794392, + 0.021236039698123932, + 0.005854478105902672, + 0.01568140834569931, + 0.029461756348609924, + 0.013808540999889374, + 0.015513770282268524, + 0.008111824281513691, + 0.014032482169568539, + 0.02738458290696144, + 0.003968833480030298, + 0.007674717344343662, + 0.008067304268479347, + 0.0062593198381364346, + 0.008780079893767834, + 0.004013171419501305, + 0.008755046874284744, + 0.005458694417029619, + 0.007909120060503483, + 0.02491881512105465, + 0.011339250020682812, + 0.01445543859153986, + 0.014672815799713135, + 0.007945461198687553, + 0.012893025763332844, + 0.01695672795176506, + 0.004166976548731327, + 0.01260622963309288, + 0.01154530979692936, + 0.013628442771732807, + 0.01849026046693325, + 0.0068361638113856316, + 0.0056875888258218765, + 0.007264933083206415, + 0.0053012496791779995, + 0.00902343075722456, + 0.008744800463318825, + 0.0178558100014925, + 0.006205388810485601, + 0.010492946952581406, + 0.009011783637106419, + 0.006723752245306969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.016079746186733246 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.07002648711204529 + ], + "opposite_logprob_tokens": [ + -2.695026397705078 + ], + "logprob_reward": -0.07002648711204529, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002648711204529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16119323670864105, + 0.5220696926116943, + 0.05694103240966797, + 0.042049311101436615, + 0.030338475480675697, + 0.0426667146384716, + 0.04900164157152176, + 0.02387317642569542, + 0.04819377511739731, + 0.023672906681895256 + ], + "credit_map": [ + 0.016492532566189766, + 0.01458316296339035, + 0.03878419101238251, + 0.010217804461717606, + 0.010516966693103313, + 0.010057386010885239, + 0.010501720011234283, + 0.05003947392106056, + 0.05676892027258873, + 0.06071944534778595, + 0.1000087559223175, + 0.06758368015289307, + 0.04896393045783043, + 0.054014068096876144, + 0.06287949532270432, + 0.07113141566514969, + 0.004616413731127977, + 0.010766197927296162, + 0.006496382411569357, + 0.007912207394838333, + 0.005241504404693842, + 0.006010129582136869, + 0.008417832665145397, + 0.007480362430214882, + 0.003914173226803541, + 0.004489435348659754, + 0.006410875823348761, + 0.005668393801897764, + 0.006179507356137037, + 0.0036888387985527515, + 0.0054468116723001, + 0.006251273211091757, + 0.0023837771732360125, + 0.004827599041163921, + 0.0039603994227945805, + 0.004624587018042803, + 0.0031870550010353327, + 0.0018762998515740037, + 0.004230549558997154, + 0.005248208064585924, + 0.0024922338780015707, + 0.005090815480798483, + 0.01006327848881483, + 0.0037097972817718983, + 0.005294034723192453, + 0.0026623141020536423, + 0.004151148721575737, + 0.009203091263771057, + 0.003789895912632346, + 0.004732249770313501, + 0.005722379311919212, + 0.00946247298270464, + 0.0073170713149011135, + 0.0036972814705222845, + 0.008956583216786385, + 0.005323701538145542, + 0.0013827496441081166, + 0.004410734865814447, + 0.002559531247243285, + 0.0046388027258217335, + 0.002676504198461771, + 0.001997321145609021, + 0.0034085758961737156, + 0.002798956586048007, + 0.0025449928361922503, + 0.00754465302452445, + 0.00896845106035471, + 0.005088998470455408, + 0.011476841755211353, + 0.003910470753908157, + 0.004001251421868801, + 0.004658115096390247, + 0.0022281098645180464, + 0.0025212643668055534, + 0.0030717875342816114, + 0.005279670935124159, + 0.0024386628065258265, + 0.003300825133919716, + 0.0024211029522120953, + 0.00241148192435503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07002648711204529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.14279012382030487 + ], + "opposite_logprob_tokens": [ + -2.0177900791168213 + ], + "logprob_reward": -0.14279012382030487, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279012382030487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14890404045581818, + 0.05319065973162651, + 0.06077025085687637, + 0.13312120735645294, + 0.06018579378724098, + 0.1252441108226776, + 0.10047999769449234, + 0.10898374766111374, + 0.16262371838092804, + 0.04649649187922478 + ], + "credit_map": [ + 0.020897522568702698, + 0.01993575319647789, + 0.021263951435685158, + 0.01703741028904915, + 0.012366925366222858, + 0.011142686940729618, + 0.013130545616149902, + 0.033129241317510605, + 0.005226365756243467, + 0.008023529313504696, + 0.007352802436798811, + 0.0054513621143996716, + 0.00951171200722456, + 0.004169335123151541, + 0.007003141101449728, + 0.006452410481870174, + 0.006338226143270731, + 0.0124141164124012, + 0.008765402249991894, + 0.00710633210837841, + 0.007580928038805723, + 0.005367270205169916, + 0.007210659794509411, + 0.005987313110381365, + 0.012243811041116714, + 0.015325598418712616, + 0.016849076375365257, + 0.022078504785895348, + 0.019978374242782593, + 0.010175489820539951, + 0.01790275052189827, + 0.018567588180303574, + 0.0044034249149262905, + 0.010847502388060093, + 0.00821992289274931, + 0.0054283663630485535, + 0.007389689330011606, + 0.007895223796367645, + 0.006177298258990049, + 0.009824362583458424, + 0.006674838252365589, + 0.008872175589203835, + 0.04032280296087265, + 0.00831112265586853, + 0.011912376619875431, + 0.005904424004256725, + 0.009741582907736301, + 0.033504799008369446, + 0.004807438235729933, + 0.011545756831765175, + 0.013314556330442429, + 0.016126055270433426, + 0.017806967720389366, + 0.008998995646834373, + 0.016911448910832405, + 0.010968772694468498, + 0.009295632131397724, + 0.02619107812643051, + 0.011909186840057373, + 0.014222300611436367, + 0.013691923581063747, + 0.007755951955914497, + 0.013851789757609367, + 0.012065878137946129, + 0.007234993390738964, + 0.021728679537773132, + 0.03963954374194145, + 0.022264808416366577, + 0.02905353158712387, + 0.011589600704610348, + 0.017399854958057404, + 0.013712700456380844, + 0.003824712010100484, + 0.007251794915646315, + 0.004942475818097591, + 0.007854670286178589, + 0.006144048646092415, + 0.006271297112107277, + 0.005437665618956089, + 0.004769830498844385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.14279012382030487 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera, starting from a low angle and slowly tilting upward, sweeps across the interior of an ancient temple hall. At the center of the frame lies a massive vermilion wooden column, its paint peeling slightly, revealing dark wood grain and the marks of age. A massive bronze dragon sculpture coils along this pillar, its muscular form imbued with power, its sharp claws digging deeply into the shaft. In the dim, flickering light of several eternal lamps burning overhead, each meticulously carved dragon scale reveals a mottled patina etched by age. As the camera ascends, the light streams across its winding surfaces, giving them a cold, mysterious metallic sheen before fading into deep shadow. The hall is empty and silent, the smooth black stone floor reflecting the vague outlines of the column and dragon. Wisps of smoke from incense burners in the distance drift slowly through the stagnant air, creating a solemn and mysterious atmosphere steeped in history.", + "seed": 23326, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_15/index_1449_seed_23326_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": -0.15119443982839584, + "vlm_logprob": -0.15119443982839584, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.15119443982839584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.48369288444519043 + ], + "opposite_logprob_tokens": [ + -0.9836928844451904 + ], + "logprob_reward": -0.48369288444519043, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.48369288444519043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19106945395469666, + 0.07629842311143875, + 0.09042973071336746, + 0.08293261379003525, + 0.19699333608150482, + 0.24543830752372742, + 0.03957297280430794, + 0.02542167343199253, + 0.02537769079208374, + 0.026465872302651405 + ], + "credit_map": [ + 0.017556946724653244, + 0.01879326067864895, + 0.035992421209812164, + 0.06419914215803146, + 0.009026139974594116, + 0.007541107013821602, + 0.01286031398922205, + 0.025100108236074448, + 0.007742476649582386, + 0.015462535433471203, + 0.007766416296362877, + 0.012335969135165215, + 0.012639612890779972, + 0.006106336135417223, + 0.007106770761311054, + 0.007138310000300407, + 0.006894019898027182, + 0.009019439108669758, + 0.01117817871272564, + 0.03288758918642998, + 0.006685599684715271, + 0.0052490984089672565, + 0.006883006542921066, + 0.01163280289620161, + 0.007304883096367121, + 0.011719748377799988, + 0.016439907252788544, + 0.02302004210650921, + 0.0062563237734138966, + 0.005594841204583645, + 0.007698423694819212, + 0.004898447077721357, + 0.020165225490927696, + 0.01607217825949192, + 0.014681375585496426, + 0.06443262100219727, + 0.02014096826314926, + 0.013333894312381744, + 0.025627577677369118, + 0.022539496421813965, + 0.028053700923919678, + 0.024029014632105827, + 0.014162590727210045, + 0.024422649294137955, + 0.08684451133012772, + 0.01692815311253071, + 0.03594616428017616, + 0.015051520429551601, + 0.006441955454647541, + 0.00474940799176693, + 0.004372202325612307, + 0.00572604313492775, + 0.00738149881362915, + 0.003931290935724974, + 0.0037458965089172125, + 0.0032246792688965797, + 0.003871465101838112, + 0.0027059228159487247, + 0.0030133668333292007, + 0.002380174584686756, + 0.0035440793726593256, + 0.002284277928993106, + 0.005198400001972914, + 0.0024239863269031048, + 0.004357864148914814, + 0.0030755787156522274, + 0.0024500233121216297, + 0.002528482349589467, + 0.0027539751026779413, + 0.003155401675030589, + 0.00361751695163548, + 0.0034388492349535227, + 0.003835478564724326, + 0.002667374210432172, + 0.004304917529225349, + 0.00560997799038887, + 0.002195541514083743, + 0.002088605659082532, + 0.002657453529536724, + 0.003106523072347045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.48369288444519043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.026378124952316284 + ], + "opposite_logprob_tokens": [ + -3.6513781547546387 + ], + "logprob_reward": -0.026378124952316284, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026378124952316284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2820265591144562, + 0.10573158413171768, + 0.08539055287837982, + 0.18644018471240997, + 0.040339820086956024, + 0.08529951423406601, + 0.046198464930057526, + 0.07198410481214523, + 0.04617850482463837, + 0.05041075497865677 + ], + "credit_map": [ + 0.027832956984639168, + 0.031572964042425156, + 0.026978949084877968, + 0.03437022119760513, + 0.04282507672905922, + 0.04155464842915535, + 0.03563201427459717, + 0.04125971719622612, + 0.01066905353218317, + 0.022138424217700958, + 0.009726905263960361, + 0.016080057248473167, + 0.01904456503689289, + 0.008011086843907833, + 0.009181085042655468, + 0.010880406945943832, + 0.007352570537477732, + 0.012854695320129395, + 0.00872420147061348, + 0.00916803628206253, + 0.019815944135189056, + 0.006426374427974224, + 0.01424799207597971, + 0.006800739094614983, + 0.018012678250670433, + 0.020741401240229607, + 0.02372053824365139, + 0.032931677997112274, + 0.020663777366280556, + 0.017459288239479065, + 0.03086436353623867, + 0.022046474739909172, + 0.00393054261803627, + 0.003468804294243455, + 0.0033985034096986055, + 0.006975272670388222, + 0.005052555352449417, + 0.004992151167243719, + 0.007313711568713188, + 0.00520827854052186, + 0.012436244636774063, + 0.00720649678260088, + 0.008826876059174538, + 0.02224811725318432, + 0.006618193816393614, + 0.007353190798312426, + 0.013117119669914246, + 0.007493279408663511, + 0.006792998407036066, + 0.005210744682699442, + 0.006023127119988203, + 0.00905811507254839, + 0.005414429120719433, + 0.005028399173170328, + 0.005172270815819502, + 0.003498381469398737, + 0.011148232966661453, + 0.006308241747319698, + 0.013328757137060165, + 0.009449083358049393, + 0.007623181212693453, + 0.0065543861128389835, + 0.01266359444707632, + 0.004908633418381214, + 0.007971694692969322, + 0.006176413036882877, + 0.005088167265057564, + 0.004536687396466732, + 0.005275801755487919, + 0.005255907773971558, + 0.006003373768180609, + 0.005870460532605648, + 0.007577256765216589, + 0.0069345757365226746, + 0.0033738596830517054, + 0.005864809267222881, + 0.007989236153662205, + 0.004910371731966734, + 0.008123539388179779, + 0.005637103226035833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.026378124952316284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.02981775999069214 + ], + "opposite_logprob_tokens": [ + -3.529817819595337 + ], + "logprob_reward": -0.02981775999069214, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981775999069214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2884089946746826, + 0.07317847013473511, + 0.06182747334241867, + 0.037024710327386856, + 0.09030390530824661, + 0.07838388532400131, + 0.10079474747180939, + 0.06502360850572586, + 0.03707768768072128, + 0.16797657310962677 + ], + "credit_map": [ + 0.019703878089785576, + 0.02715769037604332, + 0.07337778061628342, + 0.09472998231649399, + 0.02114741876721382, + 0.010168418288230896, + 0.024596361443400383, + 0.017527425661683083, + 0.007063949014991522, + 0.017520250752568245, + 0.006107032299041748, + 0.010767887346446514, + 0.01470162533223629, + 0.004578407388180494, + 0.005960515234619379, + 0.006478810217231512, + 0.0046413354575634, + 0.004292208235710859, + 0.005641178693622351, + 0.018936192616820335, + 0.007581130135804415, + 0.0037548497784882784, + 0.007065223064273596, + 0.009915359318256378, + 0.0038757440634071827, + 0.004608746152371168, + 0.005603463388979435, + 0.005687767174094915, + 0.005070456303656101, + 0.003596960799768567, + 0.004592658951878548, + 0.003988911397755146, + 0.0075166597962379456, + 0.005580062512308359, + 0.006250184495002031, + 0.01817513071000576, + 0.010232686065137386, + 0.010571600869297981, + 0.016123013570904732, + 0.01585456356406212, + 0.011381560936570168, + 0.00590625312179327, + 0.007742744404822588, + 0.02148626185953617, + 0.00657129380851984, + 0.006525394041091204, + 0.012886937707662582, + 0.005883441306650639, + 0.014921319670975208, + 0.011997113935649395, + 0.01202737633138895, + 0.015835626050829887, + 0.015219684690237045, + 0.012550835497677326, + 0.009753706865012646, + 0.008489090949296951, + 0.010759440250694752, + 0.00785126443952322, + 0.006216303911060095, + 0.005906839855015278, + 0.007932750508189201, + 0.005625338293612003, + 0.01284885685890913, + 0.00788281299173832, + 0.005398019682615995, + 0.003096406115218997, + 0.004116937518119812, + 0.0041456217877566814, + 0.003352569416165352, + 0.004976489581167698, + 0.005930050741881132, + 0.00606158934533596, + 0.024518001824617386, + 0.028159894049167633, + 0.022213898599147797, + 0.017945073544979095, + 0.02815457060933113, + 0.013568504713475704, + 0.019211553037166595, + 0.014205086044967175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.02981775999069214 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.12705792486667633 + ], + "opposite_logprob_tokens": [ + -2.1270580291748047 + ], + "logprob_reward": -0.12705792486667633, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705792486667633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11530672013759613, + 0.27027809619903564, + 0.13270409405231476, + 0.0492936335504055, + 0.08488866686820984, + 0.09271484613418579, + 0.07360457628965378, + 0.04838521406054497, + 0.05563352257013321, + 0.07719064503908157 + ], + "credit_map": [ + 0.008276238106191158, + 0.016500048339366913, + 0.020027797669172287, + 0.02125410921871662, + 0.01798873394727707, + 0.007885470986366272, + 0.014230341650545597, + 0.00914397370070219, + 0.02738436684012413, + 0.04313031956553459, + 0.02986145205795765, + 0.060551322996616364, + 0.037390053272247314, + 0.019768496975302696, + 0.02429736778140068, + 0.027894698083400726, + 0.014190715737640858, + 0.014316920191049576, + 0.018732775002717972, + 0.025841964408755302, + 0.021100815385580063, + 0.009277647361159325, + 0.016274943947792053, + 0.012968303635716438, + 0.004198022652417421, + 0.0045964098535478115, + 0.007544598076492548, + 0.012198921293020248, + 0.004981915466487408, + 0.005011022090911865, + 0.0055207922123372555, + 0.005241948179900646, + 0.006945975124835968, + 0.005754861049354076, + 0.006050556898117065, + 0.021007785573601723, + 0.007444770075380802, + 0.00856019090861082, + 0.012573741376399994, + 0.016550781205296516, + 0.011151125654578209, + 0.00923251174390316, + 0.010118355043232441, + 0.015067058615386486, + 0.020240651443600655, + 0.0077687595039606094, + 0.012366004288196564, + 0.006770370062440634, + 0.011732243932783604, + 0.009094280190765858, + 0.008522571064531803, + 0.011221159249544144, + 0.011872073635458946, + 0.008431145921349525, + 0.006505782715976238, + 0.006225308403372765, + 0.007322748191654682, + 0.005551089998334646, + 0.009737616404891014, + 0.007152573671191931, + 0.0041769505478441715, + 0.004036308266222477, + 0.006282005459070206, + 0.004125915002077818, + 0.010408314876258373, + 0.0053993109613657, + 0.012140454724431038, + 0.0080174021422863, + 0.0036276071332395077, + 0.00649030227214098, + 0.0035547411534935236, + 0.005995385814458132, + 0.012194700539112091, + 0.009011117741465569, + 0.008485854603350163, + 0.013474535197019577, + 0.008724959567189217, + 0.005732871126383543, + 0.008185027167201042, + 0.011381573975086212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.12705792486667633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.08902550488710403 + ], + "opposite_logprob_tokens": [ + -2.4640254974365234 + ], + "logprob_reward": -0.08902550488710403, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902550488710403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13174155354499817, + 0.05205898731946945, + 0.13174688816070557, + 0.14007700979709625, + 0.1438303291797638, + 0.026556339114904404, + 0.15002915263175964, + 0.12769818305969238, + 0.03402921184897423, + 0.062232352793216705 + ], + "credit_map": [ + 0.008832314983010292, + 0.010545213706791401, + 0.038358427584171295, + 0.03492558375000954, + 0.008395671844482422, + 0.004865714814513922, + 0.008731543086469173, + 0.017087087035179138, + 0.00600303104147315, + 0.006222137715667486, + 0.006726936437189579, + 0.014068033546209335, + 0.0058350032195448875, + 0.004534292500466108, + 0.004257733467966318, + 0.0044118198566138744, + 0.010529463179409504, + 0.015171326696872711, + 0.017122307792305946, + 0.03932131826877594, + 0.01205986924469471, + 0.00797349400818348, + 0.013459781184792519, + 0.016109326854348183, + 0.014197527430951595, + 0.024589935317635536, + 0.01874936744570732, + 0.020264092832803726, + 0.014049345627427101, + 0.021725203841924667, + 0.011691207997500896, + 0.014810333028435707, + 0.014285693876445293, + 0.01391179021447897, + 0.008375481702387333, + 0.05036141350865364, + 0.00790315866470337, + 0.0073017543181777, + 0.00944217387586832, + 0.03224886208772659, + 0.003675997955724597, + 0.005387532524764538, + 0.002212817082181573, + 0.004162336699664593, + 0.0023803720250725746, + 0.001626181765459478, + 0.004071071278303862, + 0.0030400303658097982, + 0.02323383092880249, + 0.017704540863633156, + 0.01798921264708042, + 0.022697966545820236, + 0.02337338775396347, + 0.017294051125645638, + 0.013584816828370094, + 0.014151337556540966, + 0.022163150832057, + 0.011707133613526821, + 0.029816532507538795, + 0.018503326922655106, + 0.010353933088481426, + 0.01022863108664751, + 0.013152685016393661, + 0.011772789992392063, + 0.005895120557397604, + 0.004547004122287035, + 0.005057542584836483, + 0.005144170485436916, + 0.0039048236794769764, + 0.003081663278862834, + 0.003583855228498578, + 0.0028150335419923067, + 0.009483086876571178, + 0.009574559517204762, + 0.00990548450499773, + 0.009270056150853634, + 0.007930132560431957, + 0.004457245115190744, + 0.006204270292073488, + 0.005407515913248062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.08902550488710403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera, starting from a low angle and slowly tilting upward, sweeps across the interior of an ancient temple hall. At the center of the frame lies a massive vermilion wooden column, its paint peeling slightly, revealing dark wood grain and the marks of age. A massive bronze dragon sculpture coils along this pillar, its muscular form imbued with power, its sharp claws digging deeply into the shaft. In the dim, flickering light of several eternal lamps burning overhead, each meticulously carved dragon scale reveals a mottled patina etched by age. As the camera ascends, the light streams across its winding surfaces, giving them a cold, mysterious metallic sheen before fading into deep shadow. The hall is empty and silent, the smooth black stone floor reflecting the vague outlines of the column and dragon. Wisps of smoke from incense burners in the distance drift slowly through the stagnant air, creating a solemn and mysterious atmosphere steeped in history.", + "seed": 23327, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_15/index_1449_seed_23327_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": -0.18961860239505768, + "vlm_logprob": -0.18961860239505768, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.18961860239505768, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.5406255125999451 + ], + "opposite_logprob_tokens": [ + -0.9156255125999451 + ], + "logprob_reward": -0.5406255125999451, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5406255125999451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07843615859746933, + 0.049988437443971634, + 0.06163549795746803, + 0.30266323685646057, + 0.023945309221744537, + 0.15192148089408875, + 0.03672409802675247, + 0.04971601814031601, + 0.14518266916275024, + 0.09978712350130081 + ], + "credit_map": [ + 0.010615772567689419, + 0.007460400927811861, + 0.015497236512601376, + 0.007921608164906502, + 0.007337820716202259, + 0.005343895871192217, + 0.01348781492561102, + 0.010771607980132103, + 0.006238792557269335, + 0.005321335978806019, + 0.0076797460205852985, + 0.004998823627829552, + 0.0038290186785161495, + 0.0036832233890891075, + 0.007120550610125065, + 0.011116948910057545, + 0.007810086943209171, + 0.006696248427033424, + 0.011523999273777008, + 0.008528882637619972, + 0.006403581704944372, + 0.0035662504378706217, + 0.008385262452065945, + 0.00872118677943945, + 0.036434900015592575, + 0.04729139804840088, + 0.05484645441174507, + 0.04779651761054993, + 0.034984856843948364, + 0.02394169755280018, + 0.035878606140613556, + 0.02148880809545517, + 0.003030026564374566, + 0.0022516492754220963, + 0.003276767209172249, + 0.0033729856368154287, + 0.0025236562360078096, + 0.0027624506037682295, + 0.002546638483181596, + 0.004181136377155781, + 0.01627337746322155, + 0.04210557043552399, + 0.059004224836826324, + 0.008894254453480244, + 0.008258841000497341, + 0.004167564678937197, + 0.005372869316488504, + 0.007844772189855576, + 0.007365134544670582, + 0.00636441633105278, + 0.00408734567463398, + 0.004239891655743122, + 0.0049995253793895245, + 0.0026275403797626495, + 0.003935635555535555, + 0.0031046061776578426, + 0.008649023249745369, + 0.00835051853209734, + 0.006498656701296568, + 0.00636428315192461, + 0.005936014465987682, + 0.004686718340963125, + 0.004771827254444361, + 0.004458972252905369, + 0.020340269431471825, + 0.016043055802583694, + 0.02021123841404915, + 0.02256273478269577, + 0.014049203135073185, + 0.01596849597990513, + 0.020468072965741158, + 0.015539583750069141, + 0.011881771497428417, + 0.010846348479390144, + 0.013557761907577515, + 0.015745345503091812, + 0.010445385240018368, + 0.011019717901945114, + 0.01677922159433365, + 0.009511570446193218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.5406255125999451 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.0298469178378582 + ], + "opposite_logprob_tokens": [ + -3.5298469066619873 + ], + "logprob_reward": -0.0298469178378582, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0298469178378582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21259871125221252, + 0.05498400703072548, + 0.03621498495340347, + 0.211663156747818, + 0.038076769560575485, + 0.17809917032718658, + 0.026000237092375755, + 0.0696028545498848, + 0.11442907899618149, + 0.05833106487989426 + ], + "credit_map": [ + 0.026158591732382774, + 0.021445870399475098, + 0.03068995475769043, + 0.02871757559478283, + 0.017156213521957397, + 0.0210300013422966, + 0.04472343996167183, + 0.022677039727568626, + 0.0088180061429739, + 0.006951690651476383, + 0.007752694189548492, + 0.0064215343445539474, + 0.005311118438839912, + 0.006709188222885132, + 0.005375418812036514, + 0.007644349243491888, + 0.004381955601274967, + 0.0034290128387510777, + 0.007578134071081877, + 0.005275774747133255, + 0.00224313884973526, + 0.0033181034959852695, + 0.005353329703211784, + 0.004635530523955822, + 0.026246123015880585, + 0.03323664516210556, + 0.03867379575967789, + 0.03340422362089157, + 0.023738190531730652, + 0.016473526135087013, + 0.023789331316947937, + 0.016101321205496788, + 0.005065637174993753, + 0.005125811789184809, + 0.007367359474301338, + 0.003992754966020584, + 0.004682888742536306, + 0.003496910911053419, + 0.0038333931006491184, + 0.004512012470513582, + 0.018299274146556854, + 0.0530715174973011, + 0.06877154111862183, + 0.009486191906034946, + 0.011295129545032978, + 0.005626989062875509, + 0.005955638829618692, + 0.005592877045273781, + 0.004621157422661781, + 0.004084484186023474, + 0.0035147725138813257, + 0.0034970759879797697, + 0.0035852533765137196, + 0.002212115563452244, + 0.002554248319938779, + 0.0019311285577714443, + 0.01039655040949583, + 0.01094289030879736, + 0.012820560485124588, + 0.01055422518402338, + 0.007640808820724487, + 0.00535614462569356, + 0.006481407675892115, + 0.005410266108810902, + 0.016260987147688866, + 0.012579387053847313, + 0.01394741702824831, + 0.017770731821656227, + 0.010543535463511944, + 0.013380472548305988, + 0.015742981806397438, + 0.014203554950654507, + 0.0060803829692304134, + 0.006143874488770962, + 0.007642937824130058, + 0.009275180287659168, + 0.006971242371946573, + 0.0065735806711018085, + 0.010190412402153015, + 0.0054534547962248325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0298469178378582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.012550774961709976 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550774961709976, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550774961709976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03845056891441345, + 0.0979616641998291, + 0.034069791436195374, + 0.4383032023906708, + 0.0609893798828125, + 0.07065191119909286, + 0.06234801560640335, + 0.09389001131057739, + 0.06610624492168427, + 0.0372292585670948 + ], + "credit_map": [ + 0.006150055676698685, + 0.005336643196642399, + 0.005425183102488518, + 0.0037548039108514786, + 0.00548027316108346, + 0.0037192271556705236, + 0.003660071874037385, + 0.004924312233924866, + 0.020951423794031143, + 0.005637710448354483, + 0.0079133166000247, + 0.0049419463612139225, + 0.005184986162930727, + 0.007878707721829414, + 0.013381007127463818, + 0.03207256644964218, + 0.0044137560762465, + 0.003570607164874673, + 0.007670198101550341, + 0.005280185956507921, + 0.0021145506761968136, + 0.0031142940279096365, + 0.0036622604820877314, + 0.004243936855345964, + 0.05442289635539055, + 0.06879983842372894, + 0.07960492372512817, + 0.06782306730747223, + 0.0493987537920475, + 0.03406734764575958, + 0.05066878721117973, + 0.03351757675409317, + 0.012073189951479435, + 0.006556784734129906, + 0.017617899924516678, + 0.005821831524372101, + 0.005111380014568567, + 0.0048066903837025166, + 0.005887569393962622, + 0.0031140344217419624, + 0.007588652428239584, + 0.023715870454907417, + 0.022119343280792236, + 0.0038260987494140863, + 0.005061762407422066, + 0.0021537430584430695, + 0.0029690763913094997, + 0.0032173600047826767, + 0.00885738804936409, + 0.011961210519075394, + 0.006703911814838648, + 0.00784257147461176, + 0.011694579385221004, + 0.004496213048696518, + 0.005340514704585075, + 0.005451630335301161, + 0.01691310480237007, + 0.0167662613093853, + 0.012766383588314056, + 0.012071216478943825, + 0.01347118429839611, + 0.006520627066493034, + 0.008450742810964584, + 0.006930491887032986, + 0.009323570877313614, + 0.007249927148222923, + 0.0076879714615643024, + 0.010463746264576912, + 0.005904147867113352, + 0.007866693660616875, + 0.009410277009010315, + 0.008199913427233696, + 0.003626057645305991, + 0.0038014475721865892, + 0.004754748195409775, + 0.005742618814110756, + 0.004319132771342993, + 0.00426339078694582, + 0.006986052729189396, + 0.003735810751095414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.012550774961709976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.2521357536315918 + ], + "opposite_logprob_tokens": [ + -1.5021357536315918 + ], + "logprob_reward": -0.2521357536315918, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521357536315918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07657548040151596, + 0.1335708200931549, + 0.12261343747377396, + 0.19142161309719086, + 0.0381384901702404, + 0.15593808889389038, + 0.025601670145988464, + 0.15404056012630463, + 0.05107172951102257, + 0.051028065383434296 + ], + "credit_map": [ + 0.01197067927569151, + 0.007632417604327202, + 0.017300086095929146, + 0.00989496149122715, + 0.0054288930259644985, + 0.006351827178150415, + 0.010425575077533722, + 0.007571043446660042, + 0.027788423001766205, + 0.006932032760232687, + 0.012553703971207142, + 0.007414937485009432, + 0.006728686857968569, + 0.011866850778460503, + 0.01761472225189209, + 0.04267147183418274, + 0.014390145428478718, + 0.01425557117909193, + 0.017316553741693497, + 0.016567640006542206, + 0.010116134770214558, + 0.00990620069205761, + 0.017814692109823227, + 0.022246504202485085, + 0.02369643747806549, + 0.02896404266357422, + 0.034518592059612274, + 0.031056534498929977, + 0.022534439340233803, + 0.014257415197789669, + 0.02342838980257511, + 0.01296578161418438, + 0.0071250260807573795, + 0.004147200845181942, + 0.009683248586952686, + 0.003746100701391697, + 0.003096499014645815, + 0.0033881112467497587, + 0.0029693334363400936, + 0.003982969094067812, + 0.015706436708569527, + 0.05098962411284447, + 0.06035325303673744, + 0.007152951788157225, + 0.00830008089542389, + 0.0048093353398144245, + 0.004194461740553379, + 0.0044319527223706245, + 0.0035616050008684397, + 0.004767309874296188, + 0.0034322314895689487, + 0.003057077992707491, + 0.004798393230885267, + 0.0017152942018583417, + 0.0023926037829369307, + 0.001877153990790248, + 0.03825269266963005, + 0.03909626230597496, + 0.02212855964899063, + 0.01884818635880947, + 0.00857651885598898, + 0.00887380912899971, + 0.009807244874536991, + 0.008457288146018982, + 0.007244989741593599, + 0.005760036874562502, + 0.007462057750672102, + 0.006529099773615599, + 0.004262672271579504, + 0.0075724562630057335, + 0.007801745552569628, + 0.004438671283423901, + 0.005213914904743433, + 0.005743062123656273, + 0.006752556189894676, + 0.007733638864010572, + 0.006506241858005524, + 0.005299913696944714, + 0.009428651072084904, + 0.004350088071078062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.2521357536315918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.11293405294418335 + ], + "opposite_logprob_tokens": [ + -2.237934112548828 + ], + "logprob_reward": -0.11293405294418335, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293405294418335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14092007279396057, + 0.06671509891748428, + 0.16180454194545746, + 0.07856280356645584, + 0.07567106187343597, + 0.07927887886762619, + 0.033332426100969315, + 0.2720445990562439, + 0.05510931834578514, + 0.03656125068664551 + ], + "credit_map": [ + 0.017854345962405205, + 0.018377220258116722, + 0.027404146268963814, + 0.015814445912837982, + 0.015485749579966068, + 0.01043345034122467, + 0.023770105093717575, + 0.01178061030805111, + 0.015846505761146545, + 0.004658084362745285, + 0.005883083678781986, + 0.004382193088531494, + 0.0030448848847299814, + 0.0052260663360357285, + 0.007172764278948307, + 0.0205015130341053, + 0.02284986712038517, + 0.015213735401630402, + 0.04156504198908806, + 0.021335871890187263, + 0.02052227593958378, + 0.00935281440615654, + 0.016843365505337715, + 0.01412157155573368, + 0.010851023718714714, + 0.011387930251657963, + 0.012472832575440407, + 0.011523592285811901, + 0.010241769254207611, + 0.005429395940154791, + 0.010361813008785248, + 0.006294445134699345, + 0.01799042336642742, + 0.00892478134483099, + 0.028268540278077126, + 0.004987917374819517, + 0.0032043808605521917, + 0.00294047431088984, + 0.003907563630491495, + 0.005446977913379669, + 0.008178740739822388, + 0.024021590128540993, + 0.027108151465654373, + 0.004482686053961515, + 0.005467064678668976, + 0.002574474783614278, + 0.0037235093768686056, + 0.003722658846527338, + 0.004257994703948498, + 0.0053903390653431416, + 0.005867638625204563, + 0.0037958784960210323, + 0.0049950373359024525, + 0.0034951004199683666, + 0.002359190955758095, + 0.003171247895807028, + 0.06402280181646347, + 0.060871049761772156, + 0.04783625155687332, + 0.025269752368330956, + 0.014986304566264153, + 0.02641342021524906, + 0.01619676686823368, + 0.01644822210073471, + 0.008603619411587715, + 0.006466525606811047, + 0.008418107405304909, + 0.007117739412933588, + 0.004056104458868504, + 0.00698104128241539, + 0.005665421485900879, + 0.007800756487995386, + 0.004511094652116299, + 0.003890734864398837, + 0.005986501462757587, + 0.004804939962923527, + 0.004552254918962717, + 0.004088441841304302, + 0.00500900624319911, + 0.0037182795349508524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.11293405294418335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera, starting from a low angle and slowly tilting upward, sweeps across the interior of an ancient temple hall. At the center of the frame lies a massive vermilion wooden column, its paint peeling slightly, revealing dark wood grain and the marks of age. A massive bronze dragon sculpture coils along this pillar, its muscular form imbued with power, its sharp claws digging deeply into the shaft. In the dim, flickering light of several eternal lamps burning overhead, each meticulously carved dragon scale reveals a mottled patina etched by age. As the camera ascends, the light streams across its winding surfaces, giving them a cold, mysterious metallic sheen before fading into deep shadow. The hall is empty and silent, the smooth black stone floor reflecting the vague outlines of the column and dragon. Wisps of smoke from incense burners in the distance drift slowly through the stagnant air, creating a solemn and mysterious atmosphere steeped in history.", + "seed": 23328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_15/index_1449_seed_23328_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": -0.09001994282007217, + "vlm_logprob": -0.09001994282007217, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.09001994282007217, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.29264411330223083 + ], + "opposite_logprob_tokens": [ + -1.4176441431045532 + ], + "logprob_reward": -0.29264411330223083, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29264411330223083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2501053512096405, + 0.06577467918395996, + 0.07456628978252411, + 0.07262375205755234, + 0.1153750866651535, + 0.08171799033880234, + 0.09310995787382126, + 0.08720562607049942, + 0.11961694806814194, + 0.03990432247519493 + ], + "credit_map": [ + 0.06011084094643593, + 0.0364191010594368, + 0.04219933971762657, + 0.019084913656115532, + 0.020029611885547638, + 0.03046059049665928, + 0.01797061786055565, + 0.023830315098166466, + 0.006173969246447086, + 0.006305229384452105, + 0.010435429401695728, + 0.006661454681307077, + 0.010647224262356758, + 0.00799745973199606, + 0.008560098707675934, + 0.00899380911141634, + 0.01097827684134245, + 0.009027757681906223, + 0.012219039723277092, + 0.00606308551505208, + 0.0098015908151865, + 0.007645157165825367, + 0.007885976694524288, + 0.010945395566523075, + 0.009267342276871204, + 0.006674077361822128, + 0.011031358502805233, + 0.004175185691565275, + 0.011677606031298637, + 0.0075660888105630875, + 0.011088439263403416, + 0.011143645271658897, + 0.012760649435222149, + 0.016551705077290535, + 0.018956990912556648, + 0.014916300773620605, + 0.010312793776392937, + 0.014140810817480087, + 0.008140716701745987, + 0.019595105201005936, + 0.00825869757682085, + 0.010780864395201206, + 0.010139175690710545, + 0.006041276268661022, + 0.014482657425105572, + 0.013474080711603165, + 0.007357473019510508, + 0.011183764785528183, + 0.010633929632604122, + 0.010816803202033043, + 0.009560241363942623, + 0.007022712379693985, + 0.012722399085760117, + 0.012274974025785923, + 0.012780416756868362, + 0.017298469319939613, + 0.006641556043177843, + 0.01696951501071453, + 0.014902551658451557, + 0.009533390402793884, + 0.011421967297792435, + 0.011107728816568851, + 0.007021541707217693, + 0.009607374668121338, + 0.011751915328204632, + 0.036283496767282486, + 0.016480712220072746, + 0.009999160654842854, + 0.013241538777947426, + 0.007054672110825777, + 0.015375979244709015, + 0.00942945946007967, + 0.005359957925975323, + 0.003871528897434473, + 0.0066283331252634525, + 0.005383090581744909, + 0.003945830278098583, + 0.0045749652199447155, + 0.00632371986284852, + 0.003816891461610794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.29264411330223083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.02060309424996376 + ], + "opposite_logprob_tokens": [ + -3.8956031799316406 + ], + "logprob_reward": -0.02060309424996376, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060309424996376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2117185890674591, + 0.2473534196615219, + 0.03877873346209526, + 0.0989142581820488, + 0.16973617672920227, + 0.049407701939344406, + 0.048293035477399826, + 0.025781195610761642, + 0.07609391212463379, + 0.033922988921403885 + ], + "credit_map": [ + 0.03511649742722511, + 0.02544758841395378, + 0.02969631925225258, + 0.0158101674169302, + 0.02808661386370659, + 0.02579660527408123, + 0.026224149391055107, + 0.02554064802825451, + 0.01345694437623024, + 0.02419741451740265, + 0.039404891431331635, + 0.02500862628221512, + 0.04498305171728134, + 0.0254681333899498, + 0.04871062934398651, + 0.02612372301518917, + 0.003841399447992444, + 0.00466768117621541, + 0.006405819673091173, + 0.0045452238991856575, + 0.005937281530350447, + 0.0033565526828169823, + 0.006026914808899164, + 0.003997858613729477, + 0.009742596186697483, + 0.010415584780275822, + 0.0152636943385005, + 0.009253062307834625, + 0.014466972090303898, + 0.010610997676849365, + 0.014214891940355301, + 0.014946462586522102, + 0.015676993876695633, + 0.02519935742020607, + 0.024414803832769394, + 0.017005348578095436, + 0.027696896344423294, + 0.016804933547973633, + 0.014148671180009842, + 0.028789179399609566, + 0.00458988081663847, + 0.0073318686336278915, + 0.006350066512823105, + 0.004539818968623877, + 0.01053403690457344, + 0.003995393868535757, + 0.006524524185806513, + 0.0055421097204089165, + 0.006549736484885216, + 0.005093120038509369, + 0.005938844755291939, + 0.0040714312344789505, + 0.006337338127195835, + 0.005866266321390867, + 0.006177095230668783, + 0.008259204216301441, + 0.0023322133347392082, + 0.003294931026175618, + 0.006524361204355955, + 0.0020419477950781584, + 0.002454433124512434, + 0.0023613839875906706, + 0.002857211511582136, + 0.003914716187864542, + 0.005888462532311678, + 0.02883009985089302, + 0.0174945630133152, + 0.0036993424873799086, + 0.007224787957966328, + 0.0038728981744498014, + 0.004710868000984192, + 0.004372893366962671, + 0.005050244741141796, + 0.004206540994346142, + 0.0065244571305811405, + 0.0039649466052651405, + 0.004214245360344648, + 0.0033245894592255354, + 0.0036004954017698765, + 0.0030374708585441113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.02060309424996376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.018195264041423798 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195264041423798, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195264041423798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07085272669792175, + 0.07486505061388016, + 0.03906827047467232, + 0.1535298228263855, + 0.1915312558412552, + 0.14366760849952698, + 0.07790198922157288, + 0.046346649527549744, + 0.11631506681442261, + 0.08592153340578079 + ], + "credit_map": [ + 0.014867386780679226, + 0.00970077607780695, + 0.011484011076390743, + 0.004679649602621794, + 0.006683464162051678, + 0.01111900806427002, + 0.00630829855799675, + 0.006010141689330339, + 0.009235492907464504, + 0.010143794119358063, + 0.011448277160525322, + 0.0067560444585978985, + 0.011812516488134861, + 0.00874279160052538, + 0.010372331365942955, + 0.006353810429573059, + 0.0037589932326227427, + 0.004852946847677231, + 0.005804150830954313, + 0.0037858339492231607, + 0.0055805835872888565, + 0.004543183371424675, + 0.006775721907615662, + 0.003966860473155975, + 0.013320490717887878, + 0.014826476573944092, + 0.024443762376904488, + 0.014554443769156933, + 0.022532641887664795, + 0.015908807516098022, + 0.022584928199648857, + 0.025358298793435097, + 0.01822011172771454, + 0.027317950502038002, + 0.02769751287996769, + 0.020201751962304115, + 0.028642334043979645, + 0.01884305477142334, + 0.01594671607017517, + 0.034661829471588135, + 0.010270890779793262, + 0.025141259655356407, + 0.018102718517184258, + 0.010770825669169426, + 0.03538757562637329, + 0.010815359652042389, + 0.016897279769182205, + 0.01628170907497406, + 0.010208225809037685, + 0.007886763662099838, + 0.010030101053416729, + 0.007216959726065397, + 0.009660353884100914, + 0.009688564576208591, + 0.009769590571522713, + 0.013441434130072594, + 0.004513176623731852, + 0.007098584435880184, + 0.00895417109131813, + 0.004241440445184708, + 0.005151866935193539, + 0.0061170789413154125, + 0.004227647557854652, + 0.006042692810297012, + 0.008660570718348026, + 0.03067968785762787, + 0.02089235559105873, + 0.011161113157868385, + 0.01231432892382145, + 0.008391373790800571, + 0.015651952475309372, + 0.008563699200749397, + 0.015073634684085846, + 0.012017746455967426, + 0.011473194696009159, + 0.00972246564924717, + 0.014185257256031036, + 0.007000986952334642, + 0.009407681412994862, + 0.00704057514667511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.018195264041423798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.048654209822416306 + ], + "opposite_logprob_tokens": [ + -3.048654317855835 + ], + "logprob_reward": -0.048654209822416306, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654209822416306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18190661072731018, + 0.08618667721748352, + 0.05919428914785385, + 0.07942170649766922, + 0.08003054559230804, + 0.09951207786798477, + 0.11133760213851929, + 0.07122641801834106, + 0.17555281519889832, + 0.05563115328550339 + ], + "credit_map": [ + 0.022950727492570877, + 0.026986239477992058, + 0.02749485895037651, + 0.018987054005265236, + 0.015860112383961678, + 0.028421742841601372, + 0.01583094149827957, + 0.02537495456635952, + 0.008886942639946938, + 0.011098233982920647, + 0.014136718586087227, + 0.008503816090524197, + 0.011270686984062195, + 0.011083576828241348, + 0.00836371909826994, + 0.012842989526689053, + 0.007573065347969532, + 0.00904647447168827, + 0.010809618048369884, + 0.006298385560512543, + 0.006006618961691856, + 0.005854559130966663, + 0.005404372233897448, + 0.008201203308999538, + 0.007682034280151129, + 0.007044091355055571, + 0.01648828573524952, + 0.008198142051696777, + 0.009437358938157558, + 0.006641201209276915, + 0.007020836230367422, + 0.01690976321697235, + 0.009293599054217339, + 0.010196415707468987, + 0.008928218856453896, + 0.0058312867768108845, + 0.01656346581876278, + 0.011948075145483017, + 0.00562694389373064, + 0.011642553843557835, + 0.007836761884391308, + 0.014314597472548485, + 0.01474726665765047, + 0.010917951352894306, + 0.02160775102674961, + 0.010275696404278278, + 0.009479732252657413, + 0.010332325473427773, + 0.01531759463250637, + 0.010876359418034554, + 0.013133401982486248, + 0.010532697662711143, + 0.011980823241174221, + 0.013395072892308235, + 0.012521733529865742, + 0.023579930886626244, + 0.006034972611814737, + 0.0099056176841259, + 0.01707378774881363, + 0.005888644140213728, + 0.007567240856587887, + 0.009442925453186035, + 0.007050219923257828, + 0.008263017982244492, + 0.014495509676635265, + 0.07263967394828796, + 0.04305309057235718, + 0.0074398936703801155, + 0.013658947311341763, + 0.009584084153175354, + 0.009061083197593689, + 0.005620559677481651, + 0.011217492632567883, + 0.0064713675528764725, + 0.009390373714268208, + 0.005142172798514366, + 0.007375241722911596, + 0.006104584783315659, + 0.004362968727946281, + 0.005566950421780348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.048654209822416306 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 a low angle and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive vermilion wooden column visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive bronze dragon sculpture coiled along the pillar with sharp claws?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the floor made of smooth black stone reflecting the column and dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of smoke from incense burners drifting through the air 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": [ + -0.07000303268432617 + ], + "opposite_logprob_tokens": [ + -2.695003032684326 + ], + "logprob_reward": -0.07000303268432617, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000303268432617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09558457881212234, + 0.0956156775355339, + 0.04686060920357704, + 0.061018455773591995, + 0.09443540126085281, + 0.11057214438915253, + 0.0730830505490303, + 0.09046068787574768, + 0.23974373936653137, + 0.09262563288211823 + ], + "credit_map": [ + 0.011726031079888344, + 0.013853579759597778, + 0.013171284459531307, + 0.00920113641768694, + 0.012471095658838749, + 0.011038529686629772, + 0.011832349002361298, + 0.01229057926684618, + 0.012645007111132145, + 0.011259626597166061, + 0.015055044554173946, + 0.011153051629662514, + 0.014040491543710232, + 0.010179994627833366, + 0.01016897615045309, + 0.011113500222563744, + 0.006050465162843466, + 0.008120561018586159, + 0.005827655550092459, + 0.0036416377406567335, + 0.005525416694581509, + 0.006363308057188988, + 0.004811080638319254, + 0.006520487368106842, + 0.0058578322641551495, + 0.00621382798999548, + 0.009504569694399834, + 0.005186638794839382, + 0.011501936241984367, + 0.006434039678424597, + 0.007408732082694769, + 0.008910886943340302, + 0.009567097760736942, + 0.013823715038597584, + 0.01261750515550375, + 0.009636937640607357, + 0.015849752351641655, + 0.009943618439137936, + 0.007631601765751839, + 0.015365193597972393, + 0.009830480441451073, + 0.021015282720327377, + 0.010829467326402664, + 0.008790621533989906, + 0.021401410922408104, + 0.013921964913606644, + 0.01103434432297945, + 0.013748588971793652, + 0.007993309758603573, + 0.008404429070651531, + 0.00849138107150793, + 0.005702638998627663, + 0.011680709198117256, + 0.008801938965916634, + 0.011123141273856163, + 0.010885505937039852, + 0.007244078908115625, + 0.015845250338315964, + 0.014237908646464348, + 0.010088544338941574, + 0.010490481741726398, + 0.012942098081111908, + 0.011924171820282936, + 0.007688173558562994, + 0.018960945308208466, + 0.04629155248403549, + 0.03352866321802139, + 0.024487284943461418, + 0.03360328823328018, + 0.023908335715532303, + 0.037423595786094666, + 0.021540101617574692, + 0.014851689338684082, + 0.010790427215397358, + 0.014757679775357246, + 0.009130405262112617, + 0.014393211342394352, + 0.008028684183955193, + 0.010418361984193325, + 0.010255181230604649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.07000303268432617 + } + ], + "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/0000058/group_2/index_tensor([5844, 5844])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_tensor([5844, 5844])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8bd4681bdc19e99c1345abf05899a365277aae1c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_tensor([5844, 5844])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro shot begins a slow, revealing advance as a vibrant lime-green caterpillar, its segmented body textured with a soft, velvety texture, methodically nibbles at the edge of a broad, dew-covered leaf. Soft, golden morning sunlight slants across the frame, illuminating the scene and causing the crystal-clear, spherical dewdrops scattered across the leaf's surface to shimmer like tiny gemstones, refracting the light. As the camera inconspicuously approaches, the extremely shallow depth of field melts the background into a creamy, painterly bokeh of lush greens and earthy browns. The caterpillar's tiny, powerful jaws take center stage, rhythmically snipping at the delicate fragments of the leaf with deliberate, focused precision, its head moving in a steady, repetitive arc. The entire scene is imbued with an atmosphere of tranquility, a quiet moment of natural life observed with intimate and breathtaking clarity.", + "seed": 14693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_5844_seed_14693_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": -0.01910946317948401, + "vlm_logprob": -0.01910946317948401, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.01910946317948401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03590713441371918 + ], + "opposite_logprob_tokens": [ + -3.410907030105591 + ], + "logprob_reward": -0.03590713441371918, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03590713441371918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06372405588626862, + 0.04320826381444931, + 0.1576511710882187, + 0.0649668276309967, + 0.35383909940719604, + 0.07197694480419159, + 0.04006403684616089, + 0.13184788823127747, + 0.03309296444058418, + 0.039628766477108 + ], + "credit_map": [ + 0.012174106203019619, + 0.005880236625671387, + 0.005820680409669876, + 0.004590889438986778, + 0.01760171353816986, + 0.009620198979973793, + 0.00469861039891839, + 0.00333762145601213, + 0.007179374806582928, + 0.00411113491281867, + 0.007078113965690136, + 0.0034251068718731403, + 0.003800170263275504, + 0.0033572178799659014, + 0.008032255806028843, + 0.006224888376891613, + 0.011230887845158577, + 0.005259013269096613, + 0.015539894811809063, + 0.007099321577697992, + 0.0617479532957077, + 0.043078482151031494, + 0.007598718162626028, + 0.006096897646784782, + 0.02125529944896698, + 0.0040525589138269424, + 0.007285141386091709, + 0.0030328487046062946, + 0.004977644421160221, + 0.016874181106686592, + 0.0035349375102669, + 0.003954211715608835, + 0.09655621647834778, + 0.02213086187839508, + 0.03795652836561203, + 0.02767038345336914, + 0.034244731068611145, + 0.06876052170991898, + 0.03575199469923973, + 0.030767876654863358, + 0.010534673929214478, + 0.009560930542647839, + 0.006117859855294228, + 0.00584487384185195, + 0.005482562817633152, + 0.015305914916098118, + 0.008147142827510834, + 0.010982978157699108, + 0.0036497269757092, + 0.008894204162061214, + 0.00395537493750453, + 0.004759357776492834, + 0.00574676925316453, + 0.005301886238157749, + 0.004091618582606316, + 0.0036650984548032284, + 0.01550951600074768, + 0.020633073523640633, + 0.017649205401539803, + 0.012821845710277557, + 0.015691306442022324, + 0.0228461641818285, + 0.010882982984185219, + 0.015813780948519707, + 0.007892762310802937, + 0.0034338480327278376, + 0.002964404411613941, + 0.0032094446942210197, + 0.002938907127827406, + 0.009002039209008217, + 0.0014228919753804803, + 0.0022286660969257355, + 0.00983521994203329, + 0.004377292934805155, + 0.00370968459174037, + 0.00347069650888443, + 0.005337066482752562, + 0.005903199780732393, + 0.0038098914083093405, + 0.0031857152935117483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.03590713441371918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371112421154976 + ], + "opposite_logprob_tokens": [ + -3.4087111949920654 + ], + "logprob_reward": -0.03371112421154976, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371112421154976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1415233016014099, + 0.0534842424094677, + 0.09191521257162094, + 0.030426718294620514, + 0.22625069320201874, + 0.14125266671180725, + 0.09006201475858688, + 0.13510535657405853, + 0.03681497275829315, + 0.05316481366753578 + ], + "credit_map": [ + 0.0269253421574831, + 0.011849579401314259, + 0.02954849787056446, + 0.011620261706411839, + 0.018004557117819786, + 0.013029256835579872, + 0.014924542978405952, + 0.015621265396475792, + 0.011311660520732403, + 0.006165284663438797, + 0.012303492985665798, + 0.0051245056092739105, + 0.0059005809016525745, + 0.004002457018941641, + 0.0042298706248402596, + 0.004446391947567463, + 0.01788339577615261, + 0.0042852978222072124, + 0.011803528293967247, + 0.004839617293328047, + 0.03120221197605133, + 0.005762467626482248, + 0.00828230008482933, + 0.007856391370296478, + 0.007747896481305361, + 0.0016771408263593912, + 0.0038565213326364756, + 0.001698532490991056, + 0.003108296077698469, + 0.0073584322817623615, + 0.002143374877050519, + 0.0028365214820951223, + 0.033776137977838516, + 0.012682690285146236, + 0.049503143876791, + 0.020230799913406372, + 0.0195061843842268, + 0.033056002110242844, + 0.02894054725766182, + 0.02855518087744713, + 0.0133616728708148, + 0.010992923751473427, + 0.020182207226753235, + 0.010969668626785278, + 0.01260034553706646, + 0.014739532954990864, + 0.019624482840299606, + 0.03878183290362358, + 0.015216114930808544, + 0.006777651607990265, + 0.008140445686876774, + 0.007262442260980606, + 0.005478918552398682, + 0.01879948191344738, + 0.012062955647706985, + 0.016324006021022797, + 0.016732024028897285, + 0.020762557163834572, + 0.017711646854877472, + 0.014302532188594341, + 0.017336975783109665, + 0.015321183949708939, + 0.012915926054120064, + 0.02002250961959362, + 0.0038483201060444117, + 0.007128998637199402, + 0.004925235640257597, + 0.004734923597425222, + 0.0056618801318109035, + 0.003593214089050889, + 0.0030516001861542463, + 0.0038708022329956293, + 0.01162505429238081, + 0.006416278891265392, + 0.0051996540278196335, + 0.00696911383420229, + 0.0069008516147732735, + 0.008047898299992085, + 0.004181216470897198, + 0.0038247446063905954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.03371112421154976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258538294583559 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258538294583559, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258538294583559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23550759255886078, + 0.06114063039422035, + 0.2235075682401657, + 0.034350987523794174, + 0.08765070885419846, + 0.08156465739011765, + 0.07533813267946243, + 0.09305844455957413, + 0.07655052840709686, + 0.03133078292012215 + ], + "credit_map": [ + 0.04907107353210449, + 0.02111414261162281, + 0.04291076213121414, + 0.02122446708381176, + 0.022490033879876137, + 0.019158190116286278, + 0.03120071440935135, + 0.02833818644285202, + 0.007521083578467369, + 0.004177649971097708, + 0.00807413924485445, + 0.005022108089178801, + 0.011099761351943016, + 0.012569059617817402, + 0.005984524264931679, + 0.006692299619317055, + 0.024690985679626465, + 0.011937689036130905, + 0.030981935560703278, + 0.015996940433979034, + 0.036108072847127914, + 0.07897236943244934, + 0.012492252513766289, + 0.012327296659350395, + 0.00954283308237791, + 0.004059420432895422, + 0.0033075669780373573, + 0.0026621101424098015, + 0.004390887450426817, + 0.006740417797118425, + 0.0018228119006380439, + 0.0018249378772452474, + 0.02292253077030182, + 0.004681659862399101, + 0.017720740288496017, + 0.0060605332255363464, + 0.008760256692767143, + 0.014033111743628979, + 0.007562494371086359, + 0.005909375846385956, + 0.012007173150777817, + 0.007806013338267803, + 0.02562205120921135, + 0.0058932011015713215, + 0.006525970995426178, + 0.011638805270195007, + 0.0049717058427631855, + 0.007099729031324387, + 0.011064494960010052, + 0.010318281129002571, + 0.007005825638771057, + 0.007451871410012245, + 0.005860508419573307, + 0.01187591627240181, + 0.008423324674367905, + 0.013337905518710613, + 0.010885891504585743, + 0.011112994514405727, + 0.011166487820446491, + 0.010607695206999779, + 0.00939131435006857, + 0.011738207191228867, + 0.014391281642019749, + 0.01376456581056118, + 0.019777696579694748, + 0.007052876520901918, + 0.00516715832054615, + 0.007528102491050959, + 0.006002402864396572, + 0.023412415757775307, + 0.0038580228574573994, + 0.0037518402095884085, + 0.004108336288481951, + 0.003814778057858348, + 0.003672899678349495, + 0.0028759895358234644, + 0.005324214696884155, + 0.004676464945077896, + 0.0022180594969540834, + 0.004640036728233099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.005258538294583559 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818285696208477 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818285696208477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818285696208477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25278669595718384, + 0.09396116435527802, + 0.154503732919693, + 0.07061199843883514, + 0.1056959480047226, + 0.12089108675718307, + 0.08469358086585999, + 0.02408972755074501, + 0.044957369565963745, + 0.0478086955845356 + ], + "credit_map": [ + 0.044188790023326874, + 0.031073754653334618, + 0.04779176041483879, + 0.023349910974502563, + 0.03167573735117912, + 0.024083487689495087, + 0.02554386667907238, + 0.02507939003407955, + 0.01296385657042265, + 0.005969924386590719, + 0.013978324830532074, + 0.006281924434006214, + 0.032903339713811874, + 0.008053460158407688, + 0.006744868587702513, + 0.007065463345497847, + 0.015006419271230698, + 0.006517082918435335, + 0.032715968787670135, + 0.008133643306791782, + 0.06184406206011772, + 0.01943989098072052, + 0.006058459635823965, + 0.004788198508322239, + 0.006040266249328852, + 0.002804493298754096, + 0.020767971873283386, + 0.005847543478012085, + 0.00493892515078187, + 0.02081364206969738, + 0.004937135614454746, + 0.004462013486772776, + 0.028874514624476433, + 0.008875762112438679, + 0.017474018037319183, + 0.006891954690217972, + 0.008834966458380222, + 0.021700678393244743, + 0.006293362472206354, + 0.006750688888132572, + 0.009356830269098282, + 0.008097057230770588, + 0.039801787585020065, + 0.008972937241196632, + 0.008063623681664467, + 0.031834788620471954, + 0.005949822254478931, + 0.008814245462417603, + 0.007974807173013687, + 0.01405088510364294, + 0.010413982905447483, + 0.008538870140910149, + 0.008532566018402576, + 0.01145174726843834, + 0.010006600059568882, + 0.013724121265113354, + 0.0026620347052812576, + 0.0040708547458052635, + 0.002743826713413, + 0.0022114913444966078, + 0.0037139502819627523, + 0.0028016222640872, + 0.0026554649230092764, + 0.003230483504012227, + 0.0056996578350663185, + 0.007625971920788288, + 0.0063446988351643085, + 0.006002086214721203, + 0.0064408802427351475, + 0.005155548453330994, + 0.003466830588877201, + 0.004221697803586721, + 0.006694743409752846, + 0.00621003657579422, + 0.004973890725523233, + 0.005248645786195993, + 0.006372789852321148, + 0.006993245333433151, + 0.0030962785240262747, + 0.008219067938625813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01818285696208477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.24067986011505127, + 0.057047877460718155, + 0.1253442019224167, + 0.059315651655197144, + 0.07087583094835281, + 0.12090126425027847, + 0.045382604002952576, + 0.062464356422424316, + 0.12511534988880157, + 0.09287295490503311 + ], + "credit_map": [ + 0.024414347484707832, + 0.02939039282500744, + 0.036526814103126526, + 0.027204787358641624, + 0.03495858982205391, + 0.022340619936585426, + 0.03203386440873146, + 0.033810440450906754, + 0.004833062179386616, + 0.00523379584774375, + 0.00731220468878746, + 0.004004966467618942, + 0.01597924344241619, + 0.010037168860435486, + 0.005012225359678268, + 0.00463521433994174, + 0.01101031992584467, + 0.008380116894841194, + 0.03168101981282234, + 0.008791562169790268, + 0.03440181538462639, + 0.010711475275456905, + 0.010859264992177486, + 0.009508630260825157, + 0.004269004799425602, + 0.0039015132933855057, + 0.018644412979483604, + 0.004865885246545076, + 0.004287279210984707, + 0.017399365082383156, + 0.002907716203480959, + 0.003040473908185959, + 0.018291642889380455, + 0.005157084669917822, + 0.021131837740540504, + 0.00564056308940053, + 0.0051654065027832985, + 0.004816650878638029, + 0.0052334885112941265, + 0.005439160391688347, + 0.014179197140038013, + 0.011824483051896095, + 0.03482386842370033, + 0.009025990962982178, + 0.009447802789509296, + 0.02062741480767727, + 0.00706061627715826, + 0.013911888003349304, + 0.007797669619321823, + 0.006109222769737244, + 0.0037603662349283695, + 0.004135689232498407, + 0.004152461886405945, + 0.00822874903678894, + 0.0037282626144587994, + 0.0074701858684420586, + 0.005807270295917988, + 0.01200422365218401, + 0.007113329600542784, + 0.00608024513348937, + 0.009045238606631756, + 0.008529704995453358, + 0.006412963382899761, + 0.007471377961337566, + 0.038732606917619705, + 0.006213037762790918, + 0.007469288073480129, + 0.012449152767658234, + 0.00615833792835474, + 0.04500827193260193, + 0.0037147279363125563, + 0.005369929596781731, + 0.017515704035758972, + 0.011494872160255909, + 0.008451846428215504, + 0.010543635115027428, + 0.011949339881539345, + 0.01540546678006649, + 0.006482643540948629, + 0.011029460467398167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0024876620154827833 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot begins a slow, revealing advance as a vibrant lime-green caterpillar, its segmented body textured with a soft, velvety texture, methodically nibbles at the edge of a broad, dew-covered leaf. Soft, golden morning sunlight slants across the frame, illuminating the scene and causing the crystal-clear, spherical dewdrops scattered across the leaf's surface to shimmer like tiny gemstones, refracting the light. As the camera inconspicuously approaches, the extremely shallow depth of field melts the background into a creamy, painterly bokeh of lush greens and earthy browns. The caterpillar's tiny, powerful jaws take center stage, rhythmically snipping at the delicate fragments of the leaf with deliberate, focused precision, its head moving in a steady, repetitive arc. The entire scene is imbued with an atmosphere of tranquility, a quiet moment of natural life observed with intimate and breathtaking clarity.", + "seed": 14694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_5844_seed_14694_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": -0.027532493020407854, + "vlm_logprob": -0.027532493020407854, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.027532493020407854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004711597226560116 + ], + "opposite_logprob_tokens": [ + -5.754711627960205 + ], + "logprob_reward": -0.004711597226560116, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004711597226560116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6880346536636353, + 0.034506745636463165, + 0.022547706961631775, + 0.029275748878717422, + 0.06667333096265793, + 0.038385216146707535, + 0.025558602064847946, + 0.028658775612711906, + 0.04768766835331917, + 0.01867164857685566 + ], + "credit_map": [ + 0.0325205959379673, + 0.27335986495018005, + 0.013463851995766163, + 0.009532764554023743, + 0.2813445031642914, + 0.026983637362718582, + 0.0209041777998209, + 0.029925160109996796, + 0.0034564982634037733, + 0.006127774715423584, + 0.004864860791712999, + 0.003643902949988842, + 0.004744038917124271, + 0.006730067543685436, + 0.0025913158897310495, + 0.002348280046135187, + 0.0022456569131463766, + 0.0028666576836258173, + 0.0034185207914561033, + 0.004002614878118038, + 0.0031810933724045753, + 0.0025303582660853863, + 0.002659906167536974, + 0.0016428983071818948, + 0.0019088519038632512, + 0.011221572756767273, + 0.008843577466905117, + 0.0018129851669073105, + 0.002664055209606886, + 0.0010594463674351573, + 0.0008919977117329836, + 0.0008732615387998521, + 0.003913187421858311, + 0.005664056167006493, + 0.018807554617524147, + 0.00287437136285007, + 0.026773059740662575, + 0.004408501088619232, + 0.002302673179656267, + 0.0019299216801300645, + 0.007585075218230486, + 0.0035847811959683895, + 0.012010917067527771, + 0.004060239065438509, + 0.0031970690470188856, + 0.0031526810489594936, + 0.0025842394679784775, + 0.002210209844633937, + 0.0039382204413414, + 0.0022475512232631445, + 0.005938246380537748, + 0.003996015060693026, + 0.0025239423848688602, + 0.002557669533416629, + 0.0022471470292657614, + 0.002109808148816228, + 0.005636756774038076, + 0.0037455037236213684, + 0.005575953051447868, + 0.002954353578388691, + 0.0028219891246408224, + 0.0032837195321917534, + 0.0023328764364123344, + 0.002307621529325843, + 0.005257023498415947, + 0.004491577390581369, + 0.007574302144348621, + 0.003660752670839429, + 0.006686416454613209, + 0.013123155571520329, + 0.002889786846935749, + 0.004004649817943573, + 0.002157346112653613, + 0.002037770114839077, + 0.0028798843268305063, + 0.0028031652327626944, + 0.0023056459613144398, + 0.0021037261467427015, + 0.0024153904523700476, + 0.001968716736882925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.004711597226560116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295055598020554 + ], + "opposite_logprob_tokens": [ + -2.237950563430786 + ], + "logprob_reward": -0.11295055598020554, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295055598020554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4472143054008484, + 0.020100722089409828, + 0.07261461019515991, + 0.061972156167030334, + 0.19565585255622864, + 0.07781708985567093, + 0.018013056367635727, + 0.04615550488233566, + 0.0416484959423542, + 0.018808219581842422 + ], + "credit_map": [ + 0.02046588622033596, + 0.17367523908615112, + 0.009172908961772919, + 0.00877088401466608, + 0.18258921802043915, + 0.018236488103866577, + 0.014489807188510895, + 0.01981387473642826, + 0.0019459977047517896, + 0.004634180571883917, + 0.0024954471737146378, + 0.0020748868118971586, + 0.003285541431978345, + 0.0023835529573261738, + 0.0017867814749479294, + 0.001494331401772797, + 0.0074243261478841305, + 0.010966537520289421, + 0.012992057017982006, + 0.013378300704061985, + 0.008488734252750874, + 0.007321469951421022, + 0.00625763600692153, + 0.005785549525171518, + 0.004359338898211718, + 0.02398882992565632, + 0.018150387331843376, + 0.0027268787380307913, + 0.006133622955530882, + 0.0031529082916677, + 0.0014176270924508572, + 0.002042565029114485, + 0.01379281934350729, + 0.01668216660618782, + 0.05937156081199646, + 0.01236086804419756, + 0.07290282845497131, + 0.011177766136825085, + 0.0042916410602629185, + 0.005076191388070583, + 0.017867228016257286, + 0.0065977126359939575, + 0.026380587369203568, + 0.004599550738930702, + 0.0073651643469929695, + 0.006326314527541399, + 0.004461788106709719, + 0.004218738526105881, + 0.003054531989619136, + 0.0020169566851109266, + 0.0023114862851798534, + 0.002600714797154069, + 0.002259699394926429, + 0.003060686867684126, + 0.001358109526336193, + 0.0013508708216249943, + 0.009751105681061745, + 0.004955240525305271, + 0.015542272478342056, + 0.003700227476656437, + 0.003379727713763714, + 0.004068450070917606, + 0.002259537810459733, + 0.002498941496014595, + 0.006258565466850996, + 0.004193668253719807, + 0.0088926637545228, + 0.00416594697162509, + 0.0053391605615615845, + 0.005340635310858488, + 0.002757409820333123, + 0.004700444173067808, + 0.0020979931578040123, + 0.0024188982788473368, + 0.0027467927429825068, + 0.0023204030003398657, + 0.0022173007018864155, + 0.003052388085052371, + 0.0020738972816616297, + 0.0018805458676069975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.11295055598020554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040998938493430614 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040998938493430614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040998938493430614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.663479745388031, + 0.042210157960653305, + 0.04738891124725342, + 0.05132710561156273, + 0.058501891791820526, + 0.036662306636571884, + 0.011031609028577805, + 0.040987927466630936, + 0.03208436444401741, + 0.016325999051332474 + ], + "credit_map": [ + 0.031050734221935272, + 0.2633644640445709, + 0.012123910710215569, + 0.009417911060154438, + 0.2726251780986786, + 0.026070816442370415, + 0.020181851461529732, + 0.028644852340221405, + 0.004846126306802034, + 0.006803108379244804, + 0.007197560276836157, + 0.006888832431286573, + 0.005214111413806677, + 0.004594835452735424, + 0.003548414446413517, + 0.003117172745987773, + 0.0041270931251347065, + 0.005236003082245588, + 0.006503480952233076, + 0.009109199978411198, + 0.006378077436238527, + 0.006505791563540697, + 0.0061317551881074905, + 0.0033975064288824797, + 0.004650425165891647, + 0.020219000056385994, + 0.010717769153416157, + 0.0031200775410979986, + 0.006291247438639402, + 0.0026651376392692327, + 0.0019247627351433039, + 0.0017386869294568896, + 0.0061717526987195015, + 0.003351285820826888, + 0.020914124324917793, + 0.0032012879382818937, + 0.01865004561841488, + 0.0027350501623004675, + 0.0016075740568339825, + 0.0018707711715251207, + 0.006108331959694624, + 0.003552396548911929, + 0.007541024591773748, + 0.004341131541877985, + 0.0034930442925542593, + 0.004734243731945753, + 0.0037383632734417915, + 0.0031537709292024374, + 0.0020988192409276962, + 0.0008876249776221812, + 0.002922477899119258, + 0.0010838956804946065, + 0.0009672672604210675, + 0.001536962459795177, + 0.00077426916686818, + 0.0007602925761602819, + 0.006215525791049004, + 0.004082537721842527, + 0.009887773543596268, + 0.004687412176281214, + 0.004365285858511925, + 0.005262186285108328, + 0.0033520585857331753, + 0.003135145641863346, + 0.003584200283512473, + 0.0028991352301090956, + 0.005038909614086151, + 0.002767130732536316, + 0.004631293937563896, + 0.008711445145308971, + 0.002000658307224512, + 0.002451592590659857, + 0.002069704467430711, + 0.0018867090111598372, + 0.0025998088531196117, + 0.0022785270120948553, + 0.001975015504285693, + 0.0018391008488833904, + 0.001996520208194852, + 0.0016806134954094887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0040998938493430614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188013039529324 + ], + "opposite_logprob_tokens": [ + -4.264187812805176 + ], + "logprob_reward": -0.014188013039529324, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188013039529324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46874386072158813, + 0.04759356752038002, + 0.04923688620328903, + 0.025045379996299744, + 0.04458007216453552, + 0.07688620686531067, + 0.0511941984295845, + 0.1544780731201172, + 0.04248058795928955, + 0.039761170744895935 + ], + "credit_map": [ + 0.02238469198346138, + 0.18559369444847107, + 0.008903316222131252, + 0.007479452062398195, + 0.18903155624866486, + 0.01922762766480446, + 0.015126598998904228, + 0.020996909588575363, + 0.005436552222818136, + 0.007231162395328283, + 0.008870227262377739, + 0.006908888462930918, + 0.006984783336520195, + 0.005590120330452919, + 0.00355870951898396, + 0.0030131321400403976, + 0.003501163562759757, + 0.005094117484986782, + 0.005065226461738348, + 0.011328499764204025, + 0.005992224905639887, + 0.006898841354995966, + 0.006555141415446997, + 0.0048016770742833614, + 0.003957855049520731, + 0.0059828124940395355, + 0.0028582962695509195, + 0.0028639035299420357, + 0.0026540483813732862, + 0.0025901857297867537, + 0.002094198716804385, + 0.0020440826192498207, + 0.0033532357774674892, + 0.004014462232589722, + 0.011375428177416325, + 0.003718406194821, + 0.01432612631469965, + 0.003810612251982093, + 0.0022359660360962152, + 0.0017458385555073619, + 0.010375014506280422, + 0.015998665243387222, + 0.009497018530964851, + 0.013886314816772938, + 0.009782818146049976, + 0.00663926312699914, + 0.004984637256711721, + 0.005722479894757271, + 0.011568734422326088, + 0.005136173218488693, + 0.015665952116250992, + 0.005276822485029697, + 0.00344547675922513, + 0.0038235289976000786, + 0.003021462121978402, + 0.0032560541294515133, + 0.02730211429297924, + 0.0159167293459177, + 0.03938092663884163, + 0.014416614547371864, + 0.015148291364312172, + 0.01994463987648487, + 0.011127601377665997, + 0.011241168715059757, + 0.004661038517951965, + 0.004703474696725607, + 0.004761268850415945, + 0.007282162085175514, + 0.004843533970415592, + 0.008999611251056194, + 0.0031066613737493753, + 0.004122836980968714, + 0.004733634646981955, + 0.004291612189263105, + 0.0060322037898004055, + 0.0069208210334181786, + 0.00372400414198637, + 0.005218698177486658, + 0.004554618615657091, + 0.004285583272576332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.014188013039529324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017124050064012408 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017124050064012408, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017124050064012408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3960935175418854, + 0.14612174034118652, + 0.04857577010989189, + 0.06942383199930191, + 0.08178158104419708, + 0.08637169748544693, + 0.022932453081011772, + 0.0617956779897213, + 0.034872811287641525, + 0.05203090235590935 + ], + "credit_map": [ + 0.019496342167258263, + 0.14795196056365967, + 0.01643214002251625, + 0.013401064090430737, + 0.14982636272907257, + 0.015522914938628674, + 0.014620296657085419, + 0.018842432647943497, + 0.016506122425198555, + 0.027438290417194366, + 0.026722989976406097, + 0.019722629338502884, + 0.01895863562822342, + 0.01769232749938965, + 0.010186860337853432, + 0.008893875405192375, + 0.004047216381877661, + 0.005524627398699522, + 0.00808179471641779, + 0.008679886348545551, + 0.0061844405718147755, + 0.0073244585655629635, + 0.004687510430812836, + 0.0040458329021930695, + 0.008309198543429375, + 0.02610839530825615, + 0.011082385666668415, + 0.0046347384341061115, + 0.008219228126108646, + 0.004519060254096985, + 0.0033628966193646193, + 0.003187925321981311, + 0.014871574006974697, + 0.012773009948432446, + 0.0039907777681946754, + 0.00961154606193304, + 0.026554135605692863, + 0.007523398846387863, + 0.0030773866456001997, + 0.003379747737199068, + 0.01849951595067978, + 0.00815424881875515, + 0.027949688956141472, + 0.006336399354040623, + 0.008693165145814419, + 0.005756468046456575, + 0.005437333602458239, + 0.005544872023165226, + 0.004145110957324505, + 0.0024528380017727613, + 0.004174920730292797, + 0.0030338687356561422, + 0.002524478128179908, + 0.003179851220920682, + 0.0016617848305031657, + 0.0017595987301319838, + 0.010622935369610786, + 0.005638140253722668, + 0.01566440239548683, + 0.005784991662949324, + 0.00652581499889493, + 0.00822995975613594, + 0.004707525484263897, + 0.004621909465640783, + 0.004673187155276537, + 0.00337926740758121, + 0.0044626519083976746, + 0.003117844695225358, + 0.00417170487344265, + 0.008979571983218193, + 0.0020243851467967033, + 0.004064198583364487, + 0.0053094071336090565, + 0.004827162250876427, + 0.00938365701586008, + 0.009858937002718449, + 0.004589795600622892, + 0.004632631316781044, + 0.005545597989112139, + 0.007883712649345398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0017124050064012408 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot begins a slow, revealing advance as a vibrant lime-green caterpillar, its segmented body textured with a soft, velvety texture, methodically nibbles at the edge of a broad, dew-covered leaf. Soft, golden morning sunlight slants across the frame, illuminating the scene and causing the crystal-clear, spherical dewdrops scattered across the leaf's surface to shimmer like tiny gemstones, refracting the light. As the camera inconspicuously approaches, the extremely shallow depth of field melts the background into a creamy, painterly bokeh of lush greens and earthy browns. The caterpillar's tiny, powerful jaws take center stage, rhythmically snipping at the delicate fragments of the leaf with deliberate, focused precision, its head moving in a steady, repetitive arc. The entire scene is imbued with an atmosphere of tranquility, a quiet moment of natural life observed with intimate and breathtaking clarity.", + "seed": 14695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_5844_seed_14695_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": -0.013222456339281052, + "vlm_logprob": -0.013222456339281052, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.013222456339281052, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004374696407467127 + ], + "opposite_logprob_tokens": [ + -5.7543745040893555 + ], + "logprob_reward": -0.004374696407467127, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004374696407467127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07751978188753128, + 0.1978447586297989, + 0.029450135305523872, + 0.059835437685251236, + 0.03353768214583397, + 0.019458279013633728, + 0.023056210950016975, + 0.048058874905109406, + 0.398627907037735, + 0.11261089891195297 + ], + "credit_map": [ + 0.007141967304050922, + 0.007694025989621878, + 0.008661016821861267, + 0.008771928027272224, + 0.013261028565466404, + 0.0112777603790164, + 0.01081598736345768, + 0.009896079078316689, + 0.017033228650689125, + 0.014209271408617496, + 0.03215824440121651, + 0.01500443834811449, + 0.022787705063819885, + 0.05241861566901207, + 0.02041763998568058, + 0.02381562627851963, + 0.002503267489373684, + 0.002959944773465395, + 0.0035340518224984407, + 0.004464471247047186, + 0.00351194036193192, + 0.0036338381469249725, + 0.003964719362556934, + 0.004877903498709202, + 0.009914975613355637, + 0.011948189698159695, + 0.012792333029210567, + 0.006831411272287369, + 0.004983129445463419, + 0.004284902475774288, + 0.004922030493617058, + 0.00415846798568964, + 0.007273409515619278, + 0.002787153236567974, + 0.013378463685512543, + 0.004349649418145418, + 0.0016959125641733408, + 0.0008823988027870655, + 0.0016033276915550232, + 0.001567373052239418, + 0.004193378146737814, + 0.002141844015568495, + 0.0032164030708372593, + 0.0030500106513500214, + 0.0017823345260694623, + 0.0017594947712495923, + 0.0014026861172169447, + 0.0019121289951726794, + 0.0048677511513233185, + 0.0017393432790413499, + 0.005933826789259911, + 0.00270171370357275, + 0.0017058684024959803, + 0.001606258563697338, + 0.001944253221154213, + 0.002557197352871299, + 0.006636901292949915, + 0.004309612791985273, + 0.00643523782491684, + 0.008581076748669147, + 0.004241136368364096, + 0.005722768604755402, + 0.004372593946754932, + 0.007759548258036375, + 0.09472127258777618, + 0.16168604791164398, + 0.07348675280809402, + 0.022921601310372353, + 0.017771676182746887, + 0.007430939935147762, + 0.009476232342422009, + 0.011133389547467232, + 0.008385302498936653, + 0.0231943316757679, + 0.027945278212428093, + 0.009437247179448605, + 0.008147040382027626, + 0.008485646918416023, + 0.01242398377507925, + 0.014592074789106846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.004374696407467127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05203006789088249, + 0.04744565114378929, + 0.017146820202469826, + 0.0736309066414833, + 0.012128404341638088, + 0.016809148713946342, + 0.023301994428038597, + 0.06107298284769058, + 0.6634517908096313, + 0.032982226461172104 + ], + "credit_map": [ + 0.004237966611981392, + 0.006754216272383928, + 0.004871398210525513, + 0.006304363254457712, + 0.010438368655741215, + 0.005645441822707653, + 0.008527176454663277, + 0.005251138936728239, + 0.003610732965171337, + 0.00315607781521976, + 0.009831961244344711, + 0.004893052391707897, + 0.00420777965337038, + 0.013748779892921448, + 0.0029149565380066633, + 0.005082309246063232, + 0.0018155454890802503, + 0.0017393982270732522, + 0.002500457689166069, + 0.002703561680391431, + 0.002555477898567915, + 0.0023873401805758476, + 0.0018042607698589563, + 0.0016407775692641735, + 0.010801860131323338, + 0.013211694546043873, + 0.01098009292036295, + 0.011193078942596912, + 0.00700587360188365, + 0.004939925391227007, + 0.008690002374351025, + 0.006808372680097818, + 0.0013806300703436136, + 0.0020425075199455023, + 0.0015409918269142509, + 0.0022971858270466328, + 0.0015712877502664924, + 0.001167764188721776, + 0.0011827033013105392, + 0.0009453329839743674, + 0.0024200736079365015, + 0.0016279572155326605, + 0.002820343477651477, + 0.0031365000177174807, + 0.002051976975053549, + 0.0011183293536305428, + 0.00199909252114594, + 0.0016348763601854444, + 0.00631092581897974, + 0.0017239365261048079, + 0.008044147863984108, + 0.0016570204170420766, + 0.001388257136568427, + 0.0013095324393361807, + 0.0014170575886964798, + 0.0014511169865727425, + 0.017812419682741165, + 0.0037937830202281475, + 0.017443358898162842, + 0.005396317690610886, + 0.004936254117637873, + 0.003718353807926178, + 0.0036100938450545073, + 0.004362399689853191, + 0.050201911479234695, + 0.24491311609745026, + 0.24605309963226318, + 0.047982025891542435, + 0.026302913203835487, + 0.01092128362506628, + 0.016465669497847557, + 0.020611803978681564, + 0.0032402914948761463, + 0.004198672249913216, + 0.004655241500586271, + 0.005621941760182381, + 0.003207592060789466, + 0.003603728022426367, + 0.0046308329328894615, + 0.0038239271380007267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.04865625500679016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641948267817497 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641948267817497, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641948267817497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10890531539916992, + 0.1416589319705963, + 0.030224548652768135, + 0.10608997195959091, + 0.057391662150621414, + 0.026753943413496017, + 0.01920390874147415, + 0.047127220779657364, + 0.35736072063446045, + 0.10528378188610077 + ], + "credit_map": [ + 0.009608948603272438, + 0.01176958903670311, + 0.012017828412353992, + 0.01279713585972786, + 0.01941297948360443, + 0.0144707802683115, + 0.01792301796376705, + 0.010905037634074688, + 0.012927203439176083, + 0.012964153662323952, + 0.01293216086924076, + 0.013463878072798252, + 0.02156534232199192, + 0.026531606912612915, + 0.02035384066402912, + 0.020920736715197563, + 0.002698274562135339, + 0.003105980809777975, + 0.005320905242115259, + 0.0036519761197268963, + 0.0034546377137303352, + 0.0039649843238294125, + 0.0031377794221043587, + 0.004890011623501778, + 0.014441302046179771, + 0.027241120114922523, + 0.015132769010961056, + 0.007952938787639141, + 0.007970539852976799, + 0.010874602012336254, + 0.008872250095009804, + 0.013604450039565563, + 0.012228124774992466, + 0.004296216182410717, + 0.020394224673509598, + 0.007993631064891815, + 0.0036540937144309282, + 0.0023660792503505945, + 0.003146311966702342, + 0.0033129847142845392, + 0.0030905345920473337, + 0.0021155395079404116, + 0.003556817304342985, + 0.004159620031714439, + 0.003389726858586073, + 0.0037176141049712896, + 0.0033473013900220394, + 0.0033767896238714457, + 0.0029130263719707727, + 0.0020654252730309963, + 0.0033253063447773457, + 0.002758336951956153, + 0.001804932253435254, + 0.0022897389717400074, + 0.0017556477105244994, + 0.0022914940491318703, + 0.008153445087373257, + 0.004978016018867493, + 0.007495434023439884, + 0.007962861098349094, + 0.0036514417733997107, + 0.0038046545814722776, + 0.003404871793463826, + 0.007676496170461178, + 0.07476357370615005, + 0.14539781212806702, + 0.07747052609920502, + 0.01867985725402832, + 0.016143258661031723, + 0.005828374531120062, + 0.008546542376279831, + 0.010530814528465271, + 0.007673972751945257, + 0.03137516975402832, + 0.03514464199542999, + 0.007306779269129038, + 0.0052686878480017185, + 0.004488008096814156, + 0.0063117011450231075, + 0.007714820560067892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.004641948267817497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626228500157595 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007626228500157595, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626228500157595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2631368935108185, + 0.1561405062675476, + 0.046867724508047104, + 0.12383891642093658, + 0.03340301290154457, + 0.016981063410639763, + 0.04425494000315666, + 0.10356380045413971, + 0.1430160403251648, + 0.0687970444560051 + ], + "credit_map": [ + 0.02545826882123947, + 0.04071049764752388, + 0.03145572915673256, + 0.024831285700201988, + 0.04374200850725174, + 0.03375430405139923, + 0.030015187337994576, + 0.033169619739055634, + 0.014929388649761677, + 0.019489914178848267, + 0.010006165131926537, + 0.015976062044501305, + 0.025824809446930885, + 0.018221737816929817, + 0.022414278239011765, + 0.02927815355360508, + 0.004588006529957056, + 0.004981183912605047, + 0.006257553584873676, + 0.007049685809761286, + 0.00702769635245204, + 0.005971807986497879, + 0.006708682514727116, + 0.004283108748495579, + 0.019599061459302902, + 0.022988462820649147, + 0.026465896517038345, + 0.01415456086397171, + 0.010830920189619064, + 0.007858077995479107, + 0.010264012031257153, + 0.01167792733758688, + 0.00652111554518342, + 0.002598552731797099, + 0.008716627955436707, + 0.00849133636802435, + 0.0017419029027223587, + 0.0013860521139577031, + 0.002232020255178213, + 0.0017154033994302154, + 0.002388258930295706, + 0.0017680261516943574, + 0.002499829977750778, + 0.0029533987399190664, + 0.0019346229964867234, + 0.001283245044760406, + 0.0018727394053712487, + 0.0022809410002082586, + 0.010708963498473167, + 0.004871781915426254, + 0.010356320068240166, + 0.004595427308231592, + 0.004309965763241053, + 0.00380504853092134, + 0.002694167895242572, + 0.0029132666531950235, + 0.026279335841536522, + 0.007374885957688093, + 0.026743296533823013, + 0.012911808677017689, + 0.008772372268140316, + 0.006062719039618969, + 0.007325039245188236, + 0.00809434149414301, + 0.036149777472019196, + 0.02131454087793827, + 0.055971909314394, + 0.009433282539248466, + 0.005116816610097885, + 0.004256837069988251, + 0.005091880448162556, + 0.005680998787283897, + 0.003987388219684362, + 0.019887087866663933, + 0.02169996313750744, + 0.005241993349045515, + 0.0041027795523405075, + 0.003928615245968103, + 0.006346576847136021, + 0.003602640237659216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.007626228500157595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008131535141728818 + ], + "opposite_logprob_tokens": [ + -7.125813007354736 + ], + "logprob_reward": -0.0008131535141728818, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008131535141728818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23557017743587494, + 0.048845451325178146, + 0.030897339805960655, + 0.06980795413255692, + 0.07510508596897125, + 0.02640271745622158, + 0.022653616964817047, + 0.05345557630062103, + 0.3968126177787781, + 0.040449414402246475 + ], + "credit_map": [ + 0.026712721213698387, + 0.025537380948662758, + 0.03158079460263252, + 0.022474749013781548, + 0.03515273705124855, + 0.03345243260264397, + 0.026105208322405815, + 0.0345541387796402, + 0.00423871586099267, + 0.004502329975366592, + 0.00838025938719511, + 0.008352138102054596, + 0.005409968085587025, + 0.007658494170755148, + 0.004913841839879751, + 0.005389705765992403, + 0.003532088827341795, + 0.003315525595098734, + 0.004461882170289755, + 0.005296580493450165, + 0.003012048779055476, + 0.00404938030987978, + 0.003817290998995304, + 0.003412543097510934, + 0.010105698369443417, + 0.008362536318600178, + 0.015943892300128937, + 0.008504790253937244, + 0.007205044850707054, + 0.005431418307125568, + 0.00709033478051424, + 0.00716423848643899, + 0.01766083389520645, + 0.0060795508325099945, + 0.03144648298621178, + 0.008276819251477718, + 0.0032743539195507765, + 0.0017553329234942794, + 0.0033005238510668278, + 0.003311184700578451, + 0.0033914647065103054, + 0.0017760838381946087, + 0.0038169503677636385, + 0.004629089031368494, + 0.00444504851475358, + 0.0025759211275726557, + 0.002645989879965782, + 0.0031221688259392977, + 0.003878629067912698, + 0.0025845186319202185, + 0.004516142886132002, + 0.0032601067796349525, + 0.0021697471383959055, + 0.0019984175451099873, + 0.002116614254191518, + 0.002129439963027835, + 0.010617496445775032, + 0.005720668938010931, + 0.010090271942317486, + 0.008541391231119633, + 0.00399185623973608, + 0.003901825984939933, + 0.0038563304115086794, + 0.0067357332445681095, + 0.04293930158019066, + 0.15774613618850708, + 0.13497070968151093, + 0.02148052677512169, + 0.01633327268064022, + 0.005763419438153505, + 0.007483469322323799, + 0.010095790959894657, + 0.004755858797580004, + 0.0058353738859295845, + 0.006213823799043894, + 0.006319797597825527, + 0.004935011267662048, + 0.002924368716776371, + 0.005043113138526678, + 0.004422069061547518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0008131535141728818 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot begins a slow, revealing advance as a vibrant lime-green caterpillar, its segmented body textured with a soft, velvety texture, methodically nibbles at the edge of a broad, dew-covered leaf. Soft, golden morning sunlight slants across the frame, illuminating the scene and causing the crystal-clear, spherical dewdrops scattered across the leaf's surface to shimmer like tiny gemstones, refracting the light. As the camera inconspicuously approaches, the extremely shallow depth of field melts the background into a creamy, painterly bokeh of lush greens and earthy browns. The caterpillar's tiny, powerful jaws take center stage, rhythmically snipping at the delicate fragments of the leaf with deliberate, focused precision, its head moving in a steady, repetitive arc. The entire scene is imbued with an atmosphere of tranquility, a quiet moment of natural life observed with intimate and breathtaking clarity.", + "seed": 14696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_2/index_5844_seed_14696_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": -0.008541926904581488, + "vlm_logprob": -0.008541926904581488, + "vlm_yes_no_margin": 5.6, + "vlm_token_credit": -0.008541926904581488, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002150010783225298 + ], + "opposite_logprob_tokens": [ + -6.627150058746338 + ], + "logprob_reward": -0.002150010783225298, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.002150010783225298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14375661313533783, + 0.12159387022256851, + 0.08725196868181229, + 0.023828942328691483, + 0.19637088477611542, + 0.09426269680261612, + 0.060991015285253525, + 0.07330215722322464, + 0.0861358568072319, + 0.11250599473714828 + ], + "credit_map": [ + 0.017811225727200508, + 0.017453700304031372, + 0.015979692339897156, + 0.01910717412829399, + 0.02577686868607998, + 0.0188345517963171, + 0.012192131020128727, + 0.016601258888840675, + 0.012756026349961758, + 0.01286371424794197, + 0.013945079408586025, + 0.017350170761346817, + 0.015255490317940712, + 0.018866052851080894, + 0.019734442234039307, + 0.010822891257703304, + 0.012157936580479145, + 0.018995314836502075, + 0.008797948248684406, + 0.004895062651485205, + 0.014247400686144829, + 0.015370357781648636, + 0.006528920494019985, + 0.006259032990783453, + 0.0028576087206602097, + 0.002240358619019389, + 0.0043326690793037415, + 0.00265585258603096, + 0.00259562861174345, + 0.003203233703970909, + 0.002825785893946886, + 0.0031178053468465805, + 0.03508437052369118, + 0.025970665737986565, + 0.028736380860209465, + 0.02633945643901825, + 0.02003934420645237, + 0.013463923707604408, + 0.018175465986132622, + 0.028561262413859367, + 0.014181065373122692, + 0.00942936073988676, + 0.012268285267055035, + 0.010845690034329891, + 0.012119830586016178, + 0.013969479128718376, + 0.012509996071457863, + 0.008938988670706749, + 0.009566733613610268, + 0.007774865720421076, + 0.008959601633250713, + 0.012452958151698112, + 0.007655501365661621, + 0.004945548716932535, + 0.004196532536298037, + 0.005439277738332748, + 0.008849520236253738, + 0.01221340149641037, + 0.010575586929917336, + 0.008917710743844509, + 0.007810921408236027, + 0.00877507496625185, + 0.007677565794438124, + 0.00848237331956625, + 0.013923336751759052, + 0.011872334405779839, + 0.008879809640347958, + 0.01229656022042036, + 0.013630274683237076, + 0.008580564521253109, + 0.008939050137996674, + 0.008013933897018433, + 0.016917221248149872, + 0.007975740358233452, + 0.014852430671453476, + 0.022248348221182823, + 0.014060724526643753, + 0.00945072527974844, + 0.00838133879005909, + 0.018619472160935402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.002150010783225298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807694256305695 + ], + "opposite_logprob_tokens": [ + -3.5298078060150146 + ], + "logprob_reward": -0.029807694256305695, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807694256305695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10452238470315933, + 0.10173293948173523, + 0.08009422570466995, + 0.044834572821855545, + 0.07250849902629852, + 0.16975244879722595, + 0.09970101714134216, + 0.1228264644742012, + 0.16254960000514984, + 0.04147784411907196 + ], + "credit_map": [ + 0.010257823392748833, + 0.012591056525707245, + 0.009857629425823689, + 0.009436183609068394, + 0.023710066452622414, + 0.016083326190710068, + 0.009169502183794975, + 0.013416795991361141, + 0.008740555495023727, + 0.012207554653286934, + 0.011433174833655357, + 0.014602260664105415, + 0.009673194959759712, + 0.021114934235811234, + 0.01111612282693386, + 0.012845142744481564, + 0.009535462595522404, + 0.013907769694924355, + 0.012063102796673775, + 0.00683022104203701, + 0.01190041285008192, + 0.010870306752622128, + 0.007339141797274351, + 0.0076478091068565845, + 0.0052915229462087154, + 0.0044917156919837, + 0.007707768119871616, + 0.004580322187393904, + 0.005613073706626892, + 0.005483080167323351, + 0.006348385475575924, + 0.0053187026642262936, + 0.010725945234298706, + 0.009081128984689713, + 0.009439906105399132, + 0.009149091318249702, + 0.009076209738850594, + 0.005629872903227806, + 0.00907671544700861, + 0.010329639539122581, + 0.021005932241678238, + 0.031002255156636238, + 0.01596061699092388, + 0.018390338867902756, + 0.024707738310098648, + 0.02266623079776764, + 0.0196668803691864, + 0.016352454200387, + 0.012491721659898758, + 0.011804020032286644, + 0.01165041420608759, + 0.01726214960217476, + 0.016012078151106834, + 0.011795197613537312, + 0.00967672374099493, + 0.009008709341287613, + 0.014216707088053226, + 0.02253837324678898, + 0.012989446520805359, + 0.018033040687441826, + 0.013511141762137413, + 0.01291958149522543, + 0.011335154063999653, + 0.01728302426636219, + 0.023584121838212013, + 0.018562912940979004, + 0.026341797783970833, + 0.035240620374679565, + 0.013357104733586311, + 0.013387532904744148, + 0.01300477422773838, + 0.019070738926529884, + 0.005387778393924236, + 0.004584226291626692, + 0.005180804058909416, + 0.004908675793558359, + 0.007772326003760099, + 0.00560217397287488, + 0.003665823256596923, + 0.004376035649329424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.029807694256305695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019449143437668681 + ], + "opposite_logprob_tokens": [ + -6.2519450187683105 + ], + "logprob_reward": -0.0019449143437668681, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019449143437668681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34477055072784424, + 0.18947814404964447, + 0.09871476143598557, + 0.02715134248137474, + 0.054513853043317795, + 0.035525936633348465, + 0.08241289854049683, + 0.04435470327734947, + 0.04223139211535454, + 0.08084634691476822 + ], + "credit_map": [ + 0.03641457110643387, + 0.04689091816544533, + 0.03428766131401062, + 0.028875142335891724, + 0.0749938115477562, + 0.042013801634311676, + 0.035096675157547, + 0.046198029071092606, + 0.023385588079690933, + 0.02238227240741253, + 0.02168460749089718, + 0.028588958084583282, + 0.025552287697792053, + 0.028917446732521057, + 0.020836548879742622, + 0.01813047006726265, + 0.015346962958574295, + 0.021946057677268982, + 0.01114238053560257, + 0.006401464343070984, + 0.016091937199234962, + 0.015130432322621346, + 0.006082763895392418, + 0.006572785321623087, + 0.003763826098293066, + 0.002495158463716507, + 0.004832379054278135, + 0.0033907024189829826, + 0.003072823164984584, + 0.0020926780998706818, + 0.003674819367006421, + 0.0038289621006697416, + 0.007606271654367447, + 0.0077070072293281555, + 0.007877885363996029, + 0.006422828882932663, + 0.007408328354358673, + 0.0060589550994336605, + 0.0056211440823972225, + 0.005811444018036127, + 0.005692785605788231, + 0.0035409685224294662, + 0.00395579356700182, + 0.005181917920708656, + 0.003868109080940485, + 0.005008707754313946, + 0.004182849545031786, + 0.004094810225069523, + 0.009707260876893997, + 0.008190842345356941, + 0.013908213935792446, + 0.012829911895096302, + 0.009864788502454758, + 0.007172492798417807, + 0.009124028496444225, + 0.011615381576120853, + 0.005502655170857906, + 0.006798956543207169, + 0.004762596450746059, + 0.004135513678193092, + 0.005299849901348352, + 0.007119637914001942, + 0.0037106992676854134, + 0.007024806458503008, + 0.006236260756850243, + 0.0029077136423438787, + 0.005902957636862993, + 0.006826253142207861, + 0.005495171528309584, + 0.004935362376272678, + 0.004255346022546291, + 0.005672339349985123, + 0.011828554794192314, + 0.006920569576323032, + 0.01147346943616867, + 0.014872653409838676, + 0.00989900529384613, + 0.005824717693030834, + 0.006246464792639017, + 0.0137809282168746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0019449143437668681 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076261102221906185 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.0076261102221906185, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076261102221906185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13963717222213745, + 0.23179839551448822, + 0.07187844812870026, + 0.05078970268368721, + 0.1646617203950882, + 0.08502496033906937, + 0.0872235968708992, + 0.07218747586011887, + 0.03965585678815842, + 0.057142652571201324 + ], + "credit_map": [ + 0.015621945261955261, + 0.018657265231013298, + 0.01661202870309353, + 0.016516495496034622, + 0.020276637747883797, + 0.0237360130995512, + 0.014987885020673275, + 0.013228902593255043, + 0.019045304507017136, + 0.02167319878935814, + 0.028314590454101562, + 0.03782761096954346, + 0.02189779095351696, + 0.029965726658701897, + 0.05211186036467552, + 0.020962320268154144, + 0.007237739395350218, + 0.009140724316239357, + 0.01046169362962246, + 0.008006446063518524, + 0.009588451124727726, + 0.010923134163022041, + 0.009807980619370937, + 0.006712276954203844, + 0.006285486277192831, + 0.0047309077344834805, + 0.006185177713632584, + 0.004012496210634708, + 0.005041309166699648, + 0.0052663059905171394, + 0.007919922471046448, + 0.01134809572249651, + 0.025443699210882187, + 0.021785471588373184, + 0.02392701804637909, + 0.020460840314626694, + 0.020111272111535072, + 0.01461258064955473, + 0.015637820586562157, + 0.022683035582304, + 0.011075128801167011, + 0.014815053902566433, + 0.00955459475517273, + 0.009342550300061703, + 0.012605980038642883, + 0.011212551034986973, + 0.010431498289108276, + 0.005987600889056921, + 0.011484937742352486, + 0.013110565021634102, + 0.013495020568370819, + 0.01768363267183304, + 0.00816067960113287, + 0.00823903363198042, + 0.007130589336156845, + 0.007919135503470898, + 0.00891595333814621, + 0.011057285591959953, + 0.006453845649957657, + 0.010872939601540565, + 0.008065604604780674, + 0.010323729366064072, + 0.005613589193671942, + 0.010884532704949379, + 0.006308263633400202, + 0.003578628646209836, + 0.00676740612834692, + 0.009020524099469185, + 0.004121387377381325, + 0.002995821414515376, + 0.00314240250736475, + 0.003721421118825674, + 0.008554176427423954, + 0.005221717990934849, + 0.009240941144526005, + 0.00752821983769536, + 0.008664830587804317, + 0.005621165502816439, + 0.006048315670341253, + 0.006263283081352711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0076261102221906185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant lime-green caterpillar with a segmented, velvety body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the caterpillar nibbling at the edge of a broad leaf covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dewdrops on the leaf appear crystal-clear, spherical, and shimmering like gemstones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot an extreme macro view that slowly advances towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background exhibit an extremely shallow depth of field with a creamy, painterly bokeh?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011809049174189568 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011809049174189568, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011809049174189568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1647089719772339, + 0.31472745537757874, + 0.042163509875535965, + 0.04874187335371971, + 0.03531843051314354, + 0.06714276969432831, + 0.09696240723133087, + 0.08457288891077042, + 0.0457858107984066, + 0.09987595677375793 + ], + "credit_map": [ + 0.0160723477602005, + 0.02244541421532631, + 0.015500848181545734, + 0.018150633201003075, + 0.02641063928604126, + 0.02434893324971199, + 0.018678350374102592, + 0.023101799190044403, + 0.033336661756038666, + 0.03619275987148285, + 0.048011407256126404, + 0.04418365657329559, + 0.03867103531956673, + 0.04163217172026634, + 0.03792214021086693, + 0.034777577966451645, + 0.004281403962522745, + 0.005188282113522291, + 0.00500243017449975, + 0.0028089454863220453, + 0.009241071529686451, + 0.008834470994770527, + 0.0033015822991728783, + 0.003505322616547346, + 0.005512690637260675, + 0.004651815630495548, + 0.005585497245192528, + 0.004760971292853355, + 0.004018353298306465, + 0.005251334980130196, + 0.007609166204929352, + 0.011352043598890305, + 0.0047510541044175625, + 0.0038830789271742105, + 0.004404947627335787, + 0.0041154674254357815, + 0.006384116597473621, + 0.0033029024489223957, + 0.0037238006480038166, + 0.004753058776259422, + 0.009965927340090275, + 0.007827871479094028, + 0.007656338158994913, + 0.009154530242085457, + 0.00838269479572773, + 0.008646578527987003, + 0.0094423433765769, + 0.006066478323191404, + 0.011790184304118156, + 0.010481220670044422, + 0.013900090008974075, + 0.01385832391679287, + 0.01589277572929859, + 0.013274762779474258, + 0.008119679056107998, + 0.009645367041230202, + 0.010575133375823498, + 0.014052684418857098, + 0.011787126772105694, + 0.011247503571212292, + 0.008176467381417751, + 0.008466335013508797, + 0.006648173090070486, + 0.013619450852274895, + 0.006956620141863823, + 0.0040845563635230064, + 0.008879702538251877, + 0.005426950287073851, + 0.006780438590794802, + 0.005028904415667057, + 0.004157588817179203, + 0.004471047315746546, + 0.0144555838778615, + 0.009838764555752277, + 0.013730536215007305, + 0.013527480885386467, + 0.01657288894057274, + 0.010994307696819305, + 0.008140241727232933, + 0.012616148218512535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0011809049174189568 + } + ], + "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/0000058/group_3/index_tensor([9291, 9291])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_tensor([9291, 9291])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7eef02ffb807e6d6482ea55cbf4f1205fec73d28 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_tensor([9291, 9291])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, shot from a high first-person perspective, focuses steadily on a pristine white rice cooker perched on a light-grained wood kitchen countertop. The cooker's sleek, pure white exterior has a porcelain-like sheen. From within, a fierce simmering swirl of steam erupts continuously from the gap in the steam valve at the top, filling the air and rising with its thick, scalding white steam. A beam of warm afternoon sunlight streams in through a nearby window, precisely piercing the swirling steam and creating a soft yet clear Tyndall-effect beam of light. Illuminated by this beam, countless tiny dust particles slowly float and swirl like golden elves, imbuing the air itself with a palpable texture. The entire scene is imbued with the warm light and delicate textures that are the hallmarks of Makoto Shinkai's aesthetic, creating an atmosphere of tranquility, healing, and a sense of life.", + "seed": 12718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_9291_seed_12718_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": -0.04542641472071409, + "vlm_logprob": -0.04542641472071409, + "vlm_yes_no_margin": 4.099999976158142, + "vlm_token_credit": -0.04542641472071409, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052209433168172836 + ], + "opposite_logprob_tokens": [ + -3.3022093772888184 + ], + "logprob_reward": -0.052209433168172836, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.052209433168172836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07303987443447113, + 0.04054097831249237, + 0.44652220606803894, + 0.06565726548433304, + 0.19522199034690857, + 0.06359750032424927, + 0.041319578886032104, + 0.037872422486543655, + 0.019641483202576637, + 0.01658673770725727 + ], + "credit_map": [ + 0.006583299953490496, + 0.008674265816807747, + 0.0045745414681732655, + 0.016106531023979187, + 0.007118743844330311, + 0.0043052551336586475, + 0.019636226817965508, + 0.006041009444743395, + 0.003092491300776601, + 0.006545358337461948, + 0.002794975182041526, + 0.0050984579138457775, + 0.008835590444505215, + 0.002434717956930399, + 0.009108717553317547, + 0.00263066659681499, + 0.012779871001839638, + 0.01621454954147339, + 0.00758989667519927, + 0.08068641275167465, + 0.02082214690744877, + 0.02960895374417305, + 0.1819167137145996, + 0.09690362960100174, + 0.003289934480562806, + 0.0079308757558465, + 0.00409302581101656, + 0.021374380216002464, + 0.0066242823377251625, + 0.0033457044046372175, + 0.013102654367685318, + 0.005896406248211861, + 0.01743866316974163, + 0.037392839789390564, + 0.014444521628320217, + 0.03709899261593819, + 0.033216990530490875, + 0.011848410591483116, + 0.030995119363069534, + 0.01278645545244217, + 0.006211783736944199, + 0.013146752491593361, + 0.004257360007613897, + 0.008391080424189568, + 0.012366149574518204, + 0.004640200641006231, + 0.009884767234325409, + 0.004699402954429388, + 0.00318694906309247, + 0.008830330334603786, + 0.00284428708255291, + 0.007465802598744631, + 0.006870267912745476, + 0.002891648793593049, + 0.004666419699788094, + 0.0045638736337423325, + 0.00817650742828846, + 0.0048531051725149155, + 0.002718847244977951, + 0.005299870856106281, + 0.005604104604572058, + 0.003813230199739337, + 0.003295241855084896, + 0.00411151722073555, + 0.003181080799549818, + 0.0025975226890295744, + 0.0023762185592204332, + 0.0024665608070790768, + 0.0029778967145830393, + 0.001422638539224863, + 0.002326489891856909, + 0.002293075667694211, + 0.0028311803471297026, + 0.0025123509112745523, + 0.001120468252338469, + 0.002618849277496338, + 0.00241740420460701, + 0.0013515467289835215, + 0.002014733152464032, + 0.0017202041344717145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.052209433168172836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003222037572413683 + ], + "opposite_logprob_tokens": [ + -5.7532219886779785 + ], + "logprob_reward": -0.003222037572413683, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003222037572413683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08444110304117203, + 0.11469155550003052, + 0.5035117268562317, + 0.06333625316619873, + 0.09810356795787811, + 0.028958948329091072, + 0.014420761726796627, + 0.03000473789870739, + 0.024611597880721092, + 0.03791975975036621 + ], + "credit_map": [ + 0.007606775499880314, + 0.012455610558390617, + 0.009359125047922134, + 0.016179325059056282, + 0.008143864572048187, + 0.006143513135612011, + 0.012669594027101994, + 0.011883292347192764, + 0.005868768319487572, + 0.006863231770694256, + 0.009600944817066193, + 0.01451017800718546, + 0.010067584924399853, + 0.008553721010684967, + 0.03381743282079697, + 0.025409692898392677, + 0.014055653475224972, + 0.016042010858654976, + 0.008193407207727432, + 0.096006378531456, + 0.020783094689249992, + 0.03254791349172592, + 0.20846201479434967, + 0.10742129385471344, + 0.0030709735583513975, + 0.008000973612070084, + 0.003538119839504361, + 0.022594988346099854, + 0.005453457590192556, + 0.003422131994739175, + 0.009969178587198257, + 0.007286427542567253, + 0.00772674148902297, + 0.021554801613092422, + 0.005576364230364561, + 0.01814875565469265, + 0.021840455010533333, + 0.006220802199095488, + 0.011536652222275734, + 0.005498991813510656, + 0.002122541656717658, + 0.004101476166397333, + 0.00211509526707232, + 0.0033559580333530903, + 0.004764222539961338, + 0.0025228674057871103, + 0.006418061442673206, + 0.003558725118637085, + 0.0015338138910010457, + 0.0016960448119789362, + 0.001049983548000455, + 0.0027079680003225803, + 0.0019212526967749, + 0.001953868195414543, + 0.0022806082852184772, + 0.001277221948839724, + 0.004372668452560902, + 0.005327271297574043, + 0.0036864092107862234, + 0.003941760398447514, + 0.0048890127800405025, + 0.0024031212087720633, + 0.0034060771577060223, + 0.001978416694328189, + 0.004614796955138445, + 0.0035470775328576565, + 0.002565034432336688, + 0.002971985610201955, + 0.003238195087760687, + 0.0023968082387000322, + 0.002997727133333683, + 0.002279972191900015, + 0.006825814489275217, + 0.00449600862339139, + 0.003212784184142947, + 0.005783218424767256, + 0.0040254974737763405, + 0.0033471877686679363, + 0.005756895057857037, + 0.0044723558239638805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.003222037572413683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024975319392979145 + ], + "opposite_logprob_tokens": [ + -6.002497673034668 + ], + "logprob_reward": -0.0024975319392979145, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024975319392979145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08988773822784424, + 0.04596010968089104, + 0.3231290876865387, + 0.0928577110171318, + 0.1234581395983696, + 0.19776754081249237, + 0.025179317221045494, + 0.034663159400224686, + 0.01985108107328415, + 0.04724614694714546 + ], + "credit_map": [ + 0.006691643502563238, + 0.006449732463806868, + 0.005066552199423313, + 0.018290532752871513, + 0.006965808104723692, + 0.003576466115191579, + 0.029564909636974335, + 0.01328209973871708, + 0.005250018555670977, + 0.0063451495952904224, + 0.0034058247692883015, + 0.006417261436581612, + 0.006820298731327057, + 0.0038703749887645245, + 0.009876132011413574, + 0.003975046798586845, + 0.00866471603512764, + 0.010733273811638355, + 0.0059209223836660385, + 0.063349649310112, + 0.013760917820036411, + 0.02028636820614338, + 0.13335712254047394, + 0.0670560896396637, + 0.007508967537432909, + 0.013483465649187565, + 0.007609914522618055, + 0.02815069444477558, + 0.009312763810157776, + 0.0050008418038487434, + 0.015357090160250664, + 0.006433970760554075, + 0.011156454682350159, + 0.021872609853744507, + 0.008064047433435917, + 0.027495406568050385, + 0.024167662486433983, + 0.007911390624940395, + 0.014779853634536266, + 0.008010711520910263, + 0.017955316230654716, + 0.04434604570269585, + 0.012183434329926968, + 0.027050357311964035, + 0.04122782498598099, + 0.011744457297027111, + 0.029538197442889214, + 0.013721909373998642, + 0.0016463554929941893, + 0.004127246793359518, + 0.0014790025306865573, + 0.00486706243827939, + 0.005361813586205244, + 0.0023155855014920235, + 0.0036376728676259518, + 0.0017445776611566544, + 0.005836916621774435, + 0.004493165295571089, + 0.0031005737837404013, + 0.004730322398245335, + 0.006228744518011808, + 0.0034364182502031326, + 0.0040657189674675465, + 0.0027712977025657892, + 0.004104772582650185, + 0.003016827395185828, + 0.001044682809151709, + 0.0026238029822707176, + 0.003030674532055855, + 0.0017845932161435485, + 0.0026523612905293703, + 0.0015933660324662924, + 0.009485996328294277, + 0.006679742131382227, + 0.0043211583979427814, + 0.006654617376625538, + 0.006950013339519501, + 0.0036226781085133553, + 0.005822993814945221, + 0.0037089460529386997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.0024975319392979145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026365933939814568 + ], + "opposite_logprob_tokens": [ + -3.6513659954071045 + ], + "logprob_reward": -0.026365933939814568, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026365933939814568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08309245854616165, + 0.20933282375335693, + 0.05660843476653099, + 0.06473831087350845, + 0.2354801893234253, + 0.15485043823719025, + 0.01948443241417408, + 0.03051111474633217, + 0.0399165153503418, + 0.10598519444465637 + ], + "credit_map": [ + 0.006404017563909292, + 0.008541399613022804, + 0.007378645706921816, + 0.019351908937096596, + 0.0071379318833351135, + 0.0038243879098445177, + 0.022134289145469666, + 0.008319876156747341, + 0.0152747742831707, + 0.022134151309728622, + 0.011184806935489178, + 0.023252181708812714, + 0.03922739252448082, + 0.010654378682374954, + 0.06498879939317703, + 0.022616354748606682, + 0.0038900054059922695, + 0.0033669713884592056, + 0.001824752427637577, + 0.009187654592096806, + 0.0027894629165530205, + 0.004098398610949516, + 0.02000889554619789, + 0.011442294344305992, + 0.006641617510467768, + 0.014285349287092686, + 0.005177205428481102, + 0.008808664046227932, + 0.010474348440766335, + 0.0047097159549593925, + 0.009660996496677399, + 0.004980415105819702, + 0.01986847072839737, + 0.04519056901335716, + 0.015654321759939194, + 0.048547036945819855, + 0.041900526732206345, + 0.01344054564833641, + 0.03612072765827179, + 0.01475799921900034, + 0.014042462222278118, + 0.03410664200782776, + 0.009108723141252995, + 0.02024811878800392, + 0.032998226583004, + 0.009242384694516659, + 0.02400236204266548, + 0.011101515032351017, + 0.0016238261014223099, + 0.0035503320395946503, + 0.001220409176312387, + 0.0031303835567086935, + 0.0028722663410007954, + 0.002321605570614338, + 0.0022747076582163572, + 0.0024909027852118015, + 0.004225300159305334, + 0.0037363055162131786, + 0.0025163451209664345, + 0.004472331143915653, + 0.0044138492085039616, + 0.004319122992455959, + 0.00450801569968462, + 0.002319843042641878, + 0.010164632461965084, + 0.006103864870965481, + 0.003205380169674754, + 0.005559017416089773, + 0.004204904194921255, + 0.002437171060591936, + 0.005652931053191423, + 0.002588613424450159, + 0.024627048522233963, + 0.014493852853775024, + 0.008212368004024029, + 0.012071507051587105, + 0.011651991866528988, + 0.010803553275763988, + 0.011822856962680817, + 0.012302025221288204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.026365933939814568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283713698387146 + ], + "opposite_logprob_tokens": [ + -2.0178370475769043 + ], + "logprob_reward": -0.14283713698387146, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283713698387146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05403725802898407, + 0.0640658438205719, + 0.6548349857330322, + 0.02389955148100853, + 0.057735003530979156, + 0.02578030712902546, + 0.009716437198221684, + 0.02117074467241764, + 0.02949810028076172, + 0.05926177650690079 + ], + "credit_map": [ + 0.005302571691572666, + 0.005305036436766386, + 0.004910981748253107, + 0.01193958055227995, + 0.004712085705250502, + 0.003307824023067951, + 0.013211198151111603, + 0.005347981117665768, + 0.004530269652605057, + 0.00907084159553051, + 0.002632566262036562, + 0.00728182028979063, + 0.01505120936781168, + 0.0029016241896897554, + 0.01838753931224346, + 0.004209974315017462, + 0.018032578751444817, + 0.020693399012088776, + 0.010120022110641003, + 0.125366672873497, + 0.027205267921090126, + 0.04190440848469734, + 0.27183523774147034, + 0.13967740535736084, + 0.0024757334031164646, + 0.004870008677244186, + 0.0018660342320799828, + 0.0040983655489981174, + 0.003406185656785965, + 0.0017481475370004773, + 0.0028079492039978504, + 0.002627127105370164, + 0.005736140534281731, + 0.01166259404271841, + 0.0045585366897284985, + 0.010551943443715572, + 0.009854214265942574, + 0.0037998072803020477, + 0.00787859596312046, + 0.003693174570798874, + 0.00225109257735312, + 0.0031669512391090393, + 0.0024744507391005754, + 0.0046205706894397736, + 0.004108881577849388, + 0.0018831200432032347, + 0.005122109316289425, + 0.0021531295496970415, + 0.0008463652920909226, + 0.0013746621552854776, + 0.0006794087821617723, + 0.0014541985001415014, + 0.0014692946569994092, + 0.001488033216446638, + 0.0014008019352331758, + 0.0010036717867478728, + 0.003722225083038211, + 0.002441791584715247, + 0.001985996961593628, + 0.002930448157712817, + 0.0023636033292859793, + 0.00327876303344965, + 0.002136384369805455, + 0.0023115333169698715, + 0.006076447200030088, + 0.004323143977671862, + 0.001694069942459464, + 0.003931919578462839, + 0.004169682040810585, + 0.002702508820220828, + 0.004563471302390099, + 0.00203685718588531, + 0.013830826617777348, + 0.008101851679384708, + 0.005056100897490978, + 0.006662358529865742, + 0.006271774414926767, + 0.005946820601820946, + 0.00757470540702343, + 0.005817336495965719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.14283713698387146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a high first-person perspective, focuses steadily on a pristine white rice cooker perched on a light-grained wood kitchen countertop. The cooker's sleek, pure white exterior has a porcelain-like sheen. From within, a fierce simmering swirl of steam erupts continuously from the gap in the steam valve at the top, filling the air and rising with its thick, scalding white steam. A beam of warm afternoon sunlight streams in through a nearby window, precisely piercing the swirling steam and creating a soft yet clear Tyndall-effect beam of light. Illuminated by this beam, countless tiny dust particles slowly float and swirl like golden elves, imbuing the air itself with a palpable texture. The entire scene is imbued with the warm light and delicate textures that are the hallmarks of Makoto Shinkai's aesthetic, creating an atmosphere of tranquility, healing, and a sense of life.", + "seed": 12719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_9291_seed_12719_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": -0.07418249743059277, + "vlm_logprob": -0.07418249743059277, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.07418249743059277, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11162307858467102 + ], + "opposite_logprob_tokens": [ + -2.6116230487823486 + ], + "logprob_reward": -0.11162307858467102, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.11162307858467102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11613751202821732, + 0.17881572246551514, + 0.058940064162015915, + 0.07147330045700073, + 0.09622011333703995, + 0.04875073581933975, + 0.07111801207065582, + 0.09746552258729935, + 0.15545819699764252, + 0.1056208536028862 + ], + "credit_map": [ + 0.008648163639008999, + 0.014322123490273952, + 0.013513322919607162, + 0.02480207569897175, + 0.019521653652191162, + 0.012441566213965416, + 0.014576025307178497, + 0.0083125876262784, + 0.00948412623256445, + 0.0388161763548851, + 0.009987443685531616, + 0.0185799989849329, + 0.062333762645721436, + 0.012986794114112854, + 0.018482938408851624, + 0.00814447458833456, + 0.00312321400269866, + 0.005864376202225685, + 0.007133474573493004, + 0.007298287935554981, + 0.008582497015595436, + 0.014909530989825726, + 0.006917121354490519, + 0.005111565347760916, + 0.004519984126091003, + 0.009713094681501389, + 0.006548707373440266, + 0.013680336065590382, + 0.009251182898879051, + 0.008572272025048733, + 0.012987739406526089, + 0.006199980620294809, + 0.005704245995730162, + 0.01532227173447609, + 0.008407976478338242, + 0.014833394438028336, + 0.022136421874165535, + 0.011819794774055481, + 0.01179832499474287, + 0.006197684444487095, + 0.003440180327743292, + 0.00672515295445919, + 0.006322541739791632, + 0.0059698657132685184, + 0.008989663794636726, + 0.0041434383019804955, + 0.010160917416214943, + 0.002998974407091737, + 0.006655631121248007, + 0.008685177192091942, + 0.005758187267929316, + 0.012622329406440258, + 0.012518544681370258, + 0.008191882632672787, + 0.01167056243866682, + 0.0050156996585428715, + 0.01053295936435461, + 0.011197683401405811, + 0.009992026723921299, + 0.011556051671504974, + 0.012730350717902184, + 0.02056681178510189, + 0.012167147360742092, + 0.008722485043108463, + 0.033548738807439804, + 0.04391293227672577, + 0.00956636667251587, + 0.01941526308655739, + 0.017146432772278786, + 0.008469474501907825, + 0.016955409198999405, + 0.006443578749895096, + 0.009685494005680084, + 0.013127828016877174, + 0.01015003677457571, + 0.01615186408162117, + 0.015630939975380898, + 0.013881528750061989, + 0.019969945773482323, + 0.007023217156529427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.11162307858467102 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003660289803519845 + ], + "opposite_logprob_tokens": [ + -5.628660202026367 + ], + "logprob_reward": -0.003660289803519845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003660289803519845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14712223410606384, + 0.07883195579051971, + 0.08190442621707916, + 0.0313015915453434, + 0.04466857388615608, + 0.027868183329701424, + 0.09918609261512756, + 0.3175625801086426, + 0.10041127353906631, + 0.07114306092262268 + ], + "credit_map": [ + 0.009687870740890503, + 0.016161000356078148, + 0.016810502856969833, + 0.018668469041585922, + 0.024517392739653587, + 0.02740694396197796, + 0.023957446217536926, + 0.00991261750459671, + 0.0039445566944777966, + 0.01441282220184803, + 0.0060672820545732975, + 0.010956049896776676, + 0.023101823404431343, + 0.007779465522617102, + 0.008860821835696697, + 0.003709129523485899, + 0.004347598180174828, + 0.008445505052804947, + 0.009344775229692459, + 0.011688155122101307, + 0.011594711802899837, + 0.018672319129109383, + 0.010995537973940372, + 0.0068158176727592945, + 0.0017333043506368995, + 0.004400084726512432, + 0.0029975883662700653, + 0.005769546143710613, + 0.004714952781796455, + 0.0033587473444640636, + 0.005632790271192789, + 0.0026945790741592646, + 0.0033765153493732214, + 0.00663151266053319, + 0.004286827053874731, + 0.005828797817230225, + 0.009295361116528511, + 0.0056360443122684956, + 0.006791342049837112, + 0.0028221753891557455, + 0.0019321945728734136, + 0.0037039744202047586, + 0.0024000857956707478, + 0.00350371515378356, + 0.005759920924901962, + 0.004653770476579666, + 0.004141137469559908, + 0.0017733876593410969, + 0.01767687499523163, + 0.014121239073574543, + 0.00932010356336832, + 0.012764646671712399, + 0.014295539818704128, + 0.01326583418995142, + 0.010482801124453545, + 0.0072590564377605915, + 0.03530198708176613, + 0.03639260306954384, + 0.035819295793771744, + 0.03295739367604256, + 0.03912706300616264, + 0.07321187853813171, + 0.033792514353990555, + 0.030959859490394592, + 0.025068141520023346, + 0.03329184278845787, + 0.005122359376400709, + 0.01017244253307581, + 0.010525302030146122, + 0.0038015558384358883, + 0.008909709751605988, + 0.0035199217963963747, + 0.009874198585748672, + 0.011319823563098907, + 0.0076153515838086605, + 0.009969105012714863, + 0.00950437132269144, + 0.009651425294578075, + 0.008173278532922268, + 0.005035507492721081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.003660289803519845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036235882434993982 + ], + "opposite_logprob_tokens": [ + -5.6286234855651855 + ], + "logprob_reward": -0.0036235882434993982, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036235882434993982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05518937110900879, + 0.09034580737352371, + 0.14051586389541626, + 0.03622187674045563, + 0.06308342516422272, + 0.03783649578690529, + 0.09309041500091553, + 0.17340153455734253, + 0.23912020027637482, + 0.07119492441415787 + ], + "credit_map": [ + 0.0029614202212542295, + 0.007861941121518612, + 0.0044813090935349464, + 0.008951588533818722, + 0.01261468231678009, + 0.008879514411091805, + 0.005799978040158749, + 0.0036389410961419344, + 0.0031694481149315834, + 0.02112940512597561, + 0.004415282513946295, + 0.007940077222883701, + 0.03493092954158783, + 0.006488249637186527, + 0.008818992413580418, + 0.003453429788351059, + 0.006842445116490126, + 0.013763511553406715, + 0.016569532454013824, + 0.01880059204995632, + 0.019580412656068802, + 0.03461770713329315, + 0.018850499764084816, + 0.011491168290376663, + 0.002500705188140273, + 0.00457061268389225, + 0.003459066851064563, + 0.006329512223601341, + 0.004388106055557728, + 0.005407183896750212, + 0.006463290192186832, + 0.0031034033745527267, + 0.003363469848409295, + 0.010658414103090763, + 0.005226427689194679, + 0.007987656630575657, + 0.016129326075315475, + 0.006729437038302422, + 0.009731685742735863, + 0.0032570071052759886, + 0.0029324672650545835, + 0.005729624070227146, + 0.0034984054509550333, + 0.003527616150677204, + 0.008619126863777637, + 0.006934324279427528, + 0.004462558776140213, + 0.00213237595744431, + 0.016567137092351913, + 0.014081542380154133, + 0.009395782835781574, + 0.01066464651376009, + 0.01783568784594536, + 0.006642990279942751, + 0.011957625858485699, + 0.005945018958300352, + 0.021381467580795288, + 0.020992442965507507, + 0.020206032320857048, + 0.018016895279288292, + 0.02086818590760231, + 0.03739435598254204, + 0.018811343237757683, + 0.015730831772089005, + 0.06152009218931198, + 0.0774402916431427, + 0.014214221388101578, + 0.02028973586857319, + 0.027776207774877548, + 0.00711038475856185, + 0.021632710471749306, + 0.009136579930782318, + 0.006814558058977127, + 0.008119785226881504, + 0.006306101568043232, + 0.010109642520546913, + 0.008776520378887653, + 0.014388433657586575, + 0.011197797954082489, + 0.005482093431055546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.0036235882434993982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362566277384758 + ], + "opposite_logprob_tokens": [ + -3.651362657546997 + ], + "logprob_reward": -0.026362566277384758, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362566277384758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12312640994787216, + 0.04770829901099205, + 0.10006029158830643, + 0.039943017065525055, + 0.061677802354097366, + 0.06614559888839722, + 0.38457024097442627, + 0.07895078510046005, + 0.03224056959152222, + 0.06557697802782059 + ], + "credit_map": [ + 0.009831132367253304, + 0.014087951742112637, + 0.014105081558227539, + 0.015439808368682861, + 0.022599829360842705, + 0.022495144978165627, + 0.015229718759655952, + 0.009337750263512135, + 0.0030711714643985033, + 0.00804214272648096, + 0.004730040207505226, + 0.004878724459558725, + 0.013051302172243595, + 0.006082764361053705, + 0.004717738833278418, + 0.003134415252134204, + 0.005200416315346956, + 0.010054484009742737, + 0.011569473892450333, + 0.014296043664216995, + 0.013887651264667511, + 0.023453786969184875, + 0.013490179553627968, + 0.008108260110020638, + 0.004622741602361202, + 0.003971575293689966, + 0.005206454079598188, + 0.005287374369800091, + 0.004295273683965206, + 0.006823592819273472, + 0.004434771370142698, + 0.005301234778016806, + 0.005381086841225624, + 0.00706895999610424, + 0.00787873100489378, + 0.009551829658448696, + 0.008817815221846104, + 0.010646152310073376, + 0.007165220100432634, + 0.00516801280900836, + 0.006249151658266783, + 0.007577940355986357, + 0.007478306535631418, + 0.007704760413616896, + 0.009534826502203941, + 0.01283517386764288, + 0.010383352637290955, + 0.004382089711725712, + 0.07256226986646652, + 0.05841168388724327, + 0.03902024030685425, + 0.04580229893326759, + 0.061309922486543655, + 0.02751140296459198, + 0.057219065725803375, + 0.022733382880687714, + 0.005573687609285116, + 0.008622129447758198, + 0.008257707580924034, + 0.011017028242349625, + 0.010407546535134315, + 0.015767807140946388, + 0.012077540159225464, + 0.00722734909504652, + 0.005567806772887707, + 0.007926096208393574, + 0.0025071552954614162, + 0.00442696874961257, + 0.003427585819736123, + 0.0032039270736277103, + 0.0034346820320934057, + 0.001746353693306446, + 0.0061010634526610374, + 0.007390007376670837, + 0.006203788798302412, + 0.010256771929562092, + 0.008563483133912086, + 0.011444992385804653, + 0.010978077538311481, + 0.00463879993185401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.026362566277384758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22564296424388885 + ], + "opposite_logprob_tokens": [ + -1.6006429195404053 + ], + "logprob_reward": -0.22564296424388885, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22564296424388885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12023386359214783, + 0.02465715818107128, + 0.05159101262688637, + 0.021642306819558144, + 0.041999299079179764, + 0.07096077501773834, + 0.12981411814689636, + 0.4216546416282654, + 0.07460565865039825, + 0.04284114018082619 + ], + "credit_map": [ + 0.009075777605175972, + 0.014736740849912167, + 0.013872824609279633, + 0.013392079621553421, + 0.023761801421642303, + 0.02103201486170292, + 0.015271277166903019, + 0.009091369807720184, + 0.002006250200793147, + 0.003509738016873598, + 0.0026904044207185507, + 0.002573456149548292, + 0.004818649962544441, + 0.004523394163697958, + 0.002697019837796688, + 0.0018382458947598934, + 0.0028256478253751993, + 0.005576668307185173, + 0.0054551903158426285, + 0.006607293616980314, + 0.008371365256607533, + 0.012405291199684143, + 0.006700867787003517, + 0.0036486901808530092, + 0.002479391172528267, + 0.0023432052694261074, + 0.0026855021715164185, + 0.0036155995912849903, + 0.0030496951658278704, + 0.0027819015085697174, + 0.0023598589468747377, + 0.0023271553218364716, + 0.002397669479250908, + 0.006968951318413019, + 0.003173328470438719, + 0.0044345855712890625, + 0.010826407931745052, + 0.007170016411691904, + 0.004354272969067097, + 0.0026740706525743008, + 0.005594582762569189, + 0.008462772704660892, + 0.007296910509467125, + 0.009011807851493359, + 0.011627297848463058, + 0.013704918324947357, + 0.011131774634122849, + 0.0041307127103209496, + 0.020285623148083687, + 0.01876186765730381, + 0.012239725328981876, + 0.016052642837166786, + 0.020891036838293076, + 0.011970218271017075, + 0.02175263501703739, + 0.00786037091165781, + 0.04321501404047012, + 0.04764175787568092, + 0.04882551357150078, + 0.041470542550086975, + 0.05291672423481941, + 0.09916304796934128, + 0.04681486263871193, + 0.04160722345113754, + 0.015386786311864853, + 0.02067188359797001, + 0.005068153142929077, + 0.008882610127329826, + 0.007885308004915714, + 0.0058768801391124725, + 0.007298982702195644, + 0.0035350583493709564, + 0.003727265167981386, + 0.005083410535007715, + 0.004855669569224119, + 0.00697373878210783, + 0.006714805494993925, + 0.005201971158385277, + 0.007700194139033556, + 0.0025840867310762405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.22564296424388885 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a high first-person perspective, focuses steadily on a pristine white rice cooker perched on a light-grained wood kitchen countertop. The cooker's sleek, pure white exterior has a porcelain-like sheen. From within, a fierce simmering swirl of steam erupts continuously from the gap in the steam valve at the top, filling the air and rising with its thick, scalding white steam. A beam of warm afternoon sunlight streams in through a nearby window, precisely piercing the swirling steam and creating a soft yet clear Tyndall-effect beam of light. Illuminated by this beam, countless tiny dust particles slowly float and swirl like golden elves, imbuing the air itself with a palpable texture. The entire scene is imbued with the warm light and delicate textures that are the hallmarks of Makoto Shinkai's aesthetic, creating an atmosphere of tranquility, healing, and a sense of life.", + "seed": 12720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_9291_seed_12720_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": -0.06567178363911808, + "vlm_logprob": -0.06567178363911808, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06567178363911808, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06615028530359268 + ], + "opposite_logprob_tokens": [ + -3.191150188446045 + ], + "logprob_reward": -0.06615028530359268, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06615028530359268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06398316472768784, + 0.0644759088754654, + 0.050233203917741776, + 0.03164653480052948, + 0.07772740721702576, + 0.26410484313964844, + 0.050239015370607376, + 0.07745920866727829, + 0.2986556887626648, + 0.02147502638399601 + ], + "credit_map": [ + 0.006052761804312468, + 0.009192722849547863, + 0.0075027961283922195, + 0.00901107583194971, + 0.013561458326876163, + 0.0050425720401108265, + 0.006514463108032942, + 0.007105317898094654, + 0.007026297505944967, + 0.008629544638097286, + 0.007113654166460037, + 0.008956804871559143, + 0.012288826517760754, + 0.005639888346195221, + 0.0077804806642234325, + 0.007040409836918116, + 0.005668927449733019, + 0.006156083662062883, + 0.007201046217232943, + 0.006812059786170721, + 0.009229430928826332, + 0.0038456732872873545, + 0.0049711172468960285, + 0.006348865572363138, + 0.003008500672876835, + 0.004431004635989666, + 0.003858119249343872, + 0.005244077183306217, + 0.004505869001150131, + 0.001964765368029475, + 0.0050390237011015415, + 0.003595175687223673, + 0.005535046570003033, + 0.014986992813646793, + 0.01016765832901001, + 0.007405675482004881, + 0.010217368602752686, + 0.007018066011369228, + 0.006430665496736765, + 0.015965929254889488, + 0.020237108692526817, + 0.05868222564458847, + 0.06973501294851303, + 0.02678702212870121, + 0.030566904693841934, + 0.012621109373867512, + 0.028120312839746475, + 0.017355136573314667, + 0.0032591326162219048, + 0.007849817164242268, + 0.005698368884623051, + 0.004248001147061586, + 0.015346769243478775, + 0.003403154667466879, + 0.005379918962717056, + 0.005053850822150707, + 0.016508126631379128, + 0.013512630015611649, + 0.008680879138410091, + 0.008736541494727135, + 0.009821115992963314, + 0.005305430386215448, + 0.008033007383346558, + 0.006861481815576553, + 0.025519294664263725, + 0.07777970284223557, + 0.02727268636226654, + 0.04387446865439415, + 0.04638464003801346, + 0.0194447860121727, + 0.03287706896662712, + 0.025503050535917282, + 0.002104640705510974, + 0.002139060525223613, + 0.0018509228248149157, + 0.0027715652249753475, + 0.004200174938887358, + 0.003227680455893278, + 0.0029945906717330217, + 0.0021863915026187897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.06615028530359268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008682594634592533 + ], + "opposite_logprob_tokens": [ + -4.758682727813721 + ], + "logprob_reward": -0.008682594634592533, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008682594634592533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1874738335609436, + 0.21264579892158508, + 0.05288791283965111, + 0.0336814820766449, + 0.07781576365232468, + 0.27330493927001953, + 0.015336062759160995, + 0.056079719215631485, + 0.06914181262254715, + 0.021632710471749306 + ], + "credit_map": [ + 0.021205848082900047, + 0.02412402629852295, + 0.020203227177262306, + 0.03149803355336189, + 0.027305247262120247, + 0.013230595737695694, + 0.02730962447822094, + 0.02259722165763378, + 0.0254986435174942, + 0.03156004101037979, + 0.023079199716448784, + 0.02473483607172966, + 0.0472823902964592, + 0.016960501670837402, + 0.02020099386572838, + 0.023329179733991623, + 0.005857220850884914, + 0.006349498406052589, + 0.007449094206094742, + 0.008505052886903286, + 0.0069388397969305515, + 0.0036467306781560183, + 0.007392105646431446, + 0.0067493668757379055, + 0.0027617181185632944, + 0.0034683318808674812, + 0.006034689024090767, + 0.0048800124786794186, + 0.003095219610258937, + 0.002951193368062377, + 0.004801386501640081, + 0.005688930861651897, + 0.007353696972131729, + 0.014829261228442192, + 0.012339620850980282, + 0.01087589468806982, + 0.008775275200605392, + 0.007260405924171209, + 0.009641538374125957, + 0.006740065291523933, + 0.02120986580848694, + 0.055902399122714996, + 0.06382361799478531, + 0.030977103859186172, + 0.031979624181985855, + 0.014067200012505054, + 0.03558498993515968, + 0.019760124385356903, + 0.0011057823430746794, + 0.0017511181067675352, + 0.0024713824968785048, + 0.002193073509261012, + 0.0027837289962917566, + 0.0009018410928547382, + 0.0023560195695608854, + 0.0017731155967339873, + 0.010678824037313461, + 0.011057138442993164, + 0.005764524918049574, + 0.007474798709154129, + 0.007658703718334436, + 0.0033391201868653297, + 0.005781735293567181, + 0.004324869252741337, + 0.004808391910046339, + 0.008078044280409813, + 0.012716865167021751, + 0.012792442925274372, + 0.013117327354848385, + 0.005648039747029543, + 0.006417509168386459, + 0.005563192069530487, + 0.0022607678547501564, + 0.0028396029956638813, + 0.0019541545771062374, + 0.0035374853760004044, + 0.0025521889328956604, + 0.0023312093690037727, + 0.004294791258871555, + 0.0018625087104737759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.008682594634592533 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003201718209311366 + ], + "opposite_logprob_tokens": [ + -5.753201484680176 + ], + "logprob_reward": -0.003201718209311366, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003201718209311366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14388929307460785, + 0.2100912630558014, + 0.09039650112390518, + 0.049583882093429565, + 0.10864940285682678, + 0.09234296530485153, + 0.040895599871873856, + 0.10346461832523346, + 0.09705183655023575, + 0.06363463401794434 + ], + "credit_map": [ + 0.012790456414222717, + 0.020145021378993988, + 0.018885282799601555, + 0.021733568981289864, + 0.02966223657131195, + 0.009834399446845055, + 0.01550841424614191, + 0.015329922549426556, + 0.02202077955007553, + 0.034098826348781586, + 0.021518884226679802, + 0.025843694806098938, + 0.04596381261944771, + 0.017680460587143898, + 0.01952398009598255, + 0.023440854623913765, + 0.010982428677380085, + 0.015010266564786434, + 0.008289463818073273, + 0.0104374336078763, + 0.019319847226142883, + 0.007067451719194651, + 0.007897043600678444, + 0.011392565444111824, + 0.006422366481274366, + 0.008098707534372807, + 0.0073488689959049225, + 0.0061973342671990395, + 0.009238856844604015, + 0.0030526011250913143, + 0.004394338000565767, + 0.004830812569707632, + 0.009346204809844494, + 0.017366262152791023, + 0.015227984637022018, + 0.017815496772527695, + 0.01623600721359253, + 0.008295029401779175, + 0.011285465210676193, + 0.013076966628432274, + 0.008538487367331982, + 0.014185571111738682, + 0.01272166334092617, + 0.016403598710894585, + 0.017573898658156395, + 0.005811804905533791, + 0.008783168159425259, + 0.008324787020683289, + 0.003374072490260005, + 0.006115954834967852, + 0.005006369668990374, + 0.005791593808680773, + 0.009228112176060677, + 0.0023128516040742397, + 0.005343554075807333, + 0.0037230930756777525, + 0.024792397394776344, + 0.015901079401373863, + 0.011224962770938873, + 0.013854481279850006, + 0.013166949152946472, + 0.007134981919080019, + 0.009127099066972733, + 0.008262679912149906, + 0.007916182279586792, + 0.022764410823583603, + 0.01268843375146389, + 0.013581070117652416, + 0.014090179465711117, + 0.006193618755787611, + 0.011462241411209106, + 0.008355701342225075, + 0.006577869411557913, + 0.007908007130026817, + 0.0060415384359657764, + 0.008242507465183735, + 0.008947287686169147, + 0.008658702485263348, + 0.011139791458845139, + 0.006118930876255035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.003201718209311366 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868168756365776 + ], + "opposite_logprob_tokens": [ + -3.0486817359924316 + ], + "logprob_reward": -0.04868168756365776, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868168756365776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13915975391864777, + 0.22598552703857422, + 0.07448841631412506, + 0.0485958531498909, + 0.19399289786815643, + 0.08562758564949036, + 0.015604459680616856, + 0.09374875575304031, + 0.09418054670095444, + 0.02861618436872959 + ], + "credit_map": [ + 0.013776028528809547, + 0.019880058243870735, + 0.0149573078379035, + 0.016455600038170815, + 0.03312276676297188, + 0.010909318923950195, + 0.014359126798808575, + 0.015699557960033417, + 0.024388160556554794, + 0.03498029336333275, + 0.027016567066311836, + 0.02255730703473091, + 0.05081777274608612, + 0.01950198970735073, + 0.02255723811686039, + 0.024166230112314224, + 0.010547562502324581, + 0.010261819697916508, + 0.007792032789438963, + 0.005987455137073994, + 0.015530289150774479, + 0.0071847341023385525, + 0.007132924627512693, + 0.010051604360342026, + 0.004246799740940332, + 0.0057970378547906876, + 0.008120543323457241, + 0.006208211183547974, + 0.006122368387877941, + 0.004438778385519981, + 0.004293865989893675, + 0.00936825480312109, + 0.01447835098952055, + 0.029114462435245514, + 0.03239237144589424, + 0.02892247959971428, + 0.02147012948989868, + 0.014813330955803394, + 0.022068019956350327, + 0.030733760446310043, + 0.0059631457552313805, + 0.01544024609029293, + 0.018599726259708405, + 0.00981250498443842, + 0.016154799610376358, + 0.004203449469059706, + 0.008566060103476048, + 0.006887659430503845, + 0.0010439425241202116, + 0.0027840312104672194, + 0.0021298283245414495, + 0.0014976469101384282, + 0.004665435757488012, + 0.001031654654070735, + 0.0011459309607744217, + 0.0013059889897704124, + 0.02286602556705475, + 0.01522226631641388, + 0.009843545034527779, + 0.011994054540991783, + 0.012264851480722427, + 0.006139841862022877, + 0.008108744397759438, + 0.00730943214148283, + 0.0057602860033512115, + 0.014880561269819736, + 0.016515633091330528, + 0.021212724968791008, + 0.012739078141748905, + 0.005654417909681797, + 0.010540952906012535, + 0.0068769003264606, + 0.0033998305443674326, + 0.0036073639057576656, + 0.0024628546088933945, + 0.0033497153781354427, + 0.004340717568993568, + 0.004093489609658718, + 0.0042940969578921795, + 0.003068121150135994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.04868168756365776 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20164263248443604 + ], + "opposite_logprob_tokens": [ + -1.701642632484436 + ], + "logprob_reward": -0.20164263248443604, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20164263248443604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1994154006242752, + 0.13730359077453613, + 0.1535198986530304, + 0.06406527757644653, + 0.1790553331375122, + 0.07024587690830231, + 0.034349944442510605, + 0.06978645920753479, + 0.060068316757678986, + 0.032189954072237015 + ], + "credit_map": [ + 0.01734219305217266, + 0.02603941783308983, + 0.019872799515724182, + 0.02682264894247055, + 0.04323647543787956, + 0.016727983951568604, + 0.028746509924530983, + 0.02062738500535488, + 0.014727028086781502, + 0.023494845256209373, + 0.013498054817318916, + 0.014422116801142693, + 0.03320058435201645, + 0.011293843388557434, + 0.01111329160630703, + 0.015553816221654415, + 0.022092269733548164, + 0.023525770753622055, + 0.01343513559550047, + 0.012990983203053474, + 0.035169877111911774, + 0.012877591885626316, + 0.012958208099007607, + 0.020470058545470238, + 0.005479150917381048, + 0.0083727752789855, + 0.011333493515849113, + 0.006737588904798031, + 0.008979303762316704, + 0.005495710764080286, + 0.005021354183554649, + 0.012645896524190903, + 0.01317811943590641, + 0.0274996068328619, + 0.027588794007897377, + 0.02468835562467575, + 0.022286107763648033, + 0.014997338876128197, + 0.021549761295318604, + 0.027267253026366234, + 0.005270503927022219, + 0.010277260094881058, + 0.01155872456729412, + 0.011107763275504112, + 0.013668252155184746, + 0.003557984484359622, + 0.008851602673530579, + 0.00595377990975976, + 0.002297493163496256, + 0.005819519981741905, + 0.00451888469979167, + 0.0035698169376701117, + 0.009501000866293907, + 0.0022401167079806328, + 0.003178994869813323, + 0.003224120009690523, + 0.015456529334187508, + 0.010060102678835392, + 0.006641994696110487, + 0.01116479467600584, + 0.009326444007456303, + 0.004784624557942152, + 0.006491237320005894, + 0.00586072588339448, + 0.0030567077919840813, + 0.01292002946138382, + 0.007239761296659708, + 0.008049612864851952, + 0.012982279062271118, + 0.004403567407280207, + 0.007084844633936882, + 0.004331517964601517, + 0.0031287174206227064, + 0.003698831656947732, + 0.0028751788195222616, + 0.003806225024163723, + 0.006692603696137667, + 0.0032964572310447693, + 0.005284120328724384, + 0.0034078180324286222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.20164263248443604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, shot from a high first-person perspective, focuses steadily on a pristine white rice cooker perched on a light-grained wood kitchen countertop. The cooker's sleek, pure white exterior has a porcelain-like sheen. From within, a fierce simmering swirl of steam erupts continuously from the gap in the steam valve at the top, filling the air and rising with its thick, scalding white steam. A beam of warm afternoon sunlight streams in through a nearby window, precisely piercing the swirling steam and creating a soft yet clear Tyndall-effect beam of light. Illuminated by this beam, countless tiny dust particles slowly float and swirl like golden elves, imbuing the air itself with a palpable texture. The entire scene is imbued with the warm light and delicate textures that are the hallmarks of Makoto Shinkai's aesthetic, creating an atmosphere of tranquility, healing, and a sense of life.", + "seed": 12721, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_3/index_9291_seed_12721_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": -0.1624766151420772, + "vlm_logprob": -0.1624766151420772, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1624766151420772, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09591592848300934 + ], + "opposite_logprob_tokens": [ + -2.7209160327911377 + ], + "logprob_reward": -0.09591592848300934, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09591592848300934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40002262592315674, + 0.046695075929164886, + 0.04347512871026993, + 0.05459924787282944, + 0.07619500905275345, + 0.05690200999379158, + 0.040213968604803085, + 0.11914120614528656, + 0.0676654502749443, + 0.09509027004241943 + ], + "credit_map": [ + 0.028984645381569862, + 0.1215396523475647, + 0.030193861573934555, + 0.018128246068954468, + 0.0766124352812767, + 0.08502110838890076, + 0.017426421865820885, + 0.02211628668010235, + 0.004479499999433756, + 0.00753019517287612, + 0.004244176670908928, + 0.005862166173756123, + 0.009441078640520573, + 0.0051469961181283, + 0.006662468891590834, + 0.0033284910023212433, + 0.008710392750799656, + 0.007229806389659643, + 0.004721686709672213, + 0.005175523925572634, + 0.005454034544527531, + 0.004484454169869423, + 0.004550902172923088, + 0.0031483250204473734, + 0.010256634093821049, + 0.009282475337386131, + 0.003855483140796423, + 0.007552743423730135, + 0.008711080998182297, + 0.0052415793761610985, + 0.004781704396009445, + 0.004917548503726721, + 0.006764495279639959, + 0.01197243295609951, + 0.006837885361164808, + 0.009594319388270378, + 0.01712278090417385, + 0.007637399714440107, + 0.007628810126334429, + 0.008636881597340107, + 0.005154407117515802, + 0.010134466923773289, + 0.005037470255047083, + 0.005739694926887751, + 0.014326208271086216, + 0.006917400751262903, + 0.004710846580564976, + 0.0048815165646374226, + 0.003711048047989607, + 0.004693503491580486, + 0.0037017876747995615, + 0.00539947347715497, + 0.01062537357211113, + 0.004311767872422934, + 0.004907666239887476, + 0.0028633468318730593, + 0.015445329248905182, + 0.021886341273784637, + 0.011900355108082294, + 0.008302101865410805, + 0.0198666974902153, + 0.019528117030858994, + 0.01015251874923706, + 0.012059736996889114, + 0.005156377330422401, + 0.015268917195498943, + 0.007270290050655603, + 0.005447286181151867, + 0.009192303754389286, + 0.01365816779434681, + 0.0064195566810667515, + 0.005252551753073931, + 0.008505621924996376, + 0.025915581732988358, + 0.011927064508199692, + 0.0050864689983427525, + 0.009525896050035954, + 0.026805568486452103, + 0.0028885239735245705, + 0.004435543902218342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.09591592848300934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011130386963486671 + ], + "opposite_logprob_tokens": [ + -4.511130332946777 + ], + "logprob_reward": -0.011130386963486671, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011130386963486671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2085607349872589, + 0.14080655574798584, + 0.059393543750047684, + 0.02419980615377426, + 0.10114455223083496, + 0.0687631219625473, + 0.04475102201104164, + 0.05613907054066658, + 0.013358042575418949, + 0.28288355469703674 + ], + "credit_map": [ + 0.023050596937537193, + 0.052121601998806, + 0.016434449702501297, + 0.013195248320698738, + 0.034749895334243774, + 0.035706717520952225, + 0.016276540234684944, + 0.01702568493783474, + 0.01335447933524847, + 0.04006674885749817, + 0.010917375795543194, + 0.008596505969762802, + 0.017179429531097412, + 0.03459590673446655, + 0.007946714758872986, + 0.008149391040205956, + 0.014959259890019894, + 0.013491233810782433, + 0.004978349432349205, + 0.004799021407961845, + 0.0068055917508900166, + 0.006320669315755367, + 0.003919453360140324, + 0.004119965247809887, + 0.004297682084143162, + 0.0034955537412315607, + 0.002471413230523467, + 0.002882124623283744, + 0.003394577419385314, + 0.0023082445841282606, + 0.0024279793724417686, + 0.002922231564298272, + 0.009954455308616161, + 0.01432389859110117, + 0.009407464414834976, + 0.013780084438621998, + 0.018105389550328255, + 0.012228444218635559, + 0.01132835540920496, + 0.012016457505524158, + 0.005170280113816261, + 0.01140299066901207, + 0.005625343881547451, + 0.0073218694888055325, + 0.01747981645166874, + 0.007991804741322994, + 0.007719630375504494, + 0.006051382981240749, + 0.006952192168682814, + 0.005166986957192421, + 0.004586188588291407, + 0.005748031660914421, + 0.009760517627000809, + 0.004753856919705868, + 0.004651684779673815, + 0.0031315635424107313, + 0.006870818790048361, + 0.007213955279439688, + 0.005345389246940613, + 0.0046706367284059525, + 0.014160825870931149, + 0.007175187114626169, + 0.005069652572274208, + 0.005632605869323015, + 0.0022251273039728403, + 0.00244506960734725, + 0.0016520109493285418, + 0.0015532518737018108, + 0.001770858303643763, + 0.0011688157683238387, + 0.0013564113760367036, + 0.0011864990228787065, + 0.02619517408311367, + 0.07866447418928146, + 0.03474104031920433, + 0.014489556662738323, + 0.026322441175580025, + 0.08098171651363373, + 0.008009761571884155, + 0.013479398563504219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.011130386963486671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263637285679579 + ], + "opposite_logprob_tokens": [ + -5.255263805389404 + ], + "logprob_reward": -0.005263637285679579, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263637285679579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41435205936431885, + 0.2424534410238266, + 0.04442533478140831, + 0.03378928080201149, + 0.06394754350185394, + 0.022734126076102257, + 0.04328330233693123, + 0.038653772324323654, + 0.011338267475366592, + 0.08502283692359924 + ], + "credit_map": [ + 0.026522014290094376, + 0.12893255054950714, + 0.0343044213950634, + 0.01520434021949768, + 0.07764894515275955, + 0.09651100635528564, + 0.012574413791298866, + 0.022654401138424873, + 0.018269799649715424, + 0.0800880417227745, + 0.01687016896903515, + 0.010292843915522099, + 0.032735779881477356, + 0.06658334285020828, + 0.005343604367226362, + 0.012269869446754456, + 0.007657404523342848, + 0.010138139128684998, + 0.0035241865552961826, + 0.003105244366452098, + 0.009310593828558922, + 0.004190069157630205, + 0.0029349965043365955, + 0.0035646974574774504, + 0.00439485814422369, + 0.0049421535804867744, + 0.0029534348286688328, + 0.005195240024477243, + 0.006472737528383732, + 0.0034590186551213264, + 0.003259910736232996, + 0.003111927304416895, + 0.008516259491443634, + 0.01135996263474226, + 0.00690305745229125, + 0.005981787107884884, + 0.010037283413112164, + 0.0073959860019385815, + 0.006813608575612307, + 0.006939600687474012, + 0.002700107404962182, + 0.00314099108800292, + 0.0026874197646975517, + 0.0029132117051631212, + 0.004173622466623783, + 0.002677180105820298, + 0.0028059929609298706, + 0.0016355989500880241, + 0.004220373462885618, + 0.003815825330093503, + 0.004844218026846647, + 0.005224277265369892, + 0.009599930606782436, + 0.005205260589718819, + 0.004558139480650425, + 0.005815278273075819, + 0.0056787109933793545, + 0.004251173231750727, + 0.004217821639031172, + 0.0044741216115653515, + 0.0062300218269228935, + 0.004537087865173817, + 0.006106728687882423, + 0.0031581043731421232, + 0.0006838055560365319, + 0.0017650933004915714, + 0.0013006948865950108, + 0.0011955122463405132, + 0.00220512761734426, + 0.002032908610999584, + 0.0013057203032076359, + 0.0008494044886901975, + 0.007520817685872316, + 0.023318246006965637, + 0.010192201472818851, + 0.00480729341506958, + 0.008259042166173458, + 0.024207336828112602, + 0.0025568963028490543, + 0.004161003511399031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005263637285679579 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22565752267837524 + ], + "opposite_logprob_tokens": [ + -1.6006574630737305 + ], + "logprob_reward": -0.22565752267837524, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22565752267837524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09952817112207413, + 0.14515139162540436, + 0.044917795807123184, + 0.04366057366132736, + 0.2537968158721924, + 0.042832955718040466, + 0.058744341135025024, + 0.10370940715074539, + 0.024408437311649323, + 0.18325012922286987 + ], + "credit_map": [ + 0.011422567069530487, + 0.02352701872587204, + 0.008147288113832474, + 0.008437375538051128, + 0.015544338151812553, + 0.016923611983656883, + 0.006729654502123594, + 0.008796312846243382, + 0.010615989565849304, + 0.04765379801392555, + 0.009916923940181732, + 0.006448662839829922, + 0.01907685399055481, + 0.04000646620988846, + 0.00391352828592062, + 0.007519185077399015, + 0.007394307758659124, + 0.007239583879709244, + 0.0050469352863729, + 0.0052510034292936325, + 0.0065222918055951595, + 0.005267176777124405, + 0.00421952735632658, + 0.0039769685827195644, + 0.0050477199256420135, + 0.005897480063140392, + 0.004492281936109066, + 0.005211479961872101, + 0.007402539253234863, + 0.004747482482343912, + 0.005034306086599827, + 0.005827281624078751, + 0.022413259372115135, + 0.03870531544089317, + 0.023984314873814583, + 0.028977323323488235, + 0.058629147708415985, + 0.026715965941548347, + 0.025833670049905777, + 0.028537815436720848, + 0.003567528212442994, + 0.007385016418993473, + 0.0037915068678557873, + 0.004452058579772711, + 0.011044930666685104, + 0.005274596158415079, + 0.0039021254051476717, + 0.0034151931758970022, + 0.005967495962977409, + 0.006288665812462568, + 0.0058804783038794994, + 0.006947850808501244, + 0.015821952372789383, + 0.006626928225159645, + 0.006268701050430536, + 0.0049422685988247395, + 0.008816621266305447, + 0.01593232899904251, + 0.0100667430087924, + 0.005955100990831852, + 0.030827734619379044, + 0.015431747771799564, + 0.007283501792699099, + 0.009395625442266464, + 0.0021551125682890415, + 0.0072543080896139145, + 0.0023586279712617397, + 0.001448649913072586, + 0.0028919707983732224, + 0.005658640060573816, + 0.0014450032031163573, + 0.001196125871501863, + 0.016458362340927124, + 0.050009310245513916, + 0.022067952901124954, + 0.010328564792871475, + 0.01587100327014923, + 0.05310414358973503, + 0.005469767842441797, + 0.009941025637090206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.22565752267837524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 from a high first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a pristine white rice cooker with a porcelain-like sheen visible on a light-grained wood countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does thick white steam continuously erupt from the steam valve gap and rise into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm sunlight visible piercing the steam to create a clear Tyndall-effect light beam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do tiny dust particles float and swirl within the light beam like golden elves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4744156002998352 + ], + "opposite_logprob_tokens": [ + -0.9744156002998352 + ], + "logprob_reward": -0.4744156002998352, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4744156002998352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19997337460517883, + 0.15311408042907715, + 0.050069406628608704, + 0.033293500542640686, + 0.14300493896007538, + 0.06103171780705452, + 0.042439162731170654, + 0.13703405857086182, + 0.03635269030928612, + 0.14368706941604614 + ], + "credit_map": [ + 0.020765183493494987, + 0.051179226487874985, + 0.01755615882575512, + 0.010453224182128906, + 0.032119229435920715, + 0.04149036854505539, + 0.011801955290138721, + 0.014608022756874561, + 0.011638662777841091, + 0.0498635359108448, + 0.009959620423614979, + 0.007779756560921669, + 0.020591244101524353, + 0.040971845388412476, + 0.004746813792735338, + 0.007562598679214716, + 0.004005381837487221, + 0.006131023168563843, + 0.005775048863142729, + 0.007828762754797935, + 0.00975044071674347, + 0.006076138932257891, + 0.004965308587998152, + 0.005537301767617464, + 0.005032440181821585, + 0.004319565836340189, + 0.0036363161634653807, + 0.0035060287918895483, + 0.005316713824868202, + 0.0036191216204315424, + 0.002937947865575552, + 0.0049253664910793304, + 0.014649040065705776, + 0.021856680512428284, + 0.012806286104023457, + 0.015459971502423286, + 0.034878164529800415, + 0.013189299963414669, + 0.015539199113845825, + 0.01462629996240139, + 0.00844984408468008, + 0.010966217145323753, + 0.005094991531223059, + 0.006859570275992155, + 0.012557785958051682, + 0.007086740806698799, + 0.005619782488793135, + 0.004396785516291857, + 0.004678127821534872, + 0.004645428620278835, + 0.005601458251476288, + 0.004079620353877544, + 0.007871047593653202, + 0.005524666979908943, + 0.0034265276044607162, + 0.006612288765609264, + 0.02260216884315014, + 0.023957978934049606, + 0.013119260780513287, + 0.010685443878173828, + 0.023420674726366997, + 0.018073873594403267, + 0.01127188466489315, + 0.01390277873724699, + 0.0038553569465875626, + 0.006665388587862253, + 0.0034243923146277666, + 0.003751137061044574, + 0.006534092128276825, + 0.00377775845117867, + 0.0059183211997151375, + 0.00242624175734818, + 0.011906088329851627, + 0.03793865442276001, + 0.016701018437743187, + 0.008931547403335571, + 0.012962565757334232, + 0.040853291749954224, + 0.005952578969299793, + 0.008441321551799774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.4744156002998352 + } + ], + "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/0000058/group_4/index_tensor([592, 592])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_4/index_tensor([592, 592])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0d7003bfb3a82ab1b991c615ecd5cadfdcb2179d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_4/index_tensor([592, 592])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the style of ultra-high-speed photography, a slightly high-angle medium shot unfolds across a block of raw maple, its pale, grainy surface still showing the rough edges of its original form, firmly mounted on a heavy-duty metal lathe in a dimly lit workshop. The block begins to spin at variable speeds, accelerating into a hypnotic blur under a single, intense overhead task light, casting dramatic, deep shadows and illuminating fine sawdust suspended in mid-air. A sharp steel chisel with a dark, polished wooden handle is steadily and deliberately guided toward the spinning surface. As its gleaming edge makes contact with the wood, the entire action transforms into breathtaking slow motion. Tiny chips erupt from the point of impact, each gossamer-thin strip of wood catching the bright light and spiraling away in elegant, suspended arcs against the dark, out-of-focus background. Meanwhile, the camera makes steady, deliberate pushes, ever closer to the moment of creation, zooming in on the violent yet graceful interaction between tool and material, revealing intricate, silky-smooth organic patterns emerging from the raw wood, creating a captivating and focused atmosphere of pure craftsmanship.", + "seed": 29863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_4/index_592_seed_29863_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": -0.025720637291669846, + "vlm_logprob": -0.025720637291669846, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.025720637291669846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.031934235244989395 + ], + "opposite_logprob_tokens": [ + -3.7819342613220215 + ], + "logprob_reward": -0.031934235244989395, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.031934235244989395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12969927489757538, + 0.05992182716727257, + 0.04650099575519562, + 0.030631138011813164, + 0.08097998797893524, + 0.016196489334106445, + 0.12468022108078003, + 0.18198010325431824, + 0.03755619749426842, + 0.29185375571250916 + ], + "credit_map": [ + 0.029811624437570572, + 0.009733003564178944, + 0.011079780757427216, + 0.0384974479675293, + 0.005852852948009968, + 0.014712659642100334, + 0.009320295415818691, + 0.010691601783037186, + 0.007020534016191959, + 0.004519276320934296, + 0.005674717947840691, + 0.011771860532462597, + 0.0028724721632897854, + 0.005655168555676937, + 0.009533876553177834, + 0.012873919680714607, + 0.0056225876323878765, + 0.009760802611708641, + 0.004395731259137392, + 0.005851970985531807, + 0.004685511346906424, + 0.003865607315674424, + 0.004118206910789013, + 0.008200574666261673, + 0.004287123214453459, + 0.0068718018010258675, + 0.0038904433604329824, + 0.0046419138088822365, + 0.001879494753666222, + 0.002397944452241063, + 0.0022553694434463978, + 0.00440704682841897, + 0.007843634113669395, + 0.0068145073018968105, + 0.02862384542822838, + 0.0081707201898098, + 0.0071335285902023315, + 0.01234250608831644, + 0.005064996890723705, + 0.004986249376088381, + 0.0021246641408652067, + 0.0015390543267130852, + 0.003588833147659898, + 0.0018559310119599104, + 0.0010430313413962722, + 0.0020231851376593113, + 0.0015073419781401753, + 0.0025144468527287245, + 0.023968704044818878, + 0.04337555542588234, + 0.009659020230174065, + 0.013236870057880878, + 0.012399046681821346, + 0.008275147527456284, + 0.007405199110507965, + 0.006360682658851147, + 0.017031462863087654, + 0.023352118209004402, + 0.03149600327014923, + 0.018109817057847977, + 0.017969220876693726, + 0.017912564799189568, + 0.03401046618819237, + 0.022098448127508163, + 0.0034834162797778845, + 0.0047769793309271336, + 0.005023485980927944, + 0.004441793542355299, + 0.003908699844032526, + 0.0036502026487141848, + 0.006985194981098175, + 0.005286426283419132, + 0.022746125236153603, + 0.02764957956969738, + 0.025109611451625824, + 0.028614507988095284, + 0.042591243982315063, + 0.030158111825585365, + 0.034370359033346176, + 0.08061420917510986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.031934235244989395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.026391245424747467 + ], + "opposite_logprob_tokens": [ + -3.6513912677764893 + ], + "logprob_reward": -0.026391245424747467, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026391245424747467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14717715978622437, + 0.021609360352158546, + 0.022923283278942108, + 0.09211046993732452, + 0.020477328449487686, + 0.013359898701310158, + 0.21042519807815552, + 0.2587744891643524, + 0.10501212626695633, + 0.10813065618276596 + ], + "credit_map": [ + 0.03103010728955269, + 0.010360910557210445, + 0.013481130823493004, + 0.04241371154785156, + 0.00687421252951026, + 0.01689869724214077, + 0.0145279411226511, + 0.011590452864766121, + 0.003268502652645111, + 0.0022014707792550325, + 0.003206344321370125, + 0.004241704475134611, + 0.0011027598520740867, + 0.0021677834447473288, + 0.0023201806470751762, + 0.0031006138306111097, + 0.002865749876946211, + 0.00529579259455204, + 0.0025247959420084953, + 0.003631448606029153, + 0.0020713794510811567, + 0.0017688106745481491, + 0.0024286326952278614, + 0.0023366722743958235, + 0.009451272897422314, + 0.020070163533091545, + 0.00851000938564539, + 0.01549643836915493, + 0.006589040160179138, + 0.007529801223427057, + 0.005794953554868698, + 0.01866879314184189, + 0.0024450772907584906, + 0.002966571133583784, + 0.0031175038311630487, + 0.003431993303820491, + 0.0018929166253656149, + 0.0013837838778272271, + 0.0026628903578966856, + 0.0025765926111489534, + 0.0015501666348427534, + 0.0020122136920690536, + 0.001910549821332097, + 0.0017441443633288145, + 0.0009077006252482533, + 0.0011944801080971956, + 0.0017662300961092114, + 0.0022744128946214914, + 0.04128340631723404, + 0.06608469784259796, + 0.02008603885769844, + 0.021817883476614952, + 0.019605346024036407, + 0.015797585248947144, + 0.013522094115614891, + 0.012228148058056831, + 0.0207529217004776, + 0.04143268242478371, + 0.06396321207284927, + 0.028372660279273987, + 0.01737985759973526, + 0.030618984252214432, + 0.03169146925210953, + 0.024562714621424675, + 0.008886764757335186, + 0.01584762893617153, + 0.017043370753526688, + 0.011972829699516296, + 0.009615747258067131, + 0.011448495090007782, + 0.018882356584072113, + 0.011314925737679005, + 0.009070833213627338, + 0.014310917817056179, + 0.00891790259629488, + 0.010563354939222336, + 0.016067899763584137, + 0.010585254058241844, + 0.015421045944094658, + 0.023193446919322014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.026391245424747467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.029815100133419037 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.029815100133419037, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815100133419037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10574087500572205, + 0.07189733535051346, + 0.02820134535431862, + 0.0146096870303154, + 0.01647711545228958, + 0.027106449007987976, + 0.16166146099567413, + 0.33729737997055054, + 0.029905665665864944, + 0.2071027308702469 + ], + "credit_map": [ + 0.022426601499319077, + 0.010237716138362885, + 0.009573436342179775, + 0.028710661455988884, + 0.0045063067227602005, + 0.010385621339082718, + 0.007175966165959835, + 0.012724553234875202, + 0.008123497478663921, + 0.0060123782604932785, + 0.007571553345769644, + 0.013537047430872917, + 0.004176679532974958, + 0.005647029262036085, + 0.01256829034537077, + 0.01426085177809, + 0.0039003475103527308, + 0.005125243682414293, + 0.004564655479043722, + 0.0036988051142543554, + 0.002408795291557908, + 0.0021238154731690884, + 0.002693132497370243, + 0.0036865496076643467, + 0.002195522887632251, + 0.0013772665988653898, + 0.0021082288585603237, + 0.0028107683174312115, + 0.0013240150874480605, + 0.0013087419793009758, + 0.0016398142324760556, + 0.0018453276716172695, + 0.0013859565369784832, + 0.0018860090058296919, + 0.002292899414896965, + 0.002441232558339834, + 0.001361209200695157, + 0.002010902389883995, + 0.0017324246000498533, + 0.0033664796501398087, + 0.0034886193461716175, + 0.0040137795731425285, + 0.004448604304343462, + 0.0036085823085159063, + 0.002588744508102536, + 0.0025746680330485106, + 0.0029267501085996628, + 0.003456700826063752, + 0.03847574442625046, + 0.02015567012131214, + 0.040190257132053375, + 0.01510942354798317, + 0.00829092413187027, + 0.017947115004062653, + 0.012558205053210258, + 0.008934114128351212, + 0.026187758892774582, + 0.03892207518219948, + 0.07335975766181946, + 0.03432532399892807, + 0.03087410144507885, + 0.039194654673337936, + 0.05844671279191971, + 0.03598695248365402, + 0.002917943987995386, + 0.003369250101968646, + 0.004073786083608866, + 0.004913690034300089, + 0.0022985960822552443, + 0.003625360084697604, + 0.004681013990193605, + 0.0040260241366922855, + 0.01922513358294964, + 0.015609567984938622, + 0.019830064848065376, + 0.018574340268969536, + 0.027584897354245186, + 0.023559652268886566, + 0.017282770946621895, + 0.06543629616498947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.029815100133419037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.033710431307554245 + ], + "opposite_logprob_tokens": [ + -3.408710479736328 + ], + "logprob_reward": -0.033710431307554245, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710431307554245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0798572301864624, + 0.04962930083274841, + 0.03217470645904541, + 0.044374242424964905, + 0.15631629526615143, + 0.017258068546652794, + 0.2792242169380188, + 0.09362655133008957, + 0.1806047111749649, + 0.06693463772535324 + ], + "credit_map": [ + 0.01748078130185604, + 0.006942342966794968, + 0.00796555820852518, + 0.021780885756015778, + 0.0036410533357411623, + 0.00831890944391489, + 0.006255839020013809, + 0.007471858989447355, + 0.005934244953095913, + 0.00636766292154789, + 0.007496639154851437, + 0.00838479958474636, + 0.0036670586559921503, + 0.00556586030870676, + 0.006170438602566719, + 0.006042599678039551, + 0.0049634166061878204, + 0.0038894671015441418, + 0.00571397552266717, + 0.005022952798753977, + 0.00138898438308388, + 0.0027072199154645205, + 0.002299889223650098, + 0.0061888019554317, + 0.00394531199708581, + 0.011130636557936668, + 0.003659440902993083, + 0.009005757980048656, + 0.0030215575825423002, + 0.0028884010389447212, + 0.003026719903573394, + 0.007696414832025766, + 0.014270882122218609, + 0.01393178105354309, + 0.054111652076244354, + 0.015308802016079426, + 0.016368918120861053, + 0.023749101907014847, + 0.009866361506283283, + 0.008708815090358257, + 0.0025125471875071526, + 0.0032138049136847258, + 0.002798329805955291, + 0.001700510736554861, + 0.0016316062537953258, + 0.0015843487344682217, + 0.0019671558402478695, + 0.0018497641431167722, + 0.03544777259230614, + 0.10859696567058563, + 0.03989594057202339, + 0.02571907453238964, + 0.029821345582604408, + 0.01826781965792179, + 0.01193235907703638, + 0.009542915970087051, + 0.007194966077804565, + 0.023098988458514214, + 0.026702960953116417, + 0.006952288094907999, + 0.005134131759405136, + 0.011798987165093422, + 0.008508776314556599, + 0.0042354511097073555, + 0.016136858612298965, + 0.023860551416873932, + 0.02628493309020996, + 0.02018832042813301, + 0.01990014687180519, + 0.015791984274983406, + 0.0323399119079113, + 0.02610201947391033, + 0.014828172512352467, + 0.008195359259843826, + 0.008514614775776863, + 0.00546792009845376, + 0.01351537462323904, + 0.0063437423668801785, + 0.00561089301481843, + 0.004458555486053228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.033710431307554245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.006752174347639084 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006752174347639084, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006752174347639084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11842331290245056, + 0.050138652324676514, + 0.02638373151421547, + 0.016886314377188683, + 0.03187277540564537, + 0.030024947598576546, + 0.19507324695587158, + 0.4223482310771942, + 0.04654823616147041, + 0.062300533056259155 + ], + "credit_map": [ + 0.020032983273267746, + 0.013174008578062057, + 0.019884321838617325, + 0.019484050571918488, + 0.006130725611001253, + 0.011845888569951057, + 0.011996560730040073, + 0.015874773263931274, + 0.005527862347662449, + 0.006127507891505957, + 0.004802178591489792, + 0.011612273752689362, + 0.0029829475097358227, + 0.004816401284188032, + 0.007235108409076929, + 0.007034376263618469, + 0.0036530678626149893, + 0.004521638620644808, + 0.003949605859816074, + 0.004025642294436693, + 0.0016651488840579987, + 0.002666739746928215, + 0.0024878783151507378, + 0.0034140089992433786, + 0.002076540607959032, + 0.003848511725664139, + 0.0023499634116888046, + 0.0033780094236135483, + 0.0015632829163223505, + 0.0011876163771376014, + 0.0012250945437699556, + 0.0012572945561259985, + 0.0032945366110652685, + 0.0038617397658526897, + 0.007155014667659998, + 0.00445490563288331, + 0.002339075319468975, + 0.0050123692490160465, + 0.0021012721117585897, + 0.0036538601852953434, + 0.0035816198214888573, + 0.004864939488470554, + 0.004174581728875637, + 0.0036157697904855013, + 0.0025047811213880777, + 0.002613859483972192, + 0.004056268837302923, + 0.004613124765455723, + 0.023040708154439926, + 0.07251803576946259, + 0.029534004628658295, + 0.01831728033721447, + 0.020452097058296204, + 0.013371296226978302, + 0.009294937364757061, + 0.00854488741606474, + 0.03537203371524811, + 0.04499669000506401, + 0.08466944098472595, + 0.04445134848356247, + 0.04056812822818756, + 0.044815849512815475, + 0.07455294579267502, + 0.05292178690433502, + 0.005680152680724859, + 0.005829459056258202, + 0.007475598249584436, + 0.006533745210617781, + 0.003789144102483988, + 0.005346547346562147, + 0.007040026132017374, + 0.0048535591922700405, + 0.009851614944636822, + 0.007005318533629179, + 0.006619233172386885, + 0.006221663672477007, + 0.0102273840457201, + 0.006921612191945314, + 0.006777992006391287, + 0.008675714023411274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.006752174347639084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of ultra-high-speed photography, a slightly high-angle medium shot unfolds across a block of raw maple, its pale, grainy surface still showing the rough edges of its original form, firmly mounted on a heavy-duty metal lathe in a dimly lit workshop. The block begins to spin at variable speeds, accelerating into a hypnotic blur under a single, intense overhead task light, casting dramatic, deep shadows and illuminating fine sawdust suspended in mid-air. A sharp steel chisel with a dark, polished wooden handle is steadily and deliberately guided toward the spinning surface. As its gleaming edge makes contact with the wood, the entire action transforms into breathtaking slow motion. Tiny chips erupt from the point of impact, each gossamer-thin strip of wood catching the bright light and spiraling away in elegant, suspended arcs against the dark, out-of-focus background. Meanwhile, the camera makes steady, deliberate pushes, ever closer to the moment of creation, zooming in on the violent yet graceful interaction between tool and material, revealing intricate, silky-smooth organic patterns emerging from the raw wood, creating a captivating and focused atmosphere of pure craftsmanship.", + "seed": 29864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_4/index_592_seed_29864_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": -0.09159697573632002, + "vlm_logprob": -0.09159697573632002, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.09159697573632002, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.21100515127182007 + ], + "opposite_logprob_tokens": [ + -1.8360052108764648 + ], + "logprob_reward": -0.21100515127182007, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21100515127182007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18764370679855347, + 0.10050173103809357, + 0.0959191843867302, + 0.08784575015306473, + 0.03404134884476662, + 0.09972484409809113, + 0.20556434988975525, + 0.0807865709066391, + 0.03730284795165062, + 0.0706697404384613 + ], + "credit_map": [ + 0.030049964785575867, + 0.012270880863070488, + 0.02590559422969818, + 0.024259429425001144, + 0.011798829771578312, + 0.029115594923496246, + 0.019763583317399025, + 0.03447983041405678, + 0.011169858276844025, + 0.008174379356205463, + 0.016132013872265816, + 0.01906602829694748, + 0.004198332782834768, + 0.0197190772742033, + 0.01276294607669115, + 0.009279097430408001, + 0.005994847044348717, + 0.004709823057055473, + 0.031018422916531563, + 0.024169795215129852, + 0.003226576605811715, + 0.009915650822222233, + 0.005264208186417818, + 0.011619864031672478, + 0.008791541680693626, + 0.005256710108369589, + 0.016794605180621147, + 0.023276271298527718, + 0.0038968471344560385, + 0.016742432489991188, + 0.008492724038660526, + 0.004594618920236826, + 0.003433435456827283, + 0.002447612350806594, + 0.004844446666538715, + 0.004207559861242771, + 0.002146510174497962, + 0.010948043316602707, + 0.0025901617482304573, + 0.0034235804341733456, + 0.011597849428653717, + 0.006535438355058432, + 0.01921561174094677, + 0.01555749960243702, + 0.007520107086747885, + 0.01478889025747776, + 0.011786500923335552, + 0.012722943909466267, + 0.017556436359882355, + 0.013112247921526432, + 0.045102208852767944, + 0.0310823954641819, + 0.013049241155385971, + 0.043994806706905365, + 0.02062968909740448, + 0.021037336438894272, + 0.008489592000842094, + 0.006388107314705849, + 0.01591360941529274, + 0.01798184961080551, + 0.00491293054074049, + 0.009203493595123291, + 0.011425240896642208, + 0.006471746135503054, + 0.003614257089793682, + 0.0030876195523887873, + 0.007751812692731619, + 0.0058059124276041985, + 0.003139694221317768, + 0.005469671916216612, + 0.00547326123341918, + 0.0029606178868561983, + 0.004561486653983593, + 0.002616764744743705, + 0.024377260357141495, + 0.009221747517585754, + 0.002516700653359294, + 0.017019446939229965, + 0.006384701933711767, + 0.003971630707383156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.21100515127182007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.062143512070178986 + ], + "opposite_logprob_tokens": [ + -2.812143564224243 + ], + "logprob_reward": -0.062143512070178986, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062143512070178986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2635258436203003, + 0.0971728265285492, + 0.19917108118534088, + 0.027445031329989433, + 0.1129034236073494, + 0.06848222762346268, + 0.0630393698811531, + 0.07533812522888184, + 0.06231967732310295, + 0.030602386221289635 + ], + "credit_map": [ + 0.03414902836084366, + 0.020120158791542053, + 0.04840303957462311, + 0.038579218089580536, + 0.013353931717574596, + 0.04566768929362297, + 0.031892504543066025, + 0.03136026859283447, + 0.008003403432667255, + 0.005370979197323322, + 0.02787039987742901, + 0.030192658305168152, + 0.003922296222299337, + 0.006588779855519533, + 0.009214943274855614, + 0.006009370554238558, + 0.011048497632145882, + 0.011771083809435368, + 0.06650744378566742, + 0.05639350041747093, + 0.008340511471033096, + 0.01820491813123226, + 0.0193856880068779, + 0.007519438397139311, + 0.0026208204217255116, + 0.0019151418237015605, + 0.005672040395438671, + 0.0038621171843260527, + 0.0014496075455099344, + 0.006708000786602497, + 0.002674048999324441, + 0.0025432526599615812, + 0.014838228933513165, + 0.010273653082549572, + 0.013689679093658924, + 0.02203173190355301, + 0.007937045767903328, + 0.011958496645092964, + 0.015335665084421635, + 0.016838926821947098, + 0.0077642896212637424, + 0.005350528284907341, + 0.013106444850564003, + 0.012087526731193066, + 0.005540112033486366, + 0.010656938888132572, + 0.007702379021793604, + 0.0062740035355091095, + 0.006039815489202738, + 0.003393081482499838, + 0.013354930095374584, + 0.01278908085078001, + 0.004506236873567104, + 0.008866953663527966, + 0.008193357847630978, + 0.005895918235182762, + 0.006909657269716263, + 0.004329435061663389, + 0.013920018449425697, + 0.012909460812807083, + 0.005408959463238716, + 0.014906106516718864, + 0.008566772565245628, + 0.008387714624404907, + 0.005494217853993177, + 0.0039004632271826267, + 0.015623917803168297, + 0.007289016153663397, + 0.0051957350224256516, + 0.010397948324680328, + 0.007507217116653919, + 0.0069111608900129795, + 0.0023989283945411444, + 0.0020801336504518986, + 0.0070554399862885475, + 0.0033375374041497707, + 0.001807724591344595, + 0.007901865988969803, + 0.003527469700202346, + 0.002493287669494748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.062143512070178986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.11297430098056793 + ], + "opposite_logprob_tokens": [ + -2.2379744052886963 + ], + "logprob_reward": -0.11297430098056793, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297430098056793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3689883351325989, + 0.1274479776620865, + 0.12158824503421783, + 0.035801395773887634, + 0.06129453331232071, + 0.06688648462295532, + 0.055691007524728775, + 0.0605040080845356, + 0.043312862515449524, + 0.058485135436058044 + ], + "credit_map": [ + 0.04580260068178177, + 0.02833643928170204, + 0.05570978298783302, + 0.04228106141090393, + 0.023899270221590996, + 0.04677077755331993, + 0.055997345596551895, + 0.07019108533859253, + 0.007827754132449627, + 0.007355010136961937, + 0.03481045737862587, + 0.03584536537528038, + 0.005533903371542692, + 0.014877021312713623, + 0.011373506858944893, + 0.009824972599744797, + 0.009074305184185505, + 0.008893492631614208, + 0.02756645157933235, + 0.024164289236068726, + 0.0049421596340835094, + 0.014452972449362278, + 0.011357487179338932, + 0.021137099713087082, + 0.0030251722782850266, + 0.0019658685196191072, + 0.009298553690314293, + 0.006357075646519661, + 0.0015754764899611473, + 0.007247440051287413, + 0.00399210536852479, + 0.002339704893529415, + 0.008125429041683674, + 0.003662111470475793, + 0.008332081139087677, + 0.011342175304889679, + 0.002844856819137931, + 0.012149919755756855, + 0.0055952477268874645, + 0.009242720901966095, + 0.008216618560254574, + 0.00405921321362257, + 0.01459671463817358, + 0.010678967460989952, + 0.0049410550855100155, + 0.011031179688870907, + 0.006606706418097019, + 0.0067560384050011635, + 0.004751117434352636, + 0.003956754691898823, + 0.01334561500698328, + 0.010097529739141464, + 0.002971070818603039, + 0.008473500609397888, + 0.005948346573859453, + 0.0061470745131373405, + 0.007353692315518856, + 0.004910742398351431, + 0.011086208745837212, + 0.01047928724437952, + 0.004106828011572361, + 0.009664672426879406, + 0.007764472626149654, + 0.005138113163411617, + 0.005220075603574514, + 0.0038515725173056126, + 0.008229658007621765, + 0.006130211520940065, + 0.003690391080453992, + 0.006486909464001656, + 0.006404075771570206, + 0.0032999760005623102, + 0.0032395096495747566, + 0.002964627230539918, + 0.015528888441622257, + 0.007290618494153023, + 0.0033376524224877357, + 0.012301436625421047, + 0.00849174428731203, + 0.005330662243068218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11297430098056793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.0620616190135479 + ], + "opposite_logprob_tokens": [ + -2.8120615482330322 + ], + "logprob_reward": -0.0620616190135479, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620616190135479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2649638056755066, + 0.08646523952484131, + 0.0803128108382225, + 0.06665633618831635, + 0.053703226149082184, + 0.090033620595932, + 0.15105149149894714, + 0.06018993631005287, + 0.1125757247209549, + 0.03404778987169266 + ], + "credit_map": [ + 0.03881808742880821, + 0.01620793156325817, + 0.05334516987204552, + 0.04989533871412277, + 0.012710154056549072, + 0.03626895323395729, + 0.0356423445045948, + 0.022075822576880455, + 0.004814003594219685, + 0.006056782323867083, + 0.02084098756313324, + 0.02182053215801716, + 0.0030430122278630733, + 0.013077929615974426, + 0.010470664128661156, + 0.006341326050460339, + 0.00545417470857501, + 0.006215360481292009, + 0.02083168365061283, + 0.01724374108016491, + 0.0037260735407471657, + 0.008847691118717194, + 0.010483498685061932, + 0.007510586641728878, + 0.005601111799478531, + 0.0036398933734744787, + 0.014318212866783142, + 0.01853475160896778, + 0.0026130732148885727, + 0.012005118653178215, + 0.006584627088159323, + 0.00335954618640244, + 0.005020256619900465, + 0.005028217565268278, + 0.006796430330723524, + 0.009684831835329533, + 0.0033877603709697723, + 0.0112104881554842, + 0.005843467079102993, + 0.0067317732609808445, + 0.010093136690557003, + 0.006958081386983395, + 0.013625146821141243, + 0.013738518580794334, + 0.008189355954527855, + 0.015778910368680954, + 0.011138909496366978, + 0.010511557571589947, + 0.013353399932384491, + 0.010286386124789715, + 0.030485710129141808, + 0.03149103745818138, + 0.010712397284805775, + 0.021964406594634056, + 0.01796755939722061, + 0.01479059737175703, + 0.008462486788630486, + 0.005157143343240023, + 0.009624217636883259, + 0.011987381614744663, + 0.004377661272883415, + 0.007500648032873869, + 0.00734968576580286, + 0.005730708595365286, + 0.014127871952950954, + 0.010947036556899548, + 0.01898401416838169, + 0.01715170592069626, + 0.012569382786750793, + 0.014774280600249767, + 0.015850534662604332, + 0.008170900866389275, + 0.0022740401327610016, + 0.0018087398493662477, + 0.010617155581712723, + 0.004060426261276007, + 0.0015935016563162208, + 0.0071329763159155846, + 0.00409212289378047, + 0.0024688239209353924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0620616190135479 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.00980029534548521 + ], + "opposite_logprob_tokens": [ + -4.634800434112549 + ], + "logprob_reward": -0.00980029534548521, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980029534548521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16802829504013062, + 0.10440566390752792, + 0.061817415058612823, + 0.06598532199859619, + 0.05319247394800186, + 0.05321090295910835, + 0.19950303435325623, + 0.10959833115339279, + 0.052554819732904434, + 0.13170374929904938 + ], + "credit_map": [ + 0.02680000476539135, + 0.016418036073446274, + 0.028913961723446846, + 0.030520154163241386, + 0.008842529729008675, + 0.02411809377372265, + 0.019256221130490303, + 0.013159315101802349, + 0.0065718586556613445, + 0.006601598113775253, + 0.022910678759217262, + 0.02362140454351902, + 0.003949911333620548, + 0.023940524086356163, + 0.009546903893351555, + 0.0072627803310751915, + 0.0043921261094510555, + 0.005309157073497772, + 0.007617943454533815, + 0.007221386302262545, + 0.004252562765032053, + 0.012843390926718712, + 0.008093266747891903, + 0.012087584473192692, + 0.007847398519515991, + 0.0031927761156111956, + 0.009191676042973995, + 0.013742025941610336, + 0.003492678515613079, + 0.016995707526803017, + 0.00613087834790349, + 0.005392177030444145, + 0.003754828590899706, + 0.005252697970718145, + 0.00709120137616992, + 0.006757291499525309, + 0.002904814900830388, + 0.01486621331423521, + 0.007025910075753927, + 0.005539515987038612, + 0.006033619400113821, + 0.0043953219428658485, + 0.008691665716469288, + 0.0073688095435500145, + 0.003926689736545086, + 0.00917012244462967, + 0.007167762145400047, + 0.006456909701228142, + 0.017464835196733475, + 0.010729133151471615, + 0.04022669419646263, + 0.03523940593004227, + 0.013394676148891449, + 0.04147765412926674, + 0.024923091754317284, + 0.016047531738877296, + 0.0085414107888937, + 0.006071307230740786, + 0.020458746701478958, + 0.0268742386251688, + 0.006925464142113924, + 0.02046545408666134, + 0.013450768776237965, + 0.0068109361454844475, + 0.005294578615576029, + 0.003430812619626522, + 0.014927076175808907, + 0.007593946531414986, + 0.0041355867870152, + 0.009302305988967419, + 0.004891563672572374, + 0.0029789467807859182, + 0.006747405976057053, + 0.0032053333707153797, + 0.05245458334684372, + 0.016781188547611237, + 0.003944130148738623, + 0.037071216851472855, + 0.006546074524521828, + 0.004953818395733833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.00980029534548521 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of ultra-high-speed photography, a slightly high-angle medium shot unfolds across a block of raw maple, its pale, grainy surface still showing the rough edges of its original form, firmly mounted on a heavy-duty metal lathe in a dimly lit workshop. The block begins to spin at variable speeds, accelerating into a hypnotic blur under a single, intense overhead task light, casting dramatic, deep shadows and illuminating fine sawdust suspended in mid-air. A sharp steel chisel with a dark, polished wooden handle is steadily and deliberately guided toward the spinning surface. As its gleaming edge makes contact with the wood, the entire action transforms into breathtaking slow motion. Tiny chips erupt from the point of impact, each gossamer-thin strip of wood catching the bright light and spiraling away in elegant, suspended arcs against the dark, out-of-focus background. Meanwhile, the camera makes steady, deliberate pushes, ever closer to the moment of creation, zooming in on the violent yet graceful interaction between tool and material, revealing intricate, silky-smooth organic patterns emerging from the raw wood, creating a captivating and focused atmosphere of pure craftsmanship.", + "seed": 29865, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_4/index_592_seed_29865_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": -0.058579342905431986, + "vlm_logprob": -0.058579342905431986, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.058579342905431986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.19474224746227264 + ], + "opposite_logprob_tokens": [ + -1.819742202758789 + ], + "logprob_reward": -0.19474224746227264, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19474224746227264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08878640085458755, + 0.06031867489218712, + 0.12812231481075287, + 0.04118531569838524, + 0.2650979161262512, + 0.07383362948894501, + 0.05240712687373161, + 0.17975331842899323, + 0.05296112969517708, + 0.05753413215279579 + ], + "credit_map": [ + 0.0125025799497962, + 0.007387688849121332, + 0.01562805287539959, + 0.0071652354672551155, + 0.007132228929549456, + 0.00939533393830061, + 0.017940644174814224, + 0.011634642258286476, + 0.008684023283421993, + 0.004216229077428579, + 0.0038249646313488483, + 0.004886890761554241, + 0.005194187629967928, + 0.005846461281180382, + 0.018143579363822937, + 0.009522336535155773, + 0.027836928144097328, + 0.007253098767250776, + 0.03273895010352135, + 0.012408970855176449, + 0.006416297983378172, + 0.007676950190216303, + 0.01119100209325552, + 0.022600119933485985, + 0.0060651893727481365, + 0.002799722133204341, + 0.002966705011203885, + 0.0041085113771259785, + 0.0040704854764044285, + 0.004208252299576998, + 0.011975331231951714, + 0.004991118796169758, + 0.03997315838932991, + 0.015284465625882149, + 0.013055856339633465, + 0.01371816173195839, + 0.023620503023266792, + 0.015319141559302807, + 0.09891090542078018, + 0.045215729624032974, + 0.01914225146174431, + 0.006048095878213644, + 0.008044368587434292, + 0.005811967886984348, + 0.005969308316707611, + 0.005745328962802887, + 0.011764987371861935, + 0.011307324282824993, + 0.006627116817981005, + 0.0029746023938059807, + 0.0040276688523590565, + 0.002583532128483057, + 0.0071548293344676495, + 0.004137515556067228, + 0.016617467626929283, + 0.008284395560622215, + 0.026216216385364532, + 0.01189311034977436, + 0.019877204671502113, + 0.008062837645411491, + 0.014067807234823704, + 0.011319431476294994, + 0.050530411303043365, + 0.03778629004955292, + 0.007184803951531649, + 0.003839761484414339, + 0.010221298784017563, + 0.004818171728402376, + 0.0055447109043598175, + 0.004415470641106367, + 0.007605031132698059, + 0.009331880137324333, + 0.0062556141056120396, + 0.0041074748151004314, + 0.013792763464152813, + 0.006773251108825207, + 0.005068364553153515, + 0.004988396540284157, + 0.008327513001859188, + 0.008220754563808441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.19474224746227264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.009815995581448078 + ], + "opposite_logprob_tokens": [ + -4.6348161697387695 + ], + "logprob_reward": -0.009815995581448078, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009815995581448078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26881828904151917, + 0.0691070407629013, + 0.07872172445058823, + 0.0444139689207077, + 0.10650740563869476, + 0.03735387697815895, + 0.014340062625706196, + 0.2979636788368225, + 0.04142477363348007, + 0.04134928435087204 + ], + "credit_map": [ + 0.03961668163537979, + 0.017763227224349976, + 0.0677870512008667, + 0.03463456407189369, + 0.019051704555749893, + 0.026194466277956963, + 0.027760976925492287, + 0.036009594798088074, + 0.011099969036877155, + 0.006497236900031567, + 0.006817704066634178, + 0.006974290125072002, + 0.006322096101939678, + 0.006169124972075224, + 0.013682203367352486, + 0.011544406414031982, + 0.012399980798363686, + 0.0047181048430502415, + 0.021600840613245964, + 0.008292688988149166, + 0.005841721780598164, + 0.009298176504671574, + 0.006250240374356508, + 0.010319964028894901, + 0.0071027688682079315, + 0.0026448844000697136, + 0.0072548482567071915, + 0.005117308814078569, + 0.004197311587631702, + 0.0055583869107067585, + 0.007974260486662388, + 0.004564195871353149, + 0.01393122784793377, + 0.006809603422880173, + 0.006676273886114359, + 0.008200231939554214, + 0.008350308984518051, + 0.007076148875057697, + 0.033770427107810974, + 0.021693173795938492, + 0.006802329793572426, + 0.003299376228824258, + 0.007577765733003616, + 0.0032464214600622654, + 0.0033928717020899057, + 0.003574006026610732, + 0.005694305524230003, + 0.003766795853152871, + 0.0019281486747786403, + 0.0010946060065180063, + 0.0019519591005519032, + 0.0018831692868843675, + 0.001766306348145008, + 0.0020470484159886837, + 0.002242129063233733, + 0.0014266945654526353, + 0.041105616837739944, + 0.018511760979890823, + 0.01012752391397953, + 0.010090778581798077, + 0.027729932218790054, + 0.016323009505867958, + 0.09980964660644531, + 0.07426537573337555, + 0.0053061433136463165, + 0.0031098793260753155, + 0.01192763727158308, + 0.004085612017661333, + 0.004542369395494461, + 0.0032908867578953505, + 0.004378251265734434, + 0.004783989861607552, + 0.005009300075471401, + 0.0026358503382653, + 0.010150223970413208, + 0.004415690433233976, + 0.003278750693425536, + 0.004779859445989132, + 0.004935821983963251, + 0.006143782287836075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.009815995581448078 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.03371435031294823 + ], + "opposite_logprob_tokens": [ + -3.4087142944335938 + ], + "logprob_reward": -0.03371435031294823, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371435031294823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31515538692474365, + 0.031684644520282745, + 0.20074845850467682, + 0.043613724410533905, + 0.050763312727212906, + 0.08178827911615372, + 0.03422194719314575, + 0.12455494701862335, + 0.03167830780148506, + 0.08579102903604507 + ], + "credit_map": [ + 0.052238717675209045, + 0.02583455666899681, + 0.06179232895374298, + 0.038742925971746445, + 0.021859437227249146, + 0.03252939507365227, + 0.04776467755436897, + 0.03439334034919739, + 0.004502157215029001, + 0.003135669743642211, + 0.004215527791529894, + 0.003552268259227276, + 0.002898287260904908, + 0.003403901122510433, + 0.006077394355088472, + 0.0038994383066892624, + 0.030958453193306923, + 0.012246230617165565, + 0.04491337016224861, + 0.02265857718884945, + 0.012154074385762215, + 0.029547810554504395, + 0.015133856795728207, + 0.033136073499917984, + 0.008140318095684052, + 0.0028089822735637426, + 0.011359654366970062, + 0.00509921507909894, + 0.003391887992620468, + 0.004432171117514372, + 0.004135800991207361, + 0.004245693329721689, + 0.007228085305541754, + 0.0037803275045007467, + 0.003926544915884733, + 0.0040574572049081326, + 0.004828168544918299, + 0.0032394391018897295, + 0.015206953510642052, + 0.008496335707604885, + 0.022759826853871346, + 0.0064470707438886166, + 0.013673524372279644, + 0.005591418594121933, + 0.004387504421174526, + 0.004928908310830593, + 0.004846648313105106, + 0.019153377041220665, + 0.006214420776814222, + 0.002042486099526286, + 0.006017585750669241, + 0.0021787185687571764, + 0.003512800205498934, + 0.004425789229571819, + 0.005063801538199186, + 0.004766346886754036, + 0.017256293445825577, + 0.007941655814647675, + 0.0050962031818926334, + 0.004562120418995619, + 0.01117525715380907, + 0.007236930541694164, + 0.04082971438765526, + 0.030456777662038803, + 0.004490169696509838, + 0.002591439988464117, + 0.009622235782444477, + 0.0026390114799141884, + 0.0028544075321406126, + 0.0027726569678634405, + 0.0030762506648898125, + 0.0036321328952908516, + 0.013911882415413857, + 0.006085090804845095, + 0.026058921590447426, + 0.0074737099930644035, + 0.007017830852419138, + 0.008523384109139442, + 0.006226730532944202, + 0.010493476875126362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.03371435031294823 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.04866227135062218 + ], + "opposite_logprob_tokens": [ + -3.0486621856689453 + ], + "logprob_reward": -0.04866227135062218, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866227135062218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2753799855709076, + 0.028487661853432655, + 0.07629620283842087, + 0.028957940638065338, + 0.24023257195949554, + 0.05513259768486023, + 0.028739016503095627, + 0.10796588659286499, + 0.08779450505971909, + 0.07101361453533173 + ], + "credit_map": [ + 0.03522222489118576, + 0.01918172836303711, + 0.06076817214488983, + 0.029406843706965446, + 0.02003524824976921, + 0.022718172520399094, + 0.05542423948645592, + 0.032623350620269775, + 0.004375105258077383, + 0.0030661316122859716, + 0.002855192171409726, + 0.003945191856473684, + 0.0035210002679377794, + 0.003100736066699028, + 0.0039261761121451855, + 0.0036981271114200354, + 0.011911374516785145, + 0.0045931111089885235, + 0.02682781033217907, + 0.009409082122147083, + 0.00434924615547061, + 0.00495870178565383, + 0.005142862442880869, + 0.009104016236960888, + 0.004541380330920219, + 0.0021896616090089083, + 0.00752056622877717, + 0.0037432743702083826, + 0.0021210177801549435, + 0.0024952504318207502, + 0.0027032666839659214, + 0.003643524833023548, + 0.03825441375374794, + 0.012557181529700756, + 0.010162522085011005, + 0.011935219168663025, + 0.022138696163892746, + 0.012588145211338997, + 0.09222401678562164, + 0.040372371673583984, + 0.007121148519217968, + 0.004653465934097767, + 0.012042750604450703, + 0.004531595390290022, + 0.003977328073233366, + 0.0057696993462741375, + 0.007805019151419401, + 0.00923158973455429, + 0.004048203118145466, + 0.0017488833982497454, + 0.004268024116754532, + 0.002029444556683302, + 0.003375200554728508, + 0.002866014139726758, + 0.0064393142238259315, + 0.003963930997997522, + 0.014925112947821617, + 0.00716090202331543, + 0.004060299601405859, + 0.00356621784158051, + 0.009332907386124134, + 0.0057706995867192745, + 0.0353909470140934, + 0.02775879204273224, + 0.012689118273556232, + 0.008159942924976349, + 0.020794302225112915, + 0.011133254505693913, + 0.00889566820114851, + 0.010092443786561489, + 0.007041555363684893, + 0.00898822396993637, + 0.011670331470668316, + 0.00464543467387557, + 0.01145370863378048, + 0.0037917683366686106, + 0.006671104114502668, + 0.005643849726766348, + 0.011797362938523293, + 0.01534005906432867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.04866227135062218 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.005961849819868803 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005961849819868803, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961849819868803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1691068410873413, + 0.0563705712556839, + 0.11698887497186661, + 0.05223766714334488, + 0.2743341326713562, + 0.04742816463112831, + 0.06965415924787521, + 0.09706103056669235, + 0.046107131987810135, + 0.0707114115357399 + ], + "credit_map": [ + 0.025922605767846107, + 0.013715684413909912, + 0.03593186289072037, + 0.026470817625522614, + 0.015432660467922688, + 0.018180333077907562, + 0.015983089804649353, + 0.017469776794314384, + 0.008061536587774754, + 0.005010000895708799, + 0.006130626425147057, + 0.004431919660419226, + 0.005939468275755644, + 0.00723875779658556, + 0.011863761581480503, + 0.007694502826780081, + 0.017717035487294197, + 0.008180743083357811, + 0.02919275127351284, + 0.013523269444704056, + 0.007898394949734211, + 0.011554044671356678, + 0.011966675519943237, + 0.016955947503447533, + 0.00964063685387373, + 0.003684757277369499, + 0.013425019569694996, + 0.004177706316113472, + 0.00440408568829298, + 0.007141427602618933, + 0.004458549898117781, + 0.00530547508969903, + 0.03919891640543938, + 0.015840040519833565, + 0.01221766322851181, + 0.013392258435487747, + 0.02532682754099369, + 0.015319946222007275, + 0.1010361760854721, + 0.052002303302288055, + 0.00957677885890007, + 0.004254611674696207, + 0.006043617147952318, + 0.003356478875502944, + 0.00504946568980813, + 0.006055270321667194, + 0.007315709721297026, + 0.005776230711489916, + 0.010436042211949825, + 0.004151404369622469, + 0.006906295195221901, + 0.003960481844842434, + 0.008322298526763916, + 0.007618810515850782, + 0.016783546656370163, + 0.011475280858576298, + 0.013874326832592487, + 0.006737749557942152, + 0.005528956651687622, + 0.004665866028517485, + 0.006714745424687862, + 0.006160547956824303, + 0.028264308348298073, + 0.025114523246884346, + 0.006613458972424269, + 0.003620765171945095, + 0.011149299331009388, + 0.004028386436402798, + 0.006283521186560392, + 0.0045687733218073845, + 0.004354990553110838, + 0.005487932823598385, + 0.00986520480364561, + 0.005779861938208342, + 0.013395093381404877, + 0.008017981424927711, + 0.0071522267535328865, + 0.00878332182765007, + 0.01013139821588993, + 0.0075863199308514595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.005961849819868803 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the style of ultra-high-speed photography, a slightly high-angle medium shot unfolds across a block of raw maple, its pale, grainy surface still showing the rough edges of its original form, firmly mounted on a heavy-duty metal lathe in a dimly lit workshop. The block begins to spin at variable speeds, accelerating into a hypnotic blur under a single, intense overhead task light, casting dramatic, deep shadows and illuminating fine sawdust suspended in mid-air. A sharp steel chisel with a dark, polished wooden handle is steadily and deliberately guided toward the spinning surface. As its gleaming edge makes contact with the wood, the entire action transforms into breathtaking slow motion. Tiny chips erupt from the point of impact, each gossamer-thin strip of wood catching the bright light and spiraling away in elegant, suspended arcs against the dark, out-of-focus background. Meanwhile, the camera makes steady, deliberate pushes, ever closer to the moment of creation, zooming in on the violent yet graceful interaction between tool and material, revealing intricate, silky-smooth organic patterns emerging from the raw wood, creating a captivating and focused atmosphere of pure craftsmanship.", + "seed": 29866, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_4/index_592_seed_29866_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": -0.05426303213462234, + "vlm_logprob": -0.05426303213462234, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05426303213462234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.17360538244247437 + ], + "opposite_logprob_tokens": [ + -1.9236054420471191 + ], + "logprob_reward": -0.17360538244247437, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17360538244247437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1630399227142334, + 0.09423080831766129, + 0.05993042513728142, + 0.14637444913387299, + 0.12417957186698914, + 0.06565175205469131, + 0.07260946184396744, + 0.11604149639606476, + 0.11117777973413467, + 0.04676428064703941 + ], + "credit_map": [ + 0.03172975033521652, + 0.01817280240356922, + 0.027766110375523567, + 0.018218737095594406, + 0.009447881951928139, + 0.036265574395656586, + 0.01136540062725544, + 0.010073672980070114, + 0.020847098901867867, + 0.007616551127284765, + 0.012991656549274921, + 0.015085747465491295, + 0.011133989319205284, + 0.012619133107364178, + 0.007836034521460533, + 0.006100609432905912, + 0.008008912205696106, + 0.0052442592568695545, + 0.007455846294760704, + 0.009260249324142933, + 0.008138506673276424, + 0.010193156078457832, + 0.005524502135813236, + 0.006104994565248489, + 0.029760275036096573, + 0.012081013061106205, + 0.023210547864437103, + 0.023682544007897377, + 0.012568286620080471, + 0.023487495258450508, + 0.01152016781270504, + 0.010064126923680305, + 0.016012193635106087, + 0.013545196503400803, + 0.023605745285749435, + 0.013387994840741158, + 0.012427338398993015, + 0.014459523372352123, + 0.019134342670440674, + 0.01160725113004446, + 0.007107968907803297, + 0.009647265076637268, + 0.008949528448283672, + 0.013458638451993465, + 0.0070504858158528805, + 0.007882194593548775, + 0.005591832101345062, + 0.005963847506791353, + 0.0094335051253438, + 0.008645555004477501, + 0.009837551973760128, + 0.009947161190211773, + 0.010728678666055202, + 0.009074577130377293, + 0.009533647447824478, + 0.005408790893852711, + 0.021166259422898293, + 0.019797587767243385, + 0.015443126671016216, + 0.015090013854205608, + 0.01666777953505516, + 0.013630975969135761, + 0.006652037613093853, + 0.007593723013997078, + 0.009265064261853695, + 0.01750217191874981, + 0.012497063726186752, + 0.008815763518214226, + 0.029010651633143425, + 0.02401968091726303, + 0.004481307230889797, + 0.005586086306720972, + 0.006627274677157402, + 0.005324901081621647, + 0.005926573183387518, + 0.005459209438413382, + 0.005114702973514795, + 0.0075471000745892525, + 0.004608844872564077, + 0.006155676674097776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.17360538244247437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.02639020048081875 + ], + "opposite_logprob_tokens": [ + -3.651390314102173 + ], + "logprob_reward": -0.02639020048081875, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02639020048081875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13355395197868347, + 0.051106665283441544, + 0.08432063460350037, + 0.11663087457418442, + 0.20798338949680328, + 0.11874022334814072, + 0.10833960026502609, + 0.06269555538892746, + 0.03118947707116604, + 0.08543964475393295 + ], + "credit_map": [ + 0.011050845496356487, + 0.011767965741455555, + 0.017115598544478416, + 0.03723021596670151, + 0.01515556126832962, + 0.018908556550741196, + 0.013075622729957104, + 0.009249577298760414, + 0.005393460858613253, + 0.005930757150053978, + 0.006398331373929977, + 0.009342173114418983, + 0.006378996651619673, + 0.006959796883165836, + 0.005952221341431141, + 0.004750927910208702, + 0.009536522440612316, + 0.007281712722033262, + 0.013032997958362103, + 0.02151969075202942, + 0.005808211397379637, + 0.011487210169434547, + 0.007579371798783541, + 0.008074911311268806, + 0.02375791035592556, + 0.009375395253300667, + 0.019243229180574417, + 0.014330580830574036, + 0.009159748442471027, + 0.023410614579916, + 0.008582974784076214, + 0.008770421147346497, + 0.020642532035708427, + 0.026212209835648537, + 0.027598707005381584, + 0.024285344406962395, + 0.027501218020915985, + 0.025271430611610413, + 0.03364551439881325, + 0.022826440632343292, + 0.01351280976086855, + 0.017267411574721336, + 0.016486456617712975, + 0.022786928340792656, + 0.012208441272377968, + 0.015055890195071697, + 0.009752573445439339, + 0.01166971679776907, + 0.018271684646606445, + 0.01218438521027565, + 0.014939706772565842, + 0.01427318062633276, + 0.01621277630329132, + 0.012853334657847881, + 0.012851675972342491, + 0.00675284955650568, + 0.01226865965873003, + 0.009301802143454552, + 0.00733236875385046, + 0.0066383411176502705, + 0.007591709494590759, + 0.012546555139124393, + 0.0032869402784854174, + 0.00372917577624321, + 0.006067286245524883, + 0.004519365727901459, + 0.0052134632132947445, + 0.0034558966290205717, + 0.0030419486574828625, + 0.003946448676288128, + 0.002005312591791153, + 0.002939754631370306, + 0.00989631935954094, + 0.009418913163244724, + 0.009556828066706657, + 0.008900209330022335, + 0.007428912911564112, + 0.01229916699230671, + 0.012415693141520023, + 0.015523596666753292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.02639020048081875 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.043084751814603806 + ], + "opposite_logprob_tokens": [ + -3.1680848598480225 + ], + "logprob_reward": -0.043084751814603806, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043084751814603806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12418577820062637, + 0.06823026388883591, + 0.06474800407886505, + 0.14126098155975342, + 0.1418197900056839, + 0.11355917900800705, + 0.14707531034946442, + 0.06887011975049973, + 0.0433214046061039, + 0.08692920953035355 + ], + "credit_map": [ + 0.022840790450572968, + 0.011616245843470097, + 0.016693660989403725, + 0.018585436046123505, + 0.009172898717224598, + 0.017452972009778023, + 0.014430561102926731, + 0.01339319534599781, + 0.008968132548034191, + 0.005900396965444088, + 0.008349505253136158, + 0.014942334964871407, + 0.008266902528703213, + 0.008936827071011066, + 0.0077297319658100605, + 0.005136425606906414, + 0.01181881595402956, + 0.005720424931496382, + 0.008938507176935673, + 0.008909513242542744, + 0.009067350998520851, + 0.009767690673470497, + 0.006514376495033503, + 0.0040113190189003944, + 0.030010001733899117, + 0.012410879135131836, + 0.023650890216231346, + 0.01738104224205017, + 0.012856422923505306, + 0.02481447346508503, + 0.010130446404218674, + 0.010006819851696491, + 0.015027577057480812, + 0.018059266731142998, + 0.022732798010110855, + 0.01842942275106907, + 0.017934445291757584, + 0.01869756355881691, + 0.01931152492761612, + 0.011627184227108955, + 0.016067730262875557, + 0.012015310116112232, + 0.013087905943393707, + 0.019996628165245056, + 0.00971824862062931, + 0.018812481313943863, + 0.010459634475409985, + 0.013401239179074764, + 0.016825150698423386, + 0.020650044083595276, + 0.02093815803527832, + 0.01762455329298973, + 0.023844748735427856, + 0.016862617805600166, + 0.02036910317838192, + 0.009960921481251717, + 0.009245740249752998, + 0.012561715207993984, + 0.009004252031445503, + 0.008938898332417011, + 0.01336114201694727, + 0.006912470329552889, + 0.004627141170203686, + 0.004218761343508959, + 0.006342368666082621, + 0.005635420326143503, + 0.0060269772075116634, + 0.00519622303545475, + 0.005295413546264172, + 0.008333035744726658, + 0.0026635394897311926, + 0.003828421700745821, + 0.010499566793441772, + 0.009041831828653812, + 0.00990335177630186, + 0.00822383165359497, + 0.009602425619959831, + 0.011099761351943016, + 0.01151252444833517, + 0.01704590395092964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.043084751814603806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.02058931440114975 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058931440114975, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058931440114975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15603695809841156, + 0.04978792741894722, + 0.028646746650338173, + 0.15585318207740784, + 0.07746172696352005, + 0.08179698139429092, + 0.1810344010591507, + 0.05056291073560715, + 0.15840019285678864, + 0.06041904538869858 + ], + "credit_map": [ + 0.0210734810680151, + 0.012148749083280563, + 0.016762221232056618, + 0.0424262136220932, + 0.015541023574769497, + 0.023080358281731606, + 0.013466496020555496, + 0.011538399383425713, + 0.0050902883522212505, + 0.003527127206325531, + 0.006529609672725201, + 0.010447792708873749, + 0.006097900215536356, + 0.0073255011811852455, + 0.005641568452119827, + 0.005128130782395601, + 0.003192100441083312, + 0.0026256449054926634, + 0.0036001023836433887, + 0.005947127938270569, + 0.003728499636054039, + 0.004068454261869192, + 0.0029342439956963062, + 0.0025505688972771168, + 0.026462117210030556, + 0.012386349029839039, + 0.02899320237338543, + 0.019231121987104416, + 0.014711556024849415, + 0.03313940763473511, + 0.011252512224018574, + 0.009676894173026085, + 0.010831648483872414, + 0.009377989917993546, + 0.012385492213070393, + 0.010275650769472122, + 0.008498500101268291, + 0.0086266053840518, + 0.010534039698541164, + 0.006931796669960022, + 0.010301763191819191, + 0.010459855198860168, + 0.011290009133517742, + 0.01220169197767973, + 0.008299428969621658, + 0.011569581925868988, + 0.007847185246646404, + 0.009827454574406147, + 0.020687121897935867, + 0.02515803463757038, + 0.02785283699631691, + 0.021384308114647865, + 0.02798706665635109, + 0.01968664675951004, + 0.0251998882740736, + 0.013078472577035427, + 0.01097526028752327, + 0.0068931360729038715, + 0.006618846207857132, + 0.005422479007393122, + 0.006130784284323454, + 0.005658593028783798, + 0.004453726578503847, + 0.004410079214721918, + 0.016598062589764595, + 0.025838728994131088, + 0.01800697110593319, + 0.016959497705101967, + 0.03504006192088127, + 0.030300553888082504, + 0.007239580620080233, + 0.00841672532260418, + 0.009097697213292122, + 0.00783996470272541, + 0.008193169720470905, + 0.006630117539316416, + 0.007127124350517988, + 0.007411709055304527, + 0.006000544875860214, + 0.008118714205920696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.02058931440114975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit workshop featuring a block of raw maple mounted on a heavy-duty metal lathe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a sharp steel chisel with a dark, polished wooden handle being guided toward the spinning wood block?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are tiny wood chips shown erupting from the point of impact and spiraling away in suspended arcs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera make steady, deliberate pushes closer to the action while zooming in on the interaction between the tool and the wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the action of the chisel making contact with the spinning wood depicted in 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": [ + -0.007645511534065008 + ], + "opposite_logprob_tokens": [ + -4.882645606994629 + ], + "logprob_reward": -0.007645511534065008, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645511534065008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2817170321941376, + 0.05075615271925926, + 0.0615813173353672, + 0.10235466063022614, + 0.09563705325126648, + 0.08689327538013458, + 0.13020239770412445, + 0.07827891409397125, + 0.08351403474807739, + 0.02906511351466179 + ], + "credit_map": [ + 0.04307312145829201, + 0.019722197204828262, + 0.03977900370955467, + 0.06004396453499794, + 0.01906386949121952, + 0.062333036214113235, + 0.019569985568523407, + 0.018131885677576065, + 0.006995546165853739, + 0.004337562248110771, + 0.007267872802913189, + 0.008830619044601917, + 0.007500358857214451, + 0.007471416145563126, + 0.00512924836948514, + 0.0032235339749604464, + 0.009476820938289165, + 0.006217555142939091, + 0.007030020002275705, + 0.010517326183617115, + 0.010302607901394367, + 0.00763289351016283, + 0.00636408431455493, + 0.004040009342133999, + 0.022803232073783875, + 0.009098962880671024, + 0.016384519636631012, + 0.012073906138539314, + 0.007555954623967409, + 0.014302574098110199, + 0.009591378271579742, + 0.01054413616657257, + 0.011790410615503788, + 0.01019157376140356, + 0.020404547452926636, + 0.010167108848690987, + 0.011087588965892792, + 0.012500549666583538, + 0.012067972682416439, + 0.007427312433719635, + 0.011896898970007896, + 0.011280464008450508, + 0.012288551777601242, + 0.013573786243796349, + 0.008620980195701122, + 0.011279454454779625, + 0.009430434554815292, + 0.00852271355688572, + 0.014368535950779915, + 0.022563623264431953, + 0.01967792771756649, + 0.014985931105911732, + 0.019722364842891693, + 0.01551730465143919, + 0.016006512567400932, + 0.007360204588621855, + 0.013925641775131226, + 0.012223247438669205, + 0.012736266478896141, + 0.008778625167906284, + 0.011508294381201267, + 0.008481164462864399, + 0.004879762884229422, + 0.005745920818299055, + 0.005783350206911564, + 0.014073768630623817, + 0.010581085458397865, + 0.00901762768626213, + 0.02167191356420517, + 0.014631859958171844, + 0.004040395841002464, + 0.0037140375934541225, + 0.00424629170447588, + 0.003280776320025325, + 0.003860931843519211, + 0.0025323666632175446, + 0.0026186967734247446, + 0.005194549914449453, + 0.003320851130411029, + 0.004010652657598257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.007645511534065008 + } + ], + "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/0000058/group_5/index_tensor([9757, 9757])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_5/index_tensor([9757, 9757])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..78731dcbbe32a31ac61e95d4438a17c0e2d1ad00 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_5/index_tensor([9757, 9757])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, low-angle, static shot captures the moment of brutal destruction in extreme slow motion. The bulldozer's massive, weathered yellow steel blade, deeply scratched and caked with dried mud, crashes unstoppably into the side of a parked, dark metallic red sedan. Under the glare of the midday sun, the initial impact is a silent ballet. The car's front fenders and hood wrinkle violently, the glossy red paint cracking and flaking as the underlying metal tears and folds inward like a sheet of paper under the immense, deliberate pressure. Simultaneously, the windshield and side windows burst open; spiderweb-like cracks flash across the glass before exploding outward in a flash of countless tiny fragments. Each crystalline shard hangs suspended in midair, individually catching and refracting the bright sunlight as dust and debris on the ground slowly swirl around the point of impact, creating an atmosphere of controlled chaos that is both visceral and terrifyingly beautiful.", + "seed": 24741, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_5/index_9757_seed_24741_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": -0.46284620724618436, + "vlm_logprob": -0.46284620724618436, + "vlm_yes_no_margin": 1.724999988079071, + "vlm_token_credit": -0.46284620724618436, + "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 captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17121732234954834 + ], + "opposite_logprob_tokens": [ + -2.296217441558838 + ], + "logprob_reward": -0.17121732234954834, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.17121732234954834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12093951553106308, + 0.061607420444488525, + 0.04194150120019913, + 0.06605284661054611, + 0.07315191626548767, + 0.17949654161930084, + 0.19905008375644684, + 0.08618415892124176, + 0.08356379717588425, + 0.08801218867301941 + ], + "credit_map": [ + 0.014592436142265797, + 0.010513979010283947, + 0.008169421926140785, + 0.010108664631843567, + 0.009183494374155998, + 0.013618675991892815, + 0.024907393380999565, + 0.02984546311199665, + 0.007524474989622831, + 0.019764810800552368, + 0.0036247684620320797, + 0.0037188411224633455, + 0.00615726737305522, + 0.005731839686632156, + 0.007815966382622719, + 0.007269452325999737, + 0.003726044436916709, + 0.004621457774192095, + 0.0035474924370646477, + 0.004431396257132292, + 0.011316105723381042, + 0.004540942143648863, + 0.004824946168810129, + 0.004933120682835579, + 0.010057945735752583, + 0.016888875514268875, + 0.005157082341611385, + 0.005684188567101955, + 0.010744408704340458, + 0.005594867281615734, + 0.006718034390360117, + 0.005207451060414314, + 0.009508905932307243, + 0.015982139855623245, + 0.012195233255624771, + 0.008945227600634098, + 0.007798964157700539, + 0.005531764589250088, + 0.006832079961895943, + 0.006357613485306501, + 0.02192290686070919, + 0.06722705811262131, + 0.009826691821217537, + 0.01604642905294895, + 0.03784796968102455, + 0.00827760435640812, + 0.010164217092096806, + 0.008183671161532402, + 0.01868635229766369, + 0.06343746930360794, + 0.039159394800662994, + 0.017312897369265556, + 0.030465859919786453, + 0.009743697941303253, + 0.009361881762742996, + 0.010882556438446045, + 0.007937818765640259, + 0.009654934518039227, + 0.01144377700984478, + 0.016459621489048004, + 0.01204879954457283, + 0.011313742958009243, + 0.010215607471764088, + 0.0071098580956459045, + 0.012984824366867542, + 0.01898673176765442, + 0.009927197359502316, + 0.007970117032527924, + 0.010753539390861988, + 0.007161307148635387, + 0.009554619900882244, + 0.00622545974329114, + 0.018573421984910965, + 0.010591442696750164, + 0.008944125846028328, + 0.00908592063933611, + 0.005909500177949667, + 0.005335836671292782, + 0.010685418732464314, + 0.01888652704656124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.17121732234954834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.7015671730041504 + ], + "opposite_logprob_tokens": [ + -0.20156711339950562 + ], + "logprob_reward": -1.7015671730041504, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015671730041504, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.038830336183309555, + 0.041660912334918976, + 0.04626774415373802, + 0.04672711342573166, + 0.021427899599075317, + 0.13390670716762543, + 0.519240140914917, + 0.04402093216776848, + 0.08558874577283859, + 0.022329537197947502 + ], + "credit_map": [ + 0.0045297518372535706, + 0.005498232785612345, + 0.002549986122176051, + 0.003929235506802797, + 0.004005802795290947, + 0.004914250690490007, + 0.007040230557322502, + 0.006362841930240393, + 0.006273030303418636, + 0.008753148838877678, + 0.00241876975633204, + 0.002964883344247937, + 0.004633890464901924, + 0.004407624714076519, + 0.006812045816332102, + 0.005397516302764416, + 0.005803052335977554, + 0.004453199449926615, + 0.0054411557503044605, + 0.0049987840466201305, + 0.00857167411595583, + 0.005335175432264805, + 0.0070632207207381725, + 0.004601474851369858, + 0.005070367828011513, + 0.008512004278600216, + 0.002606724388897419, + 0.005693809594959021, + 0.005075826775282621, + 0.003170088166370988, + 0.011364743113517761, + 0.005233545321971178, + 0.0027684206143021584, + 0.004921110346913338, + 0.002529905876144767, + 0.003321622731164098, + 0.002518568653613329, + 0.0016709137707948685, + 0.0017913805786520243, + 0.001905975048430264, + 0.014064394868910313, + 0.04529412463307381, + 0.007039355579763651, + 0.008670269511640072, + 0.032557517290115356, + 0.010343480855226517, + 0.00837274082005024, + 0.007564807776361704, + 0.04613550007343292, + 0.15586286783218384, + 0.11345911026000977, + 0.04124435782432556, + 0.07862615585327148, + 0.03274476155638695, + 0.027955370023846626, + 0.023211976513266563, + 0.0035511613823473454, + 0.004599709063768387, + 0.004312324803322554, + 0.00620535621419549, + 0.007703783456236124, + 0.007597845047712326, + 0.006254918873310089, + 0.003795830300077796, + 0.012014973908662796, + 0.015931354835629463, + 0.010311957448720932, + 0.007880465127527714, + 0.007832062430679798, + 0.0091200340539217, + 0.011203057132661343, + 0.011294828727841377, + 0.003879894968122244, + 0.00276300311088562, + 0.0030907816253602505, + 0.002869897522032261, + 0.0014047177974134684, + 0.0015526422066614032, + 0.0026067483704537153, + 0.0041618505492806435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -1.7015671730041504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38712555170059204 + ], + "opposite_logprob_tokens": [ + -1.1371254920959473 + ], + "logprob_reward": -0.38712555170059204, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38712555170059204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09200549125671387, + 0.10234656929969788, + 0.05249590054154396, + 0.06939121335744858, + 0.057669106870889664, + 0.10549529641866684, + 0.24673421680927277, + 0.06980271637439728, + 0.06166341155767441, + 0.14239603281021118 + ], + "credit_map": [ + 0.01037646271288395, + 0.014110902324318886, + 0.006210571154952049, + 0.009432223625481129, + 0.008200626820325851, + 0.007929234765470028, + 0.019782064482569695, + 0.015963416546583176, + 0.011466431431472301, + 0.020161157473921776, + 0.006349699571728706, + 0.008423413150012493, + 0.011449627578258514, + 0.011311355978250504, + 0.01837165094912052, + 0.01481323316693306, + 0.0052787549793720245, + 0.005949494428932667, + 0.004286806099116802, + 0.005495857447385788, + 0.01169910840690136, + 0.00493732700124383, + 0.00895524863153696, + 0.00589330680668354, + 0.011436093598604202, + 0.017590906471014023, + 0.006543301977217197, + 0.009005097672343254, + 0.010383658111095428, + 0.005456998944282532, + 0.005133284255862236, + 0.0038418711628764868, + 0.006585089489817619, + 0.009121297858655453, + 0.009205927141010761, + 0.010817148722708225, + 0.005866373889148235, + 0.005485601723194122, + 0.00623532896861434, + 0.004352336749434471, + 0.010380401276051998, + 0.035816773772239685, + 0.005124377552419901, + 0.006926271133124828, + 0.027423759922385216, + 0.00784603226929903, + 0.006337309256196022, + 0.005640383344143629, + 0.0219207014888525, + 0.04577219858765602, + 0.05258899927139282, + 0.022150004282593727, + 0.040113285183906555, + 0.028562799096107483, + 0.01801796816289425, + 0.017608273774385452, + 0.005402507726103067, + 0.007799494080245495, + 0.007213956210762262, + 0.01342947781085968, + 0.010645175352692604, + 0.00816368218511343, + 0.010564780794084072, + 0.006583644542843103, + 0.009273966774344444, + 0.01122059766203165, + 0.007085382007062435, + 0.005599534139037132, + 0.008198533207178116, + 0.008179208263754845, + 0.006132056005299091, + 0.005974133498966694, + 0.032673247158527374, + 0.018853923305869102, + 0.016629237681627274, + 0.01360320020467043, + 0.006934199947863817, + 0.007692602463066578, + 0.012050924822688103, + 0.03395869582891464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.38712555170059204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609749481081963 + ], + "opposite_logprob_tokens": [ + -3.8956098556518555 + ], + "logprob_reward": -0.020609749481081963, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609749481081963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08476363122463226, + 0.042220909148454666, + 0.05869539827108383, + 0.10340361297130585, + 0.13440141081809998, + 0.06375166028738022, + 0.24435466527938843, + 0.024853570386767387, + 0.08761920779943466, + 0.15593591332435608 + ], + "credit_map": [ + 0.009077274240553379, + 0.006983231753110886, + 0.008877791464328766, + 0.0078031150624156, + 0.0073087988421320915, + 0.00654194550588727, + 0.018084008246660233, + 0.020087461918592453, + 0.009130020625889301, + 0.0052436403930187225, + 0.003250215435400605, + 0.004896020982414484, + 0.005336232483386993, + 0.004027430433779955, + 0.004739290103316307, + 0.005598059855401516, + 0.004460203927010298, + 0.006060492247343063, + 0.0066499593667685986, + 0.010779913514852524, + 0.009454342536628246, + 0.005046245642006397, + 0.009748964570462704, + 0.006495278794318438, + 0.013388906605541706, + 0.015691734850406647, + 0.00778535520657897, + 0.01487177424132824, + 0.014587676152586937, + 0.006163056008517742, + 0.020908506587147713, + 0.010006608441472054, + 0.01664344035089016, + 0.024328259751200676, + 0.023108240216970444, + 0.021769369021058083, + 0.015017861500382423, + 0.010717177763581276, + 0.013514774851500988, + 0.009302286431193352, + 0.01084967516362667, + 0.007954666391015053, + 0.005466335918754339, + 0.005270160268992186, + 0.013955850154161453, + 0.010278149507939816, + 0.004883117508143187, + 0.005093705840408802, + 0.019014982506632805, + 0.0679025650024414, + 0.05130602419376373, + 0.019948730245232582, + 0.03861125558614731, + 0.020302170887589455, + 0.01434128638356924, + 0.012927651405334473, + 0.002156377537176013, + 0.0023691831156611443, + 0.0039009256288409233, + 0.00384713988751173, + 0.00430148234590888, + 0.0036131590604782104, + 0.0025156401097774506, + 0.0021496627014130354, + 0.016484476625919342, + 0.025070492178201675, + 0.007518857717514038, + 0.006781568750739098, + 0.01047931145876646, + 0.008959316648542881, + 0.006280509289354086, + 0.006044670008122921, + 0.037283290177583694, + 0.020621521398425102, + 0.014140946790575981, + 0.013299969024956226, + 0.0074445223435759544, + 0.008778837509453297, + 0.013425447978079319, + 0.040941379964351654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.020609749481081963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371123969554901 + ], + "opposite_logprob_tokens": [ + -3.4087111949920654 + ], + "logprob_reward": -0.03371123969554901, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371123969554901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08082401752471924, + 0.0462251640856266, + 0.037852462381124496, + 0.0485554113984108, + 0.05199357122182846, + 0.12365124374628067, + 0.32210773229599, + 0.021161288022994995, + 0.1313645988702774, + 0.13626451790332794 + ], + "credit_map": [ + 0.01137673668563366, + 0.009346991777420044, + 0.005046467296779156, + 0.006259493995457888, + 0.009414512664079666, + 0.008099802769720554, + 0.01323689054697752, + 0.018043117597699165, + 0.00899895653128624, + 0.006983660627156496, + 0.0031059496104717255, + 0.004459890071302652, + 0.004697814118117094, + 0.004422529134899378, + 0.006969462614506483, + 0.006586896721273661, + 0.004137846175581217, + 0.00419255206361413, + 0.005997600965201855, + 0.0049318051896989346, + 0.004554489627480507, + 0.0044070519506931305, + 0.005514364689588547, + 0.004116750322282314, + 0.004988494329154491, + 0.0070260330103337765, + 0.0020820165518671274, + 0.007604975253343582, + 0.0037593606393784285, + 0.002210983308032155, + 0.013988791033625603, + 0.0068947505205869675, + 0.005731550045311451, + 0.008119074627757072, + 0.010679955594241619, + 0.010059460997581482, + 0.005191623233258724, + 0.0034336720127612352, + 0.005000180099159479, + 0.0037780494894832373, + 0.011977842077612877, + 0.04017471894621849, + 0.006403517909348011, + 0.010253535583615303, + 0.030705133453011513, + 0.009064379148185253, + 0.007998739369213581, + 0.007073376793414354, + 0.027138980105519295, + 0.09844928979873657, + 0.06655464321374893, + 0.027116648852825165, + 0.0499800406396389, + 0.02046898379921913, + 0.01671835593879223, + 0.015680793672800064, + 0.0015427201287820935, + 0.0020272191613912582, + 0.0025485879741609097, + 0.0028051109984517097, + 0.003541812999173999, + 0.004177645314484835, + 0.0027342545799911022, + 0.0017839355859905481, + 0.016094304621219635, + 0.016576409339904785, + 0.0174117274582386, + 0.011439052410423756, + 0.012594331987202168, + 0.01746836118400097, + 0.0201056357473135, + 0.019674766808748245, + 0.029727037996053696, + 0.02008856274187565, + 0.013518070802092552, + 0.01664780266582966, + 0.010175510309636593, + 0.008652020245790482, + 0.011713995598256588, + 0.02574150450527668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.03371123969554901 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, low-angle, static shot captures the moment of brutal destruction in extreme slow motion. The bulldozer's massive, weathered yellow steel blade, deeply scratched and caked with dried mud, crashes unstoppably into the side of a parked, dark metallic red sedan. Under the glare of the midday sun, the initial impact is a silent ballet. The car's front fenders and hood wrinkle violently, the glossy red paint cracking and flaking as the underlying metal tears and folds inward like a sheet of paper under the immense, deliberate pressure. Simultaneously, the windshield and side windows burst open; spiderweb-like cracks flash across the glass before exploding outward in a flash of countless tiny fragments. Each crystalline shard hangs suspended in midair, individually catching and refracting the bright sunlight as dust and debris on the ground slowly swirl around the point of impact, creating an atmosphere of controlled chaos that is both visceral and terrifyingly beautiful.", + "seed": 24742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_5/index_9757_seed_24742_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": -0.4740244409069419, + "vlm_logprob": -0.4740244409069419, + "vlm_yes_no_margin": 1.8499999642372131, + "vlm_token_credit": -0.4740244409069419, + "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 captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.45500364899635315 + ], + "opposite_logprob_tokens": [ + -1.3300036191940308 + ], + "logprob_reward": -0.45500364899635315, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.45500364899635315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18720361590385437, + 0.14292070269584656, + 0.07410438358783722, + 0.05560416355729103, + 0.06879183650016785, + 0.05036277696490288, + 0.0683714970946312, + 0.061242226511240005, + 0.22178350389003754, + 0.06961535662412643 + ], + "credit_map": [ + 0.027417220175266266, + 0.02517598308622837, + 0.025013616308569908, + 0.03624216467142105, + 0.016322415322065353, + 0.01599552109837532, + 0.024775752797722816, + 0.01626094989478588, + 0.012041428126394749, + 0.015939081087708473, + 0.016216080635786057, + 0.03169524297118187, + 0.016481872648000717, + 0.012932192534208298, + 0.02133871242403984, + 0.016276082023978233, + 0.011763415299355984, + 0.008659331128001213, + 0.008854018524289131, + 0.008428412489593029, + 0.014212635345757008, + 0.008225551806390285, + 0.007868166081607342, + 0.006092858035117388, + 0.004272793419659138, + 0.006040508393198252, + 0.005574975162744522, + 0.00718426750972867, + 0.009118574671447277, + 0.006733966059982777, + 0.007529232185333967, + 0.009149845689535141, + 0.009894076734781265, + 0.009131531231105328, + 0.009686906822025776, + 0.011738053523004055, + 0.010313182137906551, + 0.0059681967832148075, + 0.007275228388607502, + 0.0047846594825387, + 0.006921831052750349, + 0.007010969799011946, + 0.006073070224374533, + 0.006772852502763271, + 0.007917053997516632, + 0.004859088454395533, + 0.005327325314283371, + 0.005480582360178232, + 0.00811841432005167, + 0.00966512132436037, + 0.009789548814296722, + 0.011732838116586208, + 0.008548700250685215, + 0.0056258817203342915, + 0.007757218554615974, + 0.007133772596716881, + 0.00651257811114192, + 0.007129030302166939, + 0.010844387114048004, + 0.008031868375837803, + 0.006813092157244682, + 0.007931806147098541, + 0.007775386795401573, + 0.006204076576977968, + 0.016294972971081734, + 0.011425280943512917, + 0.05185623839497566, + 0.025439603254199028, + 0.02105335146188736, + 0.016058538109064102, + 0.019844327121973038, + 0.05981118604540825, + 0.007298116106539965, + 0.006695692427456379, + 0.010900401510298252, + 0.013515726663172245, + 0.007275021634995937, + 0.009943884797394276, + 0.006734250113368034, + 0.007252258714288473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.45500364899635315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.502140998840332 + ], + "opposite_logprob_tokens": [ + -0.2521410286426544 + ], + "logprob_reward": -1.502140998840332, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502140998840332, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12176245450973511, + 0.0588303878903389, + 0.033192604780197144, + 0.07151158154010773, + 0.050318099558353424, + 0.05696582421660423, + 0.16466572880744934, + 0.052128542214632034, + 0.24311792850494385, + 0.14750683307647705 + ], + "credit_map": [ + 0.015103362500667572, + 0.012673736549913883, + 0.015653718262910843, + 0.019894979894161224, + 0.010873003862798214, + 0.009221167303621769, + 0.02325000986456871, + 0.015092484652996063, + 0.004833129234611988, + 0.0047151497565209866, + 0.005503110587596893, + 0.007023775018751621, + 0.005496383644640446, + 0.010347241535782814, + 0.011064795777201653, + 0.009846804663538933, + 0.004718024283647537, + 0.003237472614273429, + 0.0031130933202803135, + 0.002537351567298174, + 0.007846285589039326, + 0.0040633645839989185, + 0.0043679093942046165, + 0.0033091045916080475, + 0.005064430180937052, + 0.008641170337796211, + 0.0060363635420799255, + 0.008895842358469963, + 0.012236143462359905, + 0.011158316396176815, + 0.009354090318083763, + 0.010125229135155678, + 0.007180944085121155, + 0.005496159661561251, + 0.00565896974876523, + 0.007104647345840931, + 0.007014848291873932, + 0.004906000569462776, + 0.007389295846223831, + 0.005567233543843031, + 0.009307658299803734, + 0.007135051768273115, + 0.007634823676198721, + 0.007758358959108591, + 0.007442089729011059, + 0.0045686522498726845, + 0.00823470950126648, + 0.004884481430053711, + 0.0125932851806283, + 0.018551291897892952, + 0.021964432671666145, + 0.023433413356542587, + 0.03180845081806183, + 0.014174767769873142, + 0.022089915350079536, + 0.0200501699000597, + 0.0036907074972987175, + 0.0073881810531020164, + 0.008287221193313599, + 0.0075532919727265835, + 0.004700424149632454, + 0.0065133231692016125, + 0.008533027954399586, + 0.005462367553263903, + 0.02225111611187458, + 0.0171952061355114, + 0.06935349106788635, + 0.02675315923988819, + 0.0202990360558033, + 0.015936333686113358, + 0.014034955762326717, + 0.05729464814066887, + 0.01083031203597784, + 0.011738172732293606, + 0.022664831951260567, + 0.028386620804667473, + 0.016357354819774628, + 0.019844982773065567, + 0.021173907443881035, + 0.01651066727936268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -1.502140998840332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38713783025741577 + ], + "opposite_logprob_tokens": [ + -1.1371378898620605 + ], + "logprob_reward": -0.38713783025741577, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38713783025741577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07692987471818924, + 0.08545244485139847, + 0.06400620192289352, + 0.07389574497938156, + 0.1328037679195404, + 0.09241776913404465, + 0.11782531440258026, + 0.09081035852432251, + 0.18929152190685272, + 0.07656697183847427 + ], + "credit_map": [ + 0.007047855295240879, + 0.007145717274397612, + 0.010056835599243641, + 0.011158735491335392, + 0.007235834840685129, + 0.015830660238862038, + 0.010472439229488373, + 0.007981791161000729, + 0.00577658973634243, + 0.006706661079078913, + 0.009214432910084724, + 0.009551784954965115, + 0.009631942957639694, + 0.015497134067118168, + 0.014523664489388466, + 0.014550231397151947, + 0.009505156427621841, + 0.006517290137708187, + 0.006688167341053486, + 0.006729171145707369, + 0.017631009221076965, + 0.005515990778803825, + 0.0071185315027832985, + 0.004300879314541817, + 0.004762261174619198, + 0.0059310379438102245, + 0.006017213687300682, + 0.008731570094823837, + 0.014977886341512203, + 0.015684016048908234, + 0.010199603624641895, + 0.007592150010168552, + 0.01602150686085224, + 0.018142176792025566, + 0.015949763357639313, + 0.02741350047290325, + 0.020193779841065407, + 0.010296077467501163, + 0.013698729686439037, + 0.01108822226524353, + 0.008709028363227844, + 0.012254152446985245, + 0.015380660071969032, + 0.014653990045189857, + 0.012332883663475513, + 0.008581327274441719, + 0.008107448928058147, + 0.012398273684084415, + 0.009333214722573757, + 0.014436997473239899, + 0.015482017770409584, + 0.016952648758888245, + 0.02236900106072426, + 0.009306126274168491, + 0.016284413635730743, + 0.013660888187587261, + 0.007572609931230545, + 0.012788004241883755, + 0.013170876540243626, + 0.010119461454451084, + 0.011779971420764923, + 0.013656465336680412, + 0.012672731652855873, + 0.009050232358276844, + 0.022310545668005943, + 0.014103948138654232, + 0.051387250423431396, + 0.021220406517386436, + 0.01445599552243948, + 0.009392405860126019, + 0.011708201840519905, + 0.044712748378515244, + 0.006309010088443756, + 0.006304551847279072, + 0.013237345032393932, + 0.008077585138380527, + 0.01696147955954075, + 0.010969550348818302, + 0.007785266265273094, + 0.006922177504748106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.38713783025741577 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645866833627224 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.007645866833627224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645866833627224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27758505940437317, + 0.09031569212675095, + 0.08000944554805756, + 0.060027118772268295, + 0.10976174473762512, + 0.07228004932403564, + 0.13060948252677917, + 0.05687824264168739, + 0.0635257288813591, + 0.059007443487644196 + ], + "credit_map": [ + 0.03329913690686226, + 0.030603207647800446, + 0.035801149904727936, + 0.03682135045528412, + 0.02299332059919834, + 0.04745975136756897, + 0.047575242817401886, + 0.023031894117593765, + 0.006713423412293196, + 0.011527216993272305, + 0.009799383580684662, + 0.014791837893426418, + 0.00922332238405943, + 0.013247446157038212, + 0.01454827468842268, + 0.010464789345860481, + 0.007127676624804735, + 0.007803100626915693, + 0.00797047559171915, + 0.00747705390676856, + 0.01171540841460228, + 0.019121304154396057, + 0.008262722752988338, + 0.010531706735491753, + 0.0058562615886330605, + 0.005737041123211384, + 0.004848869051784277, + 0.006259813439100981, + 0.010146907530725002, + 0.012108434922993183, + 0.0063593522645533085, + 0.008710441179573536, + 0.013334177434444427, + 0.01906483992934227, + 0.01070327963680029, + 0.014450417831540108, + 0.02193574048578739, + 0.010551347397267818, + 0.01112489216029644, + 0.008597058244049549, + 0.012624218128621578, + 0.010036852210760117, + 0.012047849595546722, + 0.01025005616247654, + 0.009236943908035755, + 0.0070528294891119, + 0.0063028219155967236, + 0.004728478845208883, + 0.012862145900726318, + 0.014041190035641193, + 0.016744263470172882, + 0.01796736754477024, + 0.0229511596262455, + 0.010714022442698479, + 0.019672516733407974, + 0.015656840056180954, + 0.00589547585695982, + 0.006927859969437122, + 0.007862634025514126, + 0.007509167306125164, + 0.0047722794115543365, + 0.005799900274723768, + 0.011238258332014084, + 0.006872672121971846, + 0.006640566047281027, + 0.005055820569396019, + 0.01113592367619276, + 0.007808853406459093, + 0.006554746069014072, + 0.004652459640055895, + 0.004770089406520128, + 0.01690727286040783, + 0.00897626206278801, + 0.008249097503721714, + 0.010004057548940182, + 0.011651947163045406, + 0.005874759517610073, + 0.004991248250007629, + 0.005875053349882364, + 0.003385025542229414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.007645866833627224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193859606981277 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193859606981277, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018193859606981277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25313740968704224, + 0.050369251519441605, + 0.025205351412296295, + 0.030552441254258156, + 0.11219416558742523, + 0.03542114794254303, + 0.08229046314954758, + 0.08214372396469116, + 0.2428455948829651, + 0.08584045618772507 + ], + "credit_map": [ + 0.047115474939346313, + 0.032381847500801086, + 0.036211684346199036, + 0.0389581061899662, + 0.02264435775578022, + 0.019061066210269928, + 0.03607871010899544, + 0.02068617194890976, + 0.004581564571708441, + 0.0032609712798148394, + 0.005892091430723667, + 0.009912624955177307, + 0.005228493362665176, + 0.0060758995823562145, + 0.00905742309987545, + 0.006360182538628578, + 0.0032090716995298862, + 0.0025364498142153025, + 0.002756324363872409, + 0.002399440621957183, + 0.0028427771758288145, + 0.005175843369215727, + 0.0028671235777437687, + 0.003418323351070285, + 0.002564619993790984, + 0.0026946719735860825, + 0.0030149698723107576, + 0.00389424548484385, + 0.00524088554084301, + 0.002736912341788411, + 0.006383928935974836, + 0.004022207576781511, + 0.014711469411849976, + 0.019205374643206596, + 0.010848990641534328, + 0.015044771134853363, + 0.02002529241144657, + 0.00857019703835249, + 0.012162642553448677, + 0.011625426821410656, + 0.005784273613244295, + 0.003964610397815704, + 0.004659733735024929, + 0.004207341466099024, + 0.005143678747117519, + 0.0033310253638774157, + 0.00401578051969409, + 0.004314702935516834, + 0.008249960839748383, + 0.011460431851446629, + 0.011994791217148304, + 0.012712104246020317, + 0.011157048866152763, + 0.006568898446857929, + 0.011846546083688736, + 0.008300681598484516, + 0.009629295207560062, + 0.015514972619712353, + 0.011200394481420517, + 0.008556989952921867, + 0.006853976286947727, + 0.008756247349083424, + 0.013968882150948048, + 0.007662962190806866, + 0.03370925039052963, + 0.025762706995010376, + 0.056875694543123245, + 0.033425215631723404, + 0.025173991918563843, + 0.011538827791810036, + 0.015871839597821236, + 0.04048805683851242, + 0.006564958952367306, + 0.005723671056330204, + 0.013242886401712894, + 0.0134264612570405, + 0.013308323919773102, + 0.01314881257712841, + 0.010616926476359367, + 0.009808413684368134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.018193859606981277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, low-angle, static shot captures the moment of brutal destruction in extreme slow motion. The bulldozer's massive, weathered yellow steel blade, deeply scratched and caked with dried mud, crashes unstoppably into the side of a parked, dark metallic red sedan. Under the glare of the midday sun, the initial impact is a silent ballet. The car's front fenders and hood wrinkle violently, the glossy red paint cracking and flaking as the underlying metal tears and folds inward like a sheet of paper under the immense, deliberate pressure. Simultaneously, the windshield and side windows burst open; spiderweb-like cracks flash across the glass before exploding outward in a flash of countless tiny fragments. Each crystalline shard hangs suspended in midair, individually catching and refracting the bright sunlight as dust and debris on the ground slowly swirl around the point of impact, creating an atmosphere of controlled chaos that is both visceral and terrifyingly beautiful.", + "seed": 24743, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_5/index_9757_seed_24743_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": -0.40150677412748337, + "vlm_logprob": -0.40150677412748337, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.40150677412748337, + "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 captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22465132176876068 + ], + "opposite_logprob_tokens": [ + -2.099651336669922 + ], + "logprob_reward": -0.22465132176876068, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.22465132176876068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14100338518619537, + 0.0491211824119091, + 0.0682336837053299, + 0.3526630997657776, + 0.030219193547964096, + 0.10747750848531723, + 0.1811871975660324, + 0.029272906482219696, + 0.019862329587340355, + 0.02095949463546276 + ], + "credit_map": [ + 0.018341027200222015, + 0.031227676197886467, + 0.01944715715944767, + 0.010551191866397858, + 0.019028164446353912, + 0.01541650015860796, + 0.011250894516706467, + 0.015740767121315002, + 0.005254159681499004, + 0.014044943265616894, + 0.004881617147475481, + 0.0048065828159451485, + 0.007162509020417929, + 0.004023466259241104, + 0.0041509391739964485, + 0.004796965979039669, + 0.02034929394721985, + 0.0200734231621027, + 0.0037289082538336515, + 0.004154772963374853, + 0.00605473667383194, + 0.004188010469079018, + 0.0049362871795892715, + 0.0047482517547905445, + 0.01637575402855873, + 0.01588958315551281, + 0.007381914649158716, + 0.01273963414132595, + 0.017569484189152718, + 0.020755959674715996, + 0.1729196012020111, + 0.08903113752603531, + 0.002548511140048504, + 0.0030265124514698982, + 0.0016749280039221048, + 0.002006412483751774, + 0.0036265861708670855, + 0.005769474897533655, + 0.003979778848588467, + 0.007586990483105183, + 0.017808644101023674, + 0.008305437862873077, + 0.013000847771763802, + 0.011708439327776432, + 0.011389556340873241, + 0.011881016194820404, + 0.020039023831486702, + 0.013344542123377323, + 0.06925521045923233, + 0.02758476324379444, + 0.01619921810925007, + 0.008210507221519947, + 0.018697723746299744, + 0.024328898638486862, + 0.006299400236457586, + 0.010611476376652718, + 0.005573302507400513, + 0.0030610300600528717, + 0.0032744084019213915, + 0.0025841291062533855, + 0.0038142306730151176, + 0.0028501604683697224, + 0.0029885079711675644, + 0.005127136595547199, + 0.0038657563272863626, + 0.0021551395766437054, + 0.0017389453714713454, + 0.0012818948598578572, + 0.0029348647221922874, + 0.003530115121975541, + 0.0014933448983356357, + 0.0028622697573155165, + 0.003408011980354786, + 0.0026689779479056597, + 0.00252921343781054, + 0.001615037559531629, + 0.0029427697882056236, + 0.003209327347576618, + 0.001850277534686029, + 0.0027358769439160824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.22465132176876068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.5021119117736816 + ], + "opposite_logprob_tokens": [ + -0.25211194157600403 + ], + "logprob_reward": -1.5021119117736816, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5021119117736816, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.02628442831337452, + 0.04318712279200554, + 0.044877856969833374, + 0.05296716094017029, + 0.3317832350730896, + 0.03308464214205742, + 0.40517592430114746, + 0.035449281334877014, + 0.015724102035164833, + 0.011466286145150661 + ], + "credit_map": [ + 0.0029965443536639214, + 0.006280942820012569, + 0.002484709955751896, + 0.0017602955922484398, + 0.004814061801880598, + 0.003099390771239996, + 0.0020338476169854403, + 0.0028146354015916586, + 0.005007152445614338, + 0.011168118566274643, + 0.0041273413226008415, + 0.0034217210486531258, + 0.006464080419391394, + 0.004357683006674051, + 0.00383443059399724, + 0.004806593991816044, + 0.007237834855914116, + 0.012024843133985996, + 0.0036238390021026134, + 0.004369162954390049, + 0.005013553891330957, + 0.004090304020792246, + 0.0033834229689091444, + 0.005134893581271172, + 0.003726286580786109, + 0.003449447685852647, + 0.002411331282928586, + 0.002635339042171836, + 0.004502967465668917, + 0.00381992245092988, + 0.021468589082360268, + 0.010953276418149471, + 0.020079834386706352, + 0.01639312133193016, + 0.009312723763287067, + 0.005029302556067705, + 0.025613702833652496, + 0.12243323028087616, + 0.015350244008004665, + 0.11757107824087143, + 0.007557391654700041, + 0.0033870048355311155, + 0.0050282347947359085, + 0.0025692631024867296, + 0.0037506604567170143, + 0.0034140346106141806, + 0.0033599610906094313, + 0.004018091596662998, + 0.14980889856815338, + 0.06198926270008087, + 0.033956192433834076, + 0.017310667783021927, + 0.04497864842414856, + 0.061626072973012924, + 0.013707366771996021, + 0.021798819303512573, + 0.006868517491966486, + 0.004237918183207512, + 0.0029962966218590736, + 0.0029193127993494272, + 0.0042151762172579765, + 0.003629869781434536, + 0.005252823233604431, + 0.005329364910721779, + 0.002070765243843198, + 0.0023043956607580185, + 0.0016513318987563252, + 0.0010388623923063278, + 0.0023942841216921806, + 0.002813079860061407, + 0.0011469533201307058, + 0.0023044294212013483, + 0.0015988219529390335, + 0.0019390687812119722, + 0.0010616187937557697, + 0.0009210787247866392, + 0.001695079728960991, + 0.0015280100051313639, + 0.001003187382593751, + 0.0017194212414324284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -1.5021119117736816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25214260816574097 + ], + "opposite_logprob_tokens": [ + -1.5021426677703857 + ], + "logprob_reward": -0.25214260816574097, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25214260816574097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04931238666176796, + 0.05976058915257454, + 0.11853307485580444, + 0.025239843875169754, + 0.26673394441604614, + 0.03811151161789894, + 0.3720129430294037, + 0.03829532861709595, + 0.014216509647667408, + 0.017783893272280693 + ], + "credit_map": [ + 0.006217812187969685, + 0.011855029501020908, + 0.0047720978036522865, + 0.0035993161145597696, + 0.0065549747087061405, + 0.006883613299578428, + 0.0036623459309339523, + 0.005767198745161295, + 0.006008438300341368, + 0.013718470931053162, + 0.006250957492738962, + 0.005640963092446327, + 0.007481067441403866, + 0.006679837591946125, + 0.007021048106253147, + 0.006959802936762571, + 0.03743910789489746, + 0.04736553877592087, + 0.003741200314834714, + 0.006566074676811695, + 0.00956697203218937, + 0.0032002220395952463, + 0.00456836074590683, + 0.006085594650357962, + 0.003149735741317272, + 0.002841508947312832, + 0.001830694847740233, + 0.003080161055549979, + 0.0029680568259209394, + 0.0022898900788277388, + 0.006648831535130739, + 0.002430966356769204, + 0.017126211896538734, + 0.01427762396633625, + 0.007491603028029203, + 0.004207605961710215, + 0.020467454567551613, + 0.098526231944561, + 0.011707386933267117, + 0.09292981028556824, + 0.004333708435297012, + 0.0038017090409994125, + 0.005173469427973032, + 0.004827913362532854, + 0.003940969239920378, + 0.0037909464444965124, + 0.007427472621202469, + 0.004815319553017616, + 0.12432356178760529, + 0.05307941883802414, + 0.031313467770814896, + 0.016835087910294533, + 0.04863082990050316, + 0.06019822880625725, + 0.0142064169049263, + 0.02342594601213932, + 0.007091833278536797, + 0.004289461765438318, + 0.0030412229243665934, + 0.0029033508617430925, + 0.005351082421839237, + 0.004804793745279312, + 0.0050484174862504005, + 0.005765167530626059, + 0.0019180590752512217, + 0.0033305364195257425, + 0.0011581616709008813, + 0.0009539066813886166, + 0.0019159275107085705, + 0.0012134091230109334, + 0.001016494119539857, + 0.002710015745833516, + 0.003166249953210354, + 0.0029286318458616734, + 0.001520369085483253, + 0.0010917694307863712, + 0.0027911923825740814, + 0.002093406394124031, + 0.0016996657941490412, + 0.0024926098994910717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.25214260816574097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608537696301937 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608537696301937, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608537696301937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09061378985643387, + 0.05334634333848953, + 0.08607403188943863, + 0.13037823140621185, + 0.28752055764198303, + 0.05953086167573929, + 0.15889348089694977, + 0.10295285284519196, + 0.01114113349467516, + 0.019548749551177025 + ], + "credit_map": [ + 0.010451548732817173, + 0.023461265489459038, + 0.007434690836817026, + 0.006939907558262348, + 0.017350852489471436, + 0.009886848740279675, + 0.0066151004284620285, + 0.008473572321236134, + 0.004259075969457626, + 0.009967412799596786, + 0.007731398101896048, + 0.00843403022736311, + 0.005502732004970312, + 0.004256300628185272, + 0.006712562404572964, + 0.006482830736786127, + 0.02107037417590618, + 0.018677575513720512, + 0.005024003330618143, + 0.004586067516356707, + 0.014412891119718552, + 0.007856852374970913, + 0.006587337702512741, + 0.007858924567699432, + 0.005863565485924482, + 0.005987600423395634, + 0.0030776115600019693, + 0.005228190682828426, + 0.007233568001538515, + 0.00875675305724144, + 0.061879877001047134, + 0.0323510617017746, + 0.018405888229608536, + 0.014142831787467003, + 0.00882451981306076, + 0.004905425477772951, + 0.02192952297627926, + 0.1051541268825531, + 0.013206219300627708, + 0.10095202177762985, + 0.010823635384440422, + 0.005299365147948265, + 0.006344722583889961, + 0.006393886171281338, + 0.006705387961119413, + 0.006064745131880045, + 0.009942390955984592, + 0.007956724613904953, + 0.06413088738918304, + 0.02526683174073696, + 0.016092810779809952, + 0.007868636399507523, + 0.013348225504159927, + 0.017895512282848358, + 0.006254158914089203, + 0.008036402985453606, + 0.023003149777650833, + 0.010979030281305313, + 0.01294455025345087, + 0.008676833473145962, + 0.012768522836267948, + 0.009216628968715668, + 0.012154742144048214, + 0.01320938766002655, + 0.0020970480982214212, + 0.0013705729506909847, + 0.0010879052570089698, + 0.000809754419606179, + 0.0016059200279414654, + 0.0014267684891819954, + 0.0009785167640075088, + 0.0017646463820710778, + 0.0031668327283114195, + 0.0026229573413729668, + 0.001704417634755373, + 0.0011208945652469993, + 0.0033380540553480387, + 0.002836544532328844, + 0.0019453895511105657, + 0.002813659142702818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.01608537696301937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254265196621418 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.01254265196621418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254265196621418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027876082807779312, + 0.03430067375302315, + 0.10356247425079346, + 0.11634945869445801, + 0.38270869851112366, + 0.1293853223323822, + 0.1645001769065857, + 0.021279800683259964, + 0.0057745338417589664, + 0.014262743294239044 + ], + "credit_map": [ + 0.0027552759274840355, + 0.006164185237139463, + 0.0022667613811790943, + 0.0022951215505599976, + 0.00438527250662446, + 0.004515381995588541, + 0.0021449215710163116, + 0.003349162871018052, + 0.003790697082877159, + 0.008760801516473293, + 0.0024710760917514563, + 0.0027055947575718164, + 0.005618710536509752, + 0.00422753905877471, + 0.002901413245126605, + 0.0038248407654464245, + 0.03421661630272865, + 0.04080478847026825, + 0.0034902039915323257, + 0.004946119152009487, + 0.008291025646030903, + 0.002522999420762062, + 0.004356177058070898, + 0.004934546537697315, + 0.004900636617094278, + 0.004825621377676725, + 0.002100870246067643, + 0.004294971004128456, + 0.006435261107981205, + 0.007922807708382607, + 0.05591502785682678, + 0.029954275116324425, + 0.024472873657941818, + 0.01859678514301777, + 0.010883140377700329, + 0.006233444903045893, + 0.028895171359181404, + 0.1414441466331482, + 0.017170721665024757, + 0.13501238822937012, + 0.023097023367881775, + 0.011512148194015026, + 0.017616409808397293, + 0.01325750071555376, + 0.014454007148742676, + 0.0121618015691638, + 0.02271834947168827, + 0.01456807367503643, + 0.06174592673778534, + 0.025250637903809547, + 0.01406803634017706, + 0.007093858439475298, + 0.01751920022070408, + 0.024088837206363678, + 0.005970911122858524, + 0.008762785233557224, + 0.0066919876262545586, + 0.0025555796455591917, + 0.002346759894862771, + 0.0013376701390370727, + 0.002449501072987914, + 0.0020744691137224436, + 0.0015748818404972553, + 0.002248951233923435, + 0.001114342361688614, + 0.0006927719223313034, + 0.0007548880530521274, + 0.0004073650052305311, + 0.0009319668752141297, + 0.0006493523833341897, + 0.0004232756036799401, + 0.0008005712879821658, + 0.0028656062204390764, + 0.0019629918970167637, + 0.0014790395507588983, + 0.0007910042768344283, + 0.0018929672660306096, + 0.0020244289189577103, + 0.0011064368300139904, + 0.002140267752110958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.01254265196621418 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, low-angle, static shot captures the moment of brutal destruction in extreme slow motion. The bulldozer's massive, weathered yellow steel blade, deeply scratched and caked with dried mud, crashes unstoppably into the side of a parked, dark metallic red sedan. Under the glare of the midday sun, the initial impact is a silent ballet. The car's front fenders and hood wrinkle violently, the glossy red paint cracking and flaking as the underlying metal tears and folds inward like a sheet of paper under the immense, deliberate pressure. Simultaneously, the windshield and side windows burst open; spiderweb-like cracks flash across the glass before exploding outward in a flash of countless tiny fragments. Each crystalline shard hangs suspended in midair, individually catching and refracting the bright sunlight as dust and debris on the ground slowly swirl around the point of impact, creating an atmosphere of controlled chaos that is both visceral and terrifyingly beautiful.", + "seed": 24744, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_5/index_9757_seed_24744_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": -0.42056893594563005, + "vlm_logprob": -0.42056893594563005, + "vlm_yes_no_margin": 1.7499999523162841, + "vlm_token_credit": -0.42056893594563005, + "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 captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.30630797147750854 + ], + "opposite_logprob_tokens": [ + -1.5563080310821533 + ], + "logprob_reward": -0.30630797147750854, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.30630797147750854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17453590035438538, + 0.13603276014328003, + 0.16466835141181946, + 0.04856795072555542, + 0.17729498445987701, + 0.02072685956954956, + 0.06773009896278381, + 0.0747571587562561, + 0.06021377444267273, + 0.0754721537232399 + ], + "credit_map": [ + 0.02617502026259899, + 0.02045263536274433, + 0.01544337347149849, + 0.019582480192184448, + 0.02303212136030197, + 0.019043803215026855, + 0.028521176427602768, + 0.022285297513008118, + 0.012400691397488117, + 0.009352127090096474, + 0.017239851877093315, + 0.01576751098036766, + 0.019576333463191986, + 0.011461440473794937, + 0.030143683776259422, + 0.020091131329536438, + 0.01506641786545515, + 0.03955211862921715, + 0.04385434836149216, + 0.023814063519239426, + 0.013658033683896065, + 0.008734556846320629, + 0.008627633564174175, + 0.011361170560121536, + 0.004019052721560001, + 0.009609034284949303, + 0.005090082995593548, + 0.007164817303419113, + 0.006355012767016888, + 0.004785456694662571, + 0.005128875374794006, + 0.006415616255253553, + 0.04614867642521858, + 0.012278500013053417, + 0.05657576024532318, + 0.015910372138023376, + 0.012325508520007133, + 0.00932279508560896, + 0.010852535255253315, + 0.013880838640034199, + 0.0016962175723165274, + 0.0018889236962422729, + 0.005754874087870121, + 0.0028981228824704885, + 0.0021741425152868032, + 0.0024820624385029078, + 0.0015114244306460023, + 0.002321091014891863, + 0.007132931612432003, + 0.005380137823522091, + 0.01797819323837757, + 0.011248716153204441, + 0.004242153838276863, + 0.00486320024356246, + 0.008823443204164505, + 0.008061320520937443, + 0.006268148776143789, + 0.0069811707362532616, + 0.005920928902924061, + 0.012147015891969204, + 0.00709743844345212, + 0.006201380863785744, + 0.018319500610232353, + 0.011821574531495571, + 0.0039258380420506, + 0.0027423326391726732, + 0.006002924405038357, + 0.009914996102452278, + 0.003174379700794816, + 0.005944754928350449, + 0.011829220689833164, + 0.01667932979762554, + 0.005817546043545008, + 0.005082384217530489, + 0.021029694005846977, + 0.012990332208573818, + 0.006907431874424219, + 0.0054590399377048016, + 0.00994949508458376, + 0.00823623314499855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.30630797147750854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.406346082687378 + ], + "opposite_logprob_tokens": [ + -0.28134608268737793 + ], + "logprob_reward": -1.406346082687378, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.406346082687378, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2863404452800751, + 0.23072759807109833, + 0.09343462437391281, + 0.053848426789045334, + 0.06458969414234161, + 0.04008480906486511, + 0.05572099611163139, + 0.07371941953897476, + 0.06579660624265671, + 0.03573736920952797 + ], + "credit_map": [ + 0.04362200200557709, + 0.030598709359765053, + 0.035429518669843674, + 0.03251682221889496, + 0.031453948467969894, + 0.029586097225546837, + 0.04856157675385475, + 0.034571755677461624, + 0.021178634837269783, + 0.026731567457318306, + 0.021580040454864502, + 0.03046373650431633, + 0.04369566962122917, + 0.015821808949112892, + 0.0499122180044651, + 0.021343933418393135, + 0.011959186755120754, + 0.014425954781472683, + 0.01224943995475769, + 0.014862893149256706, + 0.011300001293420792, + 0.008525736629962921, + 0.009981545619666576, + 0.010129866190254688, + 0.005213708616793156, + 0.012177366763353348, + 0.0058698286302387714, + 0.009977760724723339, + 0.004528624936938286, + 0.0034501811023801565, + 0.005113884806632996, + 0.007517071440815926, + 0.016135744750499725, + 0.005445696413516998, + 0.017788492143154144, + 0.006122527178376913, + 0.004962657578289509, + 0.0038878361228853464, + 0.004094835836440325, + 0.006151898298412561, + 0.007651225198060274, + 0.0033104305621236563, + 0.004952490329742432, + 0.011097997426986694, + 0.002481556963175535, + 0.0020574370864778757, + 0.004914914723485708, + 0.003618754679337144, + 0.007567842025309801, + 0.0038814847357571125, + 0.008928430266678333, + 0.009226147085428238, + 0.005332580767571926, + 0.004961282014846802, + 0.008023449219763279, + 0.007799777667969465, + 0.005310024134814739, + 0.004533507861196995, + 0.023648150265216827, + 0.012470684014260769, + 0.005492617376148701, + 0.00674800667911768, + 0.009159036912024021, + 0.0063573927618563175, + 0.006169414147734642, + 0.004257870372384787, + 0.009525856003165245, + 0.009746106341481209, + 0.005151716060936451, + 0.006114862393587828, + 0.011401903815567493, + 0.013428875245153904, + 0.004500338342040777, + 0.0030845156870782375, + 0.003300976939499378, + 0.0052777789533138275, + 0.004641067236661911, + 0.004229711834341288, + 0.0051848492585122585, + 0.005518133286386728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -1.406346082687378 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486787974834442 + ], + "opposite_logprob_tokens": [ + -1.2236788272857666 + ], + "logprob_reward": -0.3486787974834442, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486787974834442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08517692983150482, + 0.08951850235462189, + 0.12616001069545746, + 0.08856045454740524, + 0.07807787507772446, + 0.10699858516454697, + 0.20559807121753693, + 0.045551080256700516, + 0.12060078233480453, + 0.05375770106911659 + ], + "credit_map": [ + 0.008172119036316872, + 0.008542504161596298, + 0.015195992775261402, + 0.012607828713953495, + 0.007517569698393345, + 0.008107003755867481, + 0.016047395765781403, + 0.008986522443592548, + 0.008952165953814983, + 0.009529473260045052, + 0.010888839140534401, + 0.009983157739043236, + 0.010691286996006966, + 0.011798174120485783, + 0.01599721610546112, + 0.011678190901875496, + 0.016694871708750725, + 0.02950936183333397, + 0.02341882698237896, + 0.01894822157919407, + 0.010206270962953568, + 0.007622926961630583, + 0.00901641696691513, + 0.010743110440671444, + 0.00847864430397749, + 0.025777151808142662, + 0.010222088545560837, + 0.008130859583616257, + 0.009609067812561989, + 0.007792327553033829, + 0.00782268587499857, + 0.010727624408900738, + 0.01770520769059658, + 0.006249774247407913, + 0.019873181357979774, + 0.007424808107316494, + 0.007106889504939318, + 0.005515472497791052, + 0.005155382677912712, + 0.009047161787748337, + 0.022936036810278893, + 0.007393036503344774, + 0.011617718264460564, + 0.034055016934871674, + 0.004631431773304939, + 0.004219430964440107, + 0.011772152036428452, + 0.010373757220804691, + 0.014156529679894447, + 0.012157732620835304, + 0.07097155600786209, + 0.037935785949230194, + 0.013562874868512154, + 0.015467499382793903, + 0.021479548886418343, + 0.019866541028022766, + 0.005811445880681276, + 0.002991892397403717, + 0.009821612387895584, + 0.007964808493852615, + 0.003860287833958864, + 0.004609793424606323, + 0.004976197145879269, + 0.005515041761100292, + 0.008837657049298286, + 0.006863722577691078, + 0.016050809994339943, + 0.019292205572128296, + 0.007501507177948952, + 0.010387424379587173, + 0.023106826469302177, + 0.028560634702444077, + 0.006634599529206753, + 0.0038139065727591515, + 0.009583542123436928, + 0.008325725793838501, + 0.005229502450674772, + 0.004712044261395931, + 0.008884290233254433, + 0.006574095226824284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.3486787974834442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209662288427353 + ], + "opposite_logprob_tokens": [ + -4.018209457397461 + ], + "logprob_reward": -0.018209662288427353, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018209662288427353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23480840027332306, + 0.1393738090991974, + 0.06958811730146408, + 0.05398700758814812, + 0.08398400247097015, + 0.10123070329427719, + 0.1020190566778183, + 0.07223080098628998, + 0.054921478033065796, + 0.08785663545131683 + ], + "credit_map": [ + 0.028519777581095695, + 0.027327200397849083, + 0.031289368867874146, + 0.02646712400019169, + 0.020754758268594742, + 0.029002342373132706, + 0.03814585506916046, + 0.03330196440219879, + 0.017035948112607002, + 0.01774713769555092, + 0.013272310607135296, + 0.01817421428859234, + 0.01646420545876026, + 0.010580720379948616, + 0.03414721414446831, + 0.011952055618166924, + 0.006402270402759314, + 0.012301267124712467, + 0.015710078179836273, + 0.009584694169461727, + 0.008955343626439571, + 0.006171097047626972, + 0.004614477511495352, + 0.005848890170454979, + 0.004262768663465977, + 0.009117409586906433, + 0.005377717316150665, + 0.00878579169511795, + 0.007160851266235113, + 0.004942275583744049, + 0.005492381285876036, + 0.008847814053297043, + 0.010140514932572842, + 0.008374000899493694, + 0.013855624943971634, + 0.011253214441239834, + 0.00798739306628704, + 0.009373249486088753, + 0.009220595471560955, + 0.013779408298432827, + 0.02026919089257717, + 0.005689734127372503, + 0.009724424220621586, + 0.033384017646312714, + 0.004129748325794935, + 0.0042883045971393585, + 0.011897295713424683, + 0.011847988702356815, + 0.005920412018895149, + 0.00615472299978137, + 0.037615370005369186, + 0.017844703048467636, + 0.007028545718640089, + 0.00870538130402565, + 0.010341571643948555, + 0.00840835366398096, + 0.006486250087618828, + 0.0038138986565172672, + 0.02428712323307991, + 0.01251962874084711, + 0.004179645329713821, + 0.005313956178724766, + 0.009132888168096542, + 0.006497413385659456, + 0.00441757682710886, + 0.00273875636048615, + 0.006366644985973835, + 0.008859348483383656, + 0.0030716510955244303, + 0.005621107295155525, + 0.009717101231217384, + 0.014129286631941795, + 0.00664684409275651, + 0.005024911370128393, + 0.02676009014248848, + 0.01590060256421566, + 0.008306784555315971, + 0.006680547259747982, + 0.01150804478675127, + 0.007028811611235142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.018209662288427353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the action depicted in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a weathered yellow steel bulldozer blade crash into a dark metallic red sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are glass shards from the car's windows suspended in midair catching sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302165791392326 + ], + "opposite_logprob_tokens": [ + -3.7733020782470703 + ], + "logprob_reward": -0.023302165791392326, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302165791392326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1750611960887909, + 0.09383686631917953, + 0.09031371772289276, + 0.06987465173006058, + 0.1464259922504425, + 0.11246845871210098, + 0.04001373052597046, + 0.07437128573656082, + 0.11734134703874588, + 0.08029279112815857 + ], + "credit_map": [ + 0.02612273022532463, + 0.02014346234500408, + 0.017893051728606224, + 0.021312402561306953, + 0.015259931795299053, + 0.022542690858244896, + 0.030799750238656998, + 0.020987166091799736, + 0.011007186025381088, + 0.016983825713396072, + 0.007026731967926025, + 0.0106387073174119, + 0.013504774309694767, + 0.006428861990571022, + 0.01722494699060917, + 0.01102183572947979, + 0.013134089298546314, + 0.0105452174320817, + 0.01181850302964449, + 0.017859214916825294, + 0.013341293670237064, + 0.0077967713586986065, + 0.00889195129275322, + 0.006926674861460924, + 0.0062424186617136, + 0.012878017500042915, + 0.007199856918305159, + 0.010960628278553486, + 0.005542424041777849, + 0.007673760876059532, + 0.007519505452364683, + 0.011858043260872364, + 0.03099382109940052, + 0.013858004473149776, + 0.03549548238515854, + 0.015047932974994183, + 0.015751196071505547, + 0.008902865462005138, + 0.014314498752355576, + 0.012062191031873226, + 0.014354393817484379, + 0.0073568313382565975, + 0.023222066462039948, + 0.02021900750696659, + 0.009320366196334362, + 0.008043087087571621, + 0.011674519628286362, + 0.018278198316693306, + 0.005507706664502621, + 0.0047492366284132, + 0.0037811065558344126, + 0.005682671908289194, + 0.0038246256299316883, + 0.0053009409457445145, + 0.005592488218098879, + 0.0055749560706317425, + 0.008635908365249634, + 0.004730156622827053, + 0.008412828668951988, + 0.012980196624994278, + 0.008710987865924835, + 0.0076897176913917065, + 0.011069378815591335, + 0.012142110615968704, + 0.008358677849173546, + 0.006638186983764172, + 0.014026626944541931, + 0.017071427777409554, + 0.012643213383853436, + 0.01242764387279749, + 0.020328423008322716, + 0.025847142562270164, + 0.010378998704254627, + 0.006458289455622435, + 0.009684574790298939, + 0.012465499341487885, + 0.013535748235881329, + 0.007281109225004911, + 0.012316946871578693, + 0.00817162450402975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.023302165791392326 + } + ], + "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/0000058/group_6/index_tensor([1185, 1185])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_tensor([1185, 1185])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4e158e930441b5c959b901faa4497b1fde21b473 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_tensor([1185, 1185])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, medium shot, shot from an upward angle, is imbued with the ethos of a martial arts film. In the frame, a warrior clad in pristine, dark iron armor is at the peak of his power. He maintains a firm horse stance, his center of gravity extremely low, his muscles bulging. His weathered armor, composed of plates of varying sizes strung together with leather cords, is covered in scratches and battle scars. The low, golden sunlight of dusk reflects a deep, metallic sheen. His long, black hair dances wildly back in the gust of wind, a few strands brushing against his stern, focused face, eyes wide and teeth clenched. With a silent roar, he thrusts his right fist forward with the force of thunder, the air around it seemingly warping slightly with the immense energy. His fist struck a massive bluestone stele before him, taller than a man, with precise and powerful force. At the moment of contact, a spiderweb of cracks centered on the center of his fist erupted across its surface. Then, with a resounding, muffled thud, the entire stele exploded, sending countless boulders and dust flying in all directions, clearly visible in the backlight. The entire scene took place on a desolate mountaintop, against a backdrop of sparse, withered trees and a sky stained orange-red by the sunset glow. The scene created an epic atmosphere of aloofness, power, and a destructive aesthetic.", + "seed": 24313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_1185_seed_24313_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": -0.08156953752040863, + "vlm_logprob": -0.08156953752040863, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.08156953752040863, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10457479953765869 + ], + "opposite_logprob_tokens": [ + -2.479574680328369 + ], + "logprob_reward": -0.10457479953765869, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10457479953765869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13063037395477295, + 0.03371933475136757, + 0.05884680151939392, + 0.1338842511177063, + 0.13118870556354523, + 0.1344754695892334, + 0.049233462661504745, + 0.10655757039785385, + 0.1351468563079834, + 0.08631710708141327 + ], + "credit_map": [ + 0.01764862611889839, + 0.020237164571881294, + 0.015557941049337387, + 0.018345726653933525, + 0.011119317263364792, + 0.014018029905855656, + 0.014163079671561718, + 0.01954049989581108, + 0.002856883918866515, + 0.00655187526717782, + 0.002949050161987543, + 0.0047552273608744144, + 0.0020090839825570583, + 0.0059456489980220795, + 0.0027588277589529753, + 0.005892736371606588, + 0.007525998633354902, + 0.008095079101622105, + 0.00449746660888195, + 0.004698700271546841, + 0.006723330821841955, + 0.008488689549267292, + 0.005412858445197344, + 0.013404673896729946, + 0.021913358941674232, + 0.022019565105438232, + 0.011569155380129814, + 0.01614333875477314, + 0.009932444430887699, + 0.020780235528945923, + 0.01156094204634428, + 0.019965218380093575, + 0.021858852356672287, + 0.019348274916410446, + 0.012844488956034184, + 0.013660162687301636, + 0.015289303846657276, + 0.019685819745063782, + 0.010949643328785896, + 0.017552150413393974, + 0.01164405606687069, + 0.020557096228003502, + 0.01139098685234785, + 0.012775483541190624, + 0.01186749991029501, + 0.03041529841721058, + 0.01002517156302929, + 0.02579987607896328, + 0.005996494088321924, + 0.00648787384852767, + 0.0067650373093783855, + 0.005072882864624262, + 0.006092962343245745, + 0.0073918541893363, + 0.004653375130146742, + 0.006772983819246292, + 0.009147622622549534, + 0.006691811140626669, + 0.006005033850669861, + 0.03226127102971077, + 0.006426273379474878, + 0.0077755446545779705, + 0.008017383515834808, + 0.03023262321949005, + 0.012875622138381004, + 0.021712787449359894, + 0.01687515899538994, + 0.021661709994077682, + 0.01643410697579384, + 0.011841841042041779, + 0.018274186179041862, + 0.015471449121832848, + 0.010812484659254551, + 0.014140618033707142, + 0.00727171590551734, + 0.012182806618511677, + 0.013567158952355385, + 0.009068922139704227, + 0.010961195454001427, + 0.008312201127409935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.10457479953765869 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903824001550674 + ], + "opposite_logprob_tokens": [ + -2.579038143157959 + ], + "logprob_reward": -0.07903824001550674, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903824001550674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0798472985625267, + 0.09119243174791336, + 0.22468656301498413, + 0.15413916110992432, + 0.06050582230091095, + 0.05855061486363411, + 0.061659086495637894, + 0.0981803834438324, + 0.07634460180997849, + 0.09489405155181885 + ], + "credit_map": [ + 0.011515727266669273, + 0.015026206150650978, + 0.01154724694788456, + 0.012652807869017124, + 0.006344009190797806, + 0.008586736395955086, + 0.005630163475871086, + 0.008544400334358215, + 0.006312164943665266, + 0.014296209439635277, + 0.011747871525585651, + 0.013761907815933228, + 0.006950440816581249, + 0.01010888535529375, + 0.015033351257443428, + 0.012981603853404522, + 0.02191806398332119, + 0.03401559963822365, + 0.036802247166633606, + 0.050747666507959366, + 0.019450388848781586, + 0.01972169429063797, + 0.019951779395341873, + 0.022079110145568848, + 0.013278179802000523, + 0.030272986739873886, + 0.01652107760310173, + 0.027531974017620087, + 0.01623232662677765, + 0.019037218764424324, + 0.014707370661199093, + 0.016558021306991577, + 0.009746394120156765, + 0.009702258743345737, + 0.007512723561376333, + 0.005816000513732433, + 0.007866105064749718, + 0.006337181199342012, + 0.006520798895508051, + 0.007004357874393463, + 0.006565581075847149, + 0.0074525512754917145, + 0.005320427007973194, + 0.007010240573436022, + 0.00809653289616108, + 0.008644556626677513, + 0.005130210425704718, + 0.010330514051020145, + 0.006873437203466892, + 0.008153491653501987, + 0.00718535203486681, + 0.0068045733496546745, + 0.009757429361343384, + 0.0070432499051094055, + 0.008567878045141697, + 0.007273677736520767, + 0.006694458425045013, + 0.005386792588979006, + 0.004337145481258631, + 0.03324782848358154, + 0.003408083226531744, + 0.00582079915329814, + 0.0077543617226183414, + 0.03153091296553612, + 0.009947145357728004, + 0.008695748634636402, + 0.01105751097202301, + 0.010875474661588669, + 0.01020674966275692, + 0.0064055402763187885, + 0.010179500095546246, + 0.008976933546364307, + 0.014031083323061466, + 0.01650378480553627, + 0.009193946607410908, + 0.01309671625494957, + 0.01871190406382084, + 0.0063655562698841095, + 0.010682683438062668, + 0.006308381911367178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.07903824001550674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705519795417786 + ], + "opposite_logprob_tokens": [ + -2.1270551681518555 + ], + "logprob_reward": -0.12705519795417786, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705519795417786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07477208971977234, + 0.08191356062889099, + 0.10576380789279938, + 0.2487841695547104, + 0.06442039459943771, + 0.10778890550136566, + 0.09040414541959763, + 0.10104162245988846, + 0.05084211006760597, + 0.07426922768354416 + ], + "credit_map": [ + 0.010829688049852848, + 0.01255242619663477, + 0.010354439727962017, + 0.011003492400050163, + 0.006018423475325108, + 0.008392393589019775, + 0.0055344621650874615, + 0.010086758993566036, + 0.006915389560163021, + 0.021887658163905144, + 0.00819537416100502, + 0.01617356948554516, + 0.00529471505433321, + 0.007041345350444317, + 0.007772020064294338, + 0.008633491583168507, + 0.01064242236316204, + 0.015447065234184265, + 0.01128220371901989, + 0.01177075132727623, + 0.014758531004190445, + 0.012514876201748848, + 0.010162885300815105, + 0.019185079261660576, + 0.030846545472741127, + 0.04103049263358116, + 0.02474028803408146, + 0.05737198144197464, + 0.027439989149570465, + 0.023139188066124916, + 0.023568879812955856, + 0.020646803081035614, + 0.010145774111151695, + 0.007476001046597958, + 0.006034689489752054, + 0.010757572017610073, + 0.005303187761455774, + 0.010519293136894703, + 0.0065674749203026295, + 0.007616402581334114, + 0.0109207509085536, + 0.020062588155269623, + 0.01143006980419159, + 0.015543080866336823, + 0.01365000568330288, + 0.016050264239311218, + 0.008150802925229073, + 0.011981344781816006, + 0.0067660072818398476, + 0.008766061626374722, + 0.006952123716473579, + 0.008675933815538883, + 0.010787513107061386, + 0.023304853588342667, + 0.007908114232122898, + 0.01724354736506939, + 0.011510043404996395, + 0.0063468581065535545, + 0.006644092500209808, + 0.026169825345277786, + 0.008380228653550148, + 0.006689210422337055, + 0.008402192033827305, + 0.026899170130491257, + 0.006801557261496782, + 0.007552356459200382, + 0.006402920000255108, + 0.008825408294796944, + 0.006911866366863251, + 0.0028244906570762396, + 0.007114943116903305, + 0.004408564418554306, + 0.014587522484362125, + 0.011624443344771862, + 0.007692660205066204, + 0.009175819344818592, + 0.015831701457500458, + 0.004062917549163103, + 0.007310665212571621, + 0.003983501810580492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.12705519795417786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818835735321045 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818835735321045, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818835735321045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10038137435913086, + 0.026173651218414307, + 0.07778774201869965, + 0.036802880465984344, + 0.03770805522799492, + 0.0556352436542511, + 0.20091886818408966, + 0.19426396489143372, + 0.18037423491477966, + 0.0899539515376091 + ], + "credit_map": [ + 0.015956122428178787, + 0.008983731269836426, + 0.009653422050178051, + 0.009313417598605156, + 0.011112013831734657, + 0.019106054678559303, + 0.00631463760510087, + 0.01994198001921177, + 0.0033926651813089848, + 0.004369846545159817, + 0.0038418376352638006, + 0.003204017411917448, + 0.001720163389109075, + 0.0027412015479058027, + 0.00295140384696424, + 0.003952515311539173, + 0.00770181929692626, + 0.014953280799090862, + 0.006990489549934864, + 0.008757245726883411, + 0.00963380467146635, + 0.010681877844035625, + 0.006093387026339769, + 0.012975835241377354, + 0.003869864856824279, + 0.004503689240664244, + 0.003410676959902048, + 0.003237084485590458, + 0.0053040627390146255, + 0.004578772466629744, + 0.004447780549526215, + 0.007450951263308525, + 0.005260789301246405, + 0.004555831663310528, + 0.0042901779524981976, + 0.006903895642608404, + 0.0039449832402169704, + 0.004745969548821449, + 0.004639980848878622, + 0.003366428893059492, + 0.0052532777190208435, + 0.01212841272354126, + 0.006864367052912712, + 0.006318607367575169, + 0.005903351120650768, + 0.006204587407410145, + 0.004690052475780249, + 0.008272584527730942, + 0.0172240249812603, + 0.023098396137356758, + 0.019903164356946945, + 0.024518942460417747, + 0.0247220229357481, + 0.03623752295970917, + 0.025493083521723747, + 0.029721707105636597, + 0.014166812412440777, + 0.012387576512992382, + 0.010502775199711323, + 0.07212695479393005, + 0.008125490508973598, + 0.010055570863187313, + 0.013829177245497704, + 0.05306960269808769, + 0.020617710426449776, + 0.027203358709812164, + 0.028131062164902687, + 0.0290092583745718, + 0.021336644887924194, + 0.011684749275445938, + 0.022615965455770493, + 0.01977548561990261, + 0.01252974197268486, + 0.01464846171438694, + 0.009902018122375011, + 0.011120876297354698, + 0.016536740586161613, + 0.00785834714770317, + 0.010626845061779022, + 0.006730919238179922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.01818835735321045 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899109274148941 + ], + "opposite_logprob_tokens": [ + -2.578991174697876 + ], + "logprob_reward": -0.07899109274148941, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899109274148941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10300343483686447, + 0.03149937465786934, + 0.06415959447622299, + 0.02872362732887268, + 0.04479765519499779, + 0.06650544703006744, + 0.04511282220482826, + 0.09655489027500153, + 0.30173927545547485, + 0.21790388226509094 + ], + "credit_map": [ + 0.01718078926205635, + 0.008630897849798203, + 0.011350029148161411, + 0.011666822247207165, + 0.010832270607352257, + 0.01749926060438156, + 0.008712850511074066, + 0.017130518332123756, + 0.0030939073767513037, + 0.0069971210323274136, + 0.00449149776250124, + 0.004222107585519552, + 0.0029450790025293827, + 0.002864264417439699, + 0.0035927302669733763, + 0.003292666980996728, + 0.006926996633410454, + 0.013094090856611729, + 0.006265420001000166, + 0.007049962878227234, + 0.009428350254893303, + 0.007828010246157646, + 0.004646608140319586, + 0.008920153602957726, + 0.0029962630942463875, + 0.0056507461704313755, + 0.00299932062625885, + 0.0036704058293253183, + 0.002876646351069212, + 0.0033913152292370796, + 0.002764860400930047, + 0.004374070093035698, + 0.006123822648078203, + 0.005044951569288969, + 0.004523868206888437, + 0.008059514686465263, + 0.004749662708491087, + 0.004899891559034586, + 0.005224136635661125, + 0.006171806715428829, + 0.0074530476704239845, + 0.008249622769653797, + 0.010074320249259472, + 0.008165187202394009, + 0.006640438921749592, + 0.011126593686640263, + 0.005506413523107767, + 0.009289818815886974, + 0.00599625101312995, + 0.003731265664100647, + 0.005851048044860363, + 0.004910165444016457, + 0.003410440869629383, + 0.010344884358346462, + 0.0036989678628742695, + 0.007169800810515881, + 0.007343027740716934, + 0.010458032600581646, + 0.006206080317497253, + 0.021680118516087532, + 0.007575865834951401, + 0.007288896478712559, + 0.011885358020663261, + 0.02411750890314579, + 0.033379506319761276, + 0.042725518345832825, + 0.04684916138648987, + 0.04845722019672394, + 0.037722472101449966, + 0.021348625421524048, + 0.03709401190280914, + 0.03416276350617409, + 0.03214476257562637, + 0.04379720613360405, + 0.027723943814635277, + 0.028957346454262733, + 0.040455661714076996, + 0.009362751618027687, + 0.023198289796710014, + 0.012263932265341282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.07899109274148941 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot, shot from an upward angle, is imbued with the ethos of a martial arts film. In the frame, a warrior clad in pristine, dark iron armor is at the peak of his power. He maintains a firm horse stance, his center of gravity extremely low, his muscles bulging. His weathered armor, composed of plates of varying sizes strung together with leather cords, is covered in scratches and battle scars. The low, golden sunlight of dusk reflects a deep, metallic sheen. His long, black hair dances wildly back in the gust of wind, a few strands brushing against his stern, focused face, eyes wide and teeth clenched. With a silent roar, he thrusts his right fist forward with the force of thunder, the air around it seemingly warping slightly with the immense energy. His fist struck a massive bluestone stele before him, taller than a man, with precise and powerful force. At the moment of contact, a spiderweb of cracks centered on the center of his fist erupted across its surface. Then, with a resounding, muffled thud, the entire stele exploded, sending countless boulders and dust flying in all directions, clearly visible in the backlight. The entire scene took place on a desolate mountaintop, against a backdrop of sparse, withered trees and a sky stained orange-red by the sunset glow. The scene created an epic atmosphere of aloofness, power, and a destructive aesthetic.", + "seed": 24314, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_1185_seed_24314_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": -0.0939678367227316, + "vlm_logprob": -0.0939678367227316, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.0939678367227316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06580857187509537 + ], + "opposite_logprob_tokens": [ + -2.9408085346221924 + ], + "logprob_reward": -0.06580857187509537, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06580857187509537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3456672728061676, + 0.060327041894197464, + 0.0707205981016159, + 0.1363421529531479, + 0.09951663762331009, + 0.045158617198467255, + 0.03348666802048683, + 0.03562748804688454, + 0.1011907160282135, + 0.07196272164583206 + ], + "credit_map": [ + 0.0933101698756218, + 0.013528651557862759, + 0.12711456418037415, + 0.02587236277759075, + 0.020761912688612938, + 0.016120227053761482, + 0.03455434367060661, + 0.01440510619431734, + 0.008730659261345863, + 0.005061453673988581, + 0.01002211682498455, + 0.0047841728664934635, + 0.006054725963622332, + 0.009085559286177158, + 0.0071152132004499435, + 0.009473154321312904, + 0.00877290591597557, + 0.010843207128345966, + 0.007049389649182558, + 0.010648736730217934, + 0.008813580498099327, + 0.0070342691615223885, + 0.010266145691275597, + 0.007292376831173897, + 0.016832273453474045, + 0.01634523645043373, + 0.019763199612498283, + 0.017238611355423927, + 0.013519161380827427, + 0.010841608047485352, + 0.024917427450418472, + 0.016884654760360718, + 0.01509999018162489, + 0.009709564037621021, + 0.021090378984808922, + 0.012265444733202457, + 0.011384019628167152, + 0.008132273331284523, + 0.012434165924787521, + 0.009400812909007072, + 0.005076149944216013, + 0.002845269162207842, + 0.011629618704319, + 0.00732778012752533, + 0.005034096539020538, + 0.003177499631419778, + 0.005076354369521141, + 0.004991855937987566, + 0.005480481777340174, + 0.003790349932387471, + 0.00457525672391057, + 0.0035990953911095858, + 0.0031886156648397446, + 0.006425112951546907, + 0.0038508446887135506, + 0.0025769169442355633, + 0.0041139982640743256, + 0.005395655985921621, + 0.003501659957692027, + 0.005111358128488064, + 0.005043169017881155, + 0.004779090639203787, + 0.004898399114608765, + 0.0027841643895953894, + 0.01287062931805849, + 0.012116659432649612, + 0.014316885732114315, + 0.013133483938872814, + 0.017427800223231316, + 0.008260922506451607, + 0.012201705016195774, + 0.01086264569312334, + 0.004024237394332886, + 0.005367726553231478, + 0.015295817516744137, + 0.0050179301761090755, + 0.009143585339188576, + 0.013262591324746609, + 0.009360511787235737, + 0.010490336455404758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.06580857187509537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790226012468338 + ], + "opposite_logprob_tokens": [ + -2.5790226459503174 + ], + "logprob_reward": -0.0790226012468338, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790226012468338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13851898908615112, + 0.04896976053714752, + 0.09042306244373322, + 0.22830133140087128, + 0.09064023941755295, + 0.03129708766937256, + 0.01533905416727066, + 0.04672008752822876, + 0.11053990572690964, + 0.1992504596710205 + ], + "credit_map": [ + 0.019746463745832443, + 0.013861914165318012, + 0.027030959725379944, + 0.02358311042189598, + 0.016064397990703583, + 0.007537466008216143, + 0.018543649464845657, + 0.012151028960943222, + 0.008988717570900917, + 0.003245784668251872, + 0.013857495039701462, + 0.004830880556255579, + 0.005027040373533964, + 0.0035238456912338734, + 0.004711173474788666, + 0.004784820135682821, + 0.013396279886364937, + 0.015473590232431889, + 0.008739246986806393, + 0.015059143304824829, + 0.009888156317174435, + 0.008681605570018291, + 0.010696838609874249, + 0.008488199673593044, + 0.024052228778600693, + 0.016428610309958458, + 0.05960211157798767, + 0.04383304342627525, + 0.01738421432673931, + 0.018011020496487617, + 0.03037833422422409, + 0.018611770123243332, + 0.015207153744995594, + 0.00847200769931078, + 0.02207997813820839, + 0.010755836963653564, + 0.008396352641284466, + 0.0072715673595666885, + 0.010394240729510784, + 0.008063104003667831, + 0.0038477175403386354, + 0.0022868334781378508, + 0.008800379000604153, + 0.006513920612633228, + 0.002214953303337097, + 0.0017778730252757668, + 0.0028538324404507875, + 0.0030015783850103617, + 0.002833941485732794, + 0.0016239364631474018, + 0.0021274369210004807, + 0.0016937608597800136, + 0.0019505323143675923, + 0.0020427179988473654, + 0.0017146550817415118, + 0.0013520725769922137, + 0.00574937229976058, + 0.00650383485481143, + 0.006345256697386503, + 0.005182045511901379, + 0.004634821321815252, + 0.007349614053964615, + 0.005587936379015446, + 0.005367207806557417, + 0.012679134495556355, + 0.015281835570931435, + 0.012779324315488338, + 0.013715648092329502, + 0.01790137216448784, + 0.01174935419112444, + 0.012838934548199177, + 0.013594305142760277, + 0.011165953241288662, + 0.014781663194298744, + 0.06253068894147873, + 0.012970262207090855, + 0.020429063588380814, + 0.030114447697997093, + 0.029579002410173416, + 0.017679385840892792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0790226012468338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252126008272171 + ], + "opposite_logprob_tokens": [ + -1.5021259784698486 + ], + "logprob_reward": -0.252126008272171, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252126008272171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40980225801467896, + 0.04119434207677841, + 0.09660876542329788, + 0.09695757180452347, + 0.07201838493347168, + 0.07623492926359177, + 0.016525615006685257, + 0.039140183478593826, + 0.08765746653079987, + 0.06386058032512665 + ], + "credit_map": [ + 0.11994276195764542, + 0.01505293883383274, + 0.15104541182518005, + 0.023190857842564583, + 0.020464424043893814, + 0.022975344210863113, + 0.03899603709578514, + 0.018134448677301407, + 0.006110964808613062, + 0.0036174750421196222, + 0.007220959756523371, + 0.0035492139868438244, + 0.0045136199332773685, + 0.0068491483107209206, + 0.0038602182175964117, + 0.00547273363918066, + 0.011352487839758396, + 0.007774559780955315, + 0.009159154258668423, + 0.010885735973715782, + 0.01806601509451866, + 0.00695611210539937, + 0.019376354292035103, + 0.013038326986134052, + 0.010116268880665302, + 0.007679823786020279, + 0.013124179095029831, + 0.01335021760314703, + 0.013059730641543865, + 0.008894383907318115, + 0.017909787595272064, + 0.012823168188333511, + 0.011943629011511803, + 0.005832947790622711, + 0.019550390541553497, + 0.007882199250161648, + 0.009127338416874409, + 0.004804167430847883, + 0.007789638824760914, + 0.005088061559945345, + 0.007465079892426729, + 0.005525798071175814, + 0.01808636449277401, + 0.014454699121415615, + 0.004677975550293922, + 0.010561246424913406, + 0.007190766744315624, + 0.008272992447018623, + 0.003131651086732745, + 0.002045213943347335, + 0.0021309913136065006, + 0.0015246837865561247, + 0.0024477129336446524, + 0.002144987927749753, + 0.001722480054013431, + 0.0013778909342363477, + 0.0043347724713385105, + 0.004834010265767574, + 0.007685840595513582, + 0.0035729475785046816, + 0.005410885903984308, + 0.00564975431188941, + 0.004295449238270521, + 0.003356517292559147, + 0.011876281350851059, + 0.010426468215882778, + 0.011718755587935448, + 0.009712796658277512, + 0.01520464662462473, + 0.0076749552972614765, + 0.011279786005616188, + 0.009763763286173344, + 0.0033367995638400316, + 0.004492879845201969, + 0.02005373314023018, + 0.00432346248999238, + 0.005805703811347485, + 0.009813331067562103, + 0.010020260699093342, + 0.006014401093125343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.252126008272171 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306979477405548 + ], + "opposite_logprob_tokens": [ + -3.168069839477539 + ], + "logprob_reward": -0.04306979477405548, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306979477405548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19131721556186676, + 0.04739023372530937, + 0.083504319190979, + 0.13019973039627075, + 0.06401480734348297, + 0.043426115065813065, + 0.03829297795891762, + 0.10113026201725006, + 0.20656725764274597, + 0.09415709972381592 + ], + "credit_map": [ + 0.020701909437775612, + 0.02075950987637043, + 0.04286261647939682, + 0.030070895329117775, + 0.021695498377084732, + 0.01725941337645054, + 0.021244658157229424, + 0.016722703352570534, + 0.010476479306817055, + 0.003177892416715622, + 0.013101402670145035, + 0.004543513059616089, + 0.0036084598395973444, + 0.0027726844418793917, + 0.005338235758244991, + 0.00437156530097127, + 0.010545583441853523, + 0.013426763005554676, + 0.010631011798977852, + 0.014460638165473938, + 0.007426599971950054, + 0.011567778885364532, + 0.008395396173000336, + 0.0070505463518202305, + 0.013366270810365677, + 0.010850883089005947, + 0.011737775057554245, + 0.016104841604828835, + 0.020961778238415718, + 0.013346190564334393, + 0.027215586975216866, + 0.01661640591919422, + 0.007493889890611172, + 0.007826445624232292, + 0.0099150650203228, + 0.007747537922114134, + 0.007610793225467205, + 0.0062981280498206615, + 0.008749040775001049, + 0.008373904041945934, + 0.0066017527133226395, + 0.0022057804744690657, + 0.011932008899748325, + 0.009027855470776558, + 0.0024722442030906677, + 0.002223397372290492, + 0.004702912177890539, + 0.004260162357240915, + 0.0054597374983131886, + 0.0037423393223434687, + 0.004545911215245724, + 0.0047333077527582645, + 0.003407888114452362, + 0.0062468647956848145, + 0.005060900934040546, + 0.00509603088721633, + 0.012147864326834679, + 0.01424960047006607, + 0.013406893238425255, + 0.01225188560783863, + 0.011568012647330761, + 0.018774067983031273, + 0.009214883670210838, + 0.009517050348222256, + 0.023010071367025375, + 0.026967445388436317, + 0.024810953065752983, + 0.028473425656557083, + 0.0356755293905735, + 0.018600735813379288, + 0.024205688387155533, + 0.024823401123285294, + 0.005631226114928722, + 0.01827530935406685, + 0.009133958257734776, + 0.008830969221889973, + 0.01791871339082718, + 0.011231428943574429, + 0.015438683331012726, + 0.0076968055218458176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.04306979477405548 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981220744550228 + ], + "opposite_logprob_tokens": [ + -3.5298120975494385 + ], + "logprob_reward": -0.02981220744550228, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981220744550228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15044021606445312, + 0.03286829963326454, + 0.05152690038084984, + 0.10688962042331696, + 0.10767824947834015, + 0.015523001551628113, + 0.016589809209108353, + 0.08353663980960846, + 0.3182262182235718, + 0.11672105640172958 + ], + "credit_map": [ + 0.04466572403907776, + 0.009660433977842331, + 0.029044363647699356, + 0.010562402196228504, + 0.014553770422935486, + 0.007731682620942593, + 0.023463409394025803, + 0.010758443735539913, + 0.006071955431252718, + 0.002559241373091936, + 0.009032976813614368, + 0.0025803078897297382, + 0.0027090460062026978, + 0.004169607535004616, + 0.0028850610833615065, + 0.0028601037338376045, + 0.008264259435236454, + 0.00918592419475317, + 0.006058812141418457, + 0.009192198514938354, + 0.005368612240999937, + 0.005245789419859648, + 0.004221129231154919, + 0.003990171942859888, + 0.01541505940258503, + 0.012522827833890915, + 0.028038203716278076, + 0.019988827407360077, + 0.006095362361520529, + 0.006225021556019783, + 0.011846226640045643, + 0.0067580873146653175, + 0.017036836594343185, + 0.009987779892981052, + 0.024486029520630836, + 0.012828349135816097, + 0.011831747367978096, + 0.008817014284431934, + 0.01284121349453926, + 0.009849282912909985, + 0.0014514727517962456, + 0.0013290708884596825, + 0.001511021051555872, + 0.0012287257704883814, + 0.002984049264341593, + 0.0019886991940438747, + 0.002259705914184451, + 0.002770256018266082, + 0.0036535600665956736, + 0.001270180568099022, + 0.0046193525195121765, + 0.0011098368559032679, + 0.0011318400502204895, + 0.0015999391907826066, + 0.0015701416414231062, + 0.0016349583165720105, + 0.009763652458786964, + 0.011686770245432854, + 0.009692718274891376, + 0.009960992261767387, + 0.008874195627868176, + 0.01674315333366394, + 0.008716762997210026, + 0.008098392747342587, + 0.0282550398260355, + 0.04369417577981949, + 0.04437946528196335, + 0.04469061270356178, + 0.05279030278325081, + 0.03087465465068817, + 0.038200702518224716, + 0.03534125164151192, + 0.006960141472518444, + 0.016288157552480698, + 0.017452098429203033, + 0.008324977941811085, + 0.013940194621682167, + 0.016540374606847763, + 0.018229037523269653, + 0.018986079841852188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.02981220744550228 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot, shot from an upward angle, is imbued with the ethos of a martial arts film. In the frame, a warrior clad in pristine, dark iron armor is at the peak of his power. He maintains a firm horse stance, his center of gravity extremely low, his muscles bulging. His weathered armor, composed of plates of varying sizes strung together with leather cords, is covered in scratches and battle scars. The low, golden sunlight of dusk reflects a deep, metallic sheen. His long, black hair dances wildly back in the gust of wind, a few strands brushing against his stern, focused face, eyes wide and teeth clenched. With a silent roar, he thrusts his right fist forward with the force of thunder, the air around it seemingly warping slightly with the immense energy. His fist struck a massive bluestone stele before him, taller than a man, with precise and powerful force. At the moment of contact, a spiderweb of cracks centered on the center of his fist erupted across its surface. Then, with a resounding, muffled thud, the entire stele exploded, sending countless boulders and dust flying in all directions, clearly visible in the backlight. The entire scene took place on a desolate mountaintop, against a backdrop of sparse, withered trees and a sky stained orange-red by the sunset glow. The scene created an epic atmosphere of aloofness, power, and a destructive aesthetic.", + "seed": 24315, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_1185_seed_24315_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": -0.06051939260214567, + "vlm_logprob": -0.06051939260214567, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.06051939260214567, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0539972223341465 + ], + "opposite_logprob_tokens": [ + -3.303997278213501 + ], + "logprob_reward": -0.0539972223341465, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0539972223341465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1569727659225464, + 0.11529551446437836, + 0.12125897407531738, + 0.03637887164950371, + 0.09250717610120773, + 0.09850005060434341, + 0.16894054412841797, + 0.08126886934041977, + 0.03643226623535156, + 0.09244505316019058 + ], + "credit_map": [ + 0.015496153384447098, + 0.01838715188205242, + 0.01745152473449707, + 0.020889829844236374, + 0.016069291159510612, + 0.028092725202441216, + 0.018274793401360512, + 0.022311272099614143, + 0.017781883478164673, + 0.008885149843990803, + 0.021821560338139534, + 0.010627051815390587, + 0.01045769453048706, + 0.016660228371620178, + 0.013339613564312458, + 0.01572231762111187, + 0.014240406453609467, + 0.007084663026034832, + 0.012295695021748543, + 0.010598918423056602, + 0.00786598864942789, + 0.03402072191238403, + 0.011494015343487263, + 0.02365855500102043, + 0.00399185111746192, + 0.0025665026623755693, + 0.004494437947869301, + 0.0034523142967373133, + 0.003082048613578081, + 0.008578216657042503, + 0.0038004741072654724, + 0.006413024850189686, + 0.01454315334558487, + 0.007349403575062752, + 0.021391887217760086, + 0.008662411943078041, + 0.00667553348466754, + 0.01284336019307375, + 0.010852260515093803, + 0.010189161635935307, + 0.015190929174423218, + 0.006350470706820488, + 0.011638302356004715, + 0.007038533221930265, + 0.005415825638920069, + 0.013937694020569324, + 0.01017062645405531, + 0.028757654130458832, + 0.01578158512711525, + 0.009450147859752178, + 0.014602030627429485, + 0.007972321473062038, + 0.007058382034301758, + 0.04847307503223419, + 0.011712629348039627, + 0.05389035493135452, + 0.009592758491635323, + 0.010648541152477264, + 0.010030888020992279, + 0.015863219276070595, + 0.009613841772079468, + 0.009314018301665783, + 0.00880010798573494, + 0.00740548362955451, + 0.005697827786207199, + 0.0048933858051896095, + 0.005044678691774607, + 0.00908351968973875, + 0.002130760345607996, + 0.003272095462307334, + 0.0030892749782651663, + 0.003220720449462533, + 0.013086656108498573, + 0.007820751518011093, + 0.011068136431276798, + 0.013940868899226189, + 0.008061661384999752, + 0.008737619034945965, + 0.01697660982608795, + 0.012752740643918514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0539972223341465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710453569889069 + ], + "opposite_logprob_tokens": [ + -2.1271045207977295 + ], + "logprob_reward": -0.12710453569889069, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710453569889069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18593844771385193, + 0.08148445934057236, + 0.07872984558343887, + 0.2172822207212448, + 0.047611042857170105, + 0.15880317986011505, + 0.06080055609345436, + 0.06097470596432686, + 0.05779183283448219, + 0.05058376118540764 + ], + "credit_map": [ + 0.02051856555044651, + 0.015177707187831402, + 0.030235782265663147, + 0.02565917745232582, + 0.015203213319182396, + 0.02375493384897709, + 0.030786888673901558, + 0.024602176621556282, + 0.011655445210635662, + 0.006661238614469767, + 0.014554272405803204, + 0.010414703749120235, + 0.00881106499582529, + 0.009458617307245731, + 0.010112592950463295, + 0.009816517122089863, + 0.01037564780563116, + 0.0038425761740654707, + 0.007809814531356096, + 0.006915220990777016, + 0.005101401824504137, + 0.022506020963191986, + 0.007864100858569145, + 0.014315048232674599, + 0.019068041816353798, + 0.004840684588998556, + 0.018956027925014496, + 0.011418668553233147, + 0.005567688029259443, + 0.061136502772569656, + 0.007085815537720919, + 0.08920877426862717, + 0.007699279580265284, + 0.003690929850563407, + 0.008805813267827034, + 0.003892291570082307, + 0.005279085133224726, + 0.004744820296764374, + 0.006611207965761423, + 0.006887604482471943, + 0.02848156914114952, + 0.01158334780484438, + 0.013604617677628994, + 0.009526688605546951, + 0.008556663990020752, + 0.025040242820978165, + 0.010330135934054852, + 0.05167989432811737, + 0.011574559845030308, + 0.004242243245244026, + 0.007489820942282677, + 0.004448118153959513, + 0.0033038451801985502, + 0.011286607012152672, + 0.004705383442342281, + 0.01374997291713953, + 0.006848513148725033, + 0.007868235930800438, + 0.00838820543140173, + 0.012479759752750397, + 0.0058268834836781025, + 0.006145944818854332, + 0.006754160858690739, + 0.006662999279797077, + 0.006742868106812239, + 0.007251264993101358, + 0.006622408516705036, + 0.014939108863472939, + 0.004357786383479834, + 0.005859934724867344, + 0.0054602413438260555, + 0.006558212451636791, + 0.0075365277007222176, + 0.005552820395678282, + 0.005934025160968304, + 0.007572727277874947, + 0.004067261703312397, + 0.008303210139274597, + 0.005604677367955446, + 0.006012509111315012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.12710453569889069 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030989348888397 + ], + "opposite_logprob_tokens": [ + -2.3503098487854004 + ], + "logprob_reward": -0.10030989348888397, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030989348888397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25670126080513, + 0.08579155802726746, + 0.16405129432678223, + 0.09069617092609406, + 0.027035607025027275, + 0.16035360097885132, + 0.05158035829663277, + 0.05431407690048218, + 0.027438046410679817, + 0.08203807473182678 + ], + "credit_map": [ + 0.02130972594022751, + 0.033586353063583374, + 0.03166457638144493, + 0.03968611732125282, + 0.02868320792913437, + 0.03334473446011543, + 0.03876679018139839, + 0.029659735038876534, + 0.010718525387346745, + 0.008357580751180649, + 0.011931100860238075, + 0.009123989380896091, + 0.008415229618549347, + 0.01079417485743761, + 0.014790193177759647, + 0.011660763062536716, + 0.019968388602137566, + 0.007777382154017687, + 0.01574009843170643, + 0.011425061151385307, + 0.008399770595133305, + 0.055119588971138, + 0.012364326976239681, + 0.03325666859745979, + 0.007848388515412807, + 0.0047936211340129375, + 0.00823611207306385, + 0.008344694040715694, + 0.004586734808981419, + 0.01963106542825699, + 0.00770202511921525, + 0.029553519561886787, + 0.0029141176491975784, + 0.002379491925239563, + 0.006100805010646582, + 0.002435956383123994, + 0.002442209515720606, + 0.004017693921923637, + 0.003119808854535222, + 0.003625522367656231, + 0.02824026718735695, + 0.011949916370213032, + 0.011017590761184692, + 0.010583956725895405, + 0.006446004379540682, + 0.02726648934185505, + 0.007381667848676443, + 0.057467713952064514, + 0.007885477505624294, + 0.004370925948023796, + 0.006987690459936857, + 0.004215978551656008, + 0.004251815844327211, + 0.009088348597288132, + 0.004317181184887886, + 0.01046293880790472, + 0.005828186869621277, + 0.007170505356043577, + 0.006317196413874626, + 0.011967601254582405, + 0.006675460375845432, + 0.005495407618582249, + 0.00629281485453248, + 0.004566899035125971, + 0.003558405674993992, + 0.004405813291668892, + 0.0036600835155695677, + 0.004781510680913925, + 0.0022393222898244858, + 0.0024918646086007357, + 0.002798047149553895, + 0.003502999432384968, + 0.011313038878142834, + 0.00784105435013771, + 0.010213096626102924, + 0.014866361394524574, + 0.007055508904159069, + 0.011742976494133472, + 0.0111387949436903, + 0.007867231965065002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.10030989348888397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540061958134174 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012540061958134174, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540061958134174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06028369069099426, + 0.053134191781282425, + 0.11940845847129822, + 0.19141632318496704, + 0.0368252769112587, + 0.1813204139471054, + 0.07521331310272217, + 0.1386140137910843, + 0.09519088268280029, + 0.04859340563416481 + ], + "credit_map": [ + 0.005858474411070347, + 0.005975223146378994, + 0.006982933264225721, + 0.00729794055223465, + 0.005886181257665157, + 0.008303819224238396, + 0.006743462756276131, + 0.013235662132501602, + 0.007910340093076229, + 0.004783085081726313, + 0.00797091145068407, + 0.005545127205550671, + 0.005145269446074963, + 0.006308003794401884, + 0.009022981859743595, + 0.006448470056056976, + 0.014440701343119144, + 0.0063417754136025906, + 0.010779437609016895, + 0.007739245891571045, + 0.006223280914127827, + 0.03925130516290665, + 0.010265674442052841, + 0.02436704933643341, + 0.016826460137963295, + 0.005001812241971493, + 0.021242553368210793, + 0.011124606244266033, + 0.007608841639012098, + 0.04914805293083191, + 0.008903195150196552, + 0.0715608298778534, + 0.005860432982444763, + 0.0030272172298282385, + 0.006967947818338871, + 0.0032830345444381237, + 0.004491238854825497, + 0.0033461994025856256, + 0.004904367029666901, + 0.004944842774420977, + 0.032427024096250534, + 0.011233200319111347, + 0.015865961089730263, + 0.011900905519723892, + 0.007860173471271992, + 0.029778294265270233, + 0.010212424211204052, + 0.06204245239496231, + 0.006931768264621496, + 0.0033346519339829683, + 0.008233032189309597, + 0.004441050346940756, + 0.005169621203094721, + 0.023122871294617653, + 0.005494529381394386, + 0.018485791981220245, + 0.015343832783401012, + 0.019674427807331085, + 0.019034424796700478, + 0.021841678768396378, + 0.018263138830661774, + 0.012775356881320477, + 0.01590144634246826, + 0.015779724344611168, + 0.012430968694388866, + 0.0104078259319067, + 0.010120267048478127, + 0.029850099235773087, + 0.006187898572534323, + 0.009546111337840557, + 0.009050235152244568, + 0.007597486022859812, + 0.007134216837584972, + 0.004726397339254618, + 0.005388500634580851, + 0.0062672789208590984, + 0.005776751320809126, + 0.006858851294964552, + 0.005959568545222282, + 0.00648184260353446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.012540061958134174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645249530673027 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645249530673027, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645249530673027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04051723703742027, + 0.05498639866709709, + 0.1407897174358368, + 0.05126849189400673, + 0.19984693825244904, + 0.04340345785021782, + 0.1093815416097641, + 0.10723688453435898, + 0.06750623881816864, + 0.18506306409835815 + ], + "credit_map": [ + 0.004583875183016062, + 0.0052507175132632256, + 0.0045303418301045895, + 0.0055102091282606125, + 0.004667987115681171, + 0.005050417501479387, + 0.005377328488975763, + 0.005546363536268473, + 0.007721832487732172, + 0.004126283340156078, + 0.010686970315873623, + 0.0061172121204435825, + 0.003854679176583886, + 0.007144202943891287, + 0.006278401240706444, + 0.009056825190782547, + 0.01842423342168331, + 0.007495497353374958, + 0.013092194683849812, + 0.008569478057324886, + 0.007712339051067829, + 0.04631277546286583, + 0.01039781142026186, + 0.028785405680537224, + 0.007766338065266609, + 0.0035359454341232777, + 0.0061990562826395035, + 0.005412082187831402, + 0.003817533142864704, + 0.008308911696076393, + 0.00518072908744216, + 0.011047899723052979, + 0.03749953955411911, + 0.013213548809289932, + 0.04357115551829338, + 0.017404548823833466, + 0.017698984593153, + 0.016186321154236794, + 0.027651794254779816, + 0.02662106789648533, + 0.003787185763940215, + 0.0038789359387010336, + 0.005611992906779051, + 0.004887358285486698, + 0.004403422586619854, + 0.008537719026207924, + 0.00774755934253335, + 0.004549289587885141, + 0.016751745715737343, + 0.007528373505920172, + 0.013251847587525845, + 0.006252102553844452, + 0.0054438309744000435, + 0.02215050905942917, + 0.007670371327549219, + 0.030332770198583603, + 0.010854716412723064, + 0.011867506429553032, + 0.013938783667981625, + 0.017879556864500046, + 0.015390969812870026, + 0.014643938280642033, + 0.012380809523165226, + 0.010280613787472248, + 0.010922779329121113, + 0.011439650319516659, + 0.007764262147247791, + 0.016159873455762863, + 0.005590121727436781, + 0.006432054564356804, + 0.005932967644184828, + 0.0032645389437675476, + 0.027636531740427017, + 0.025782650336623192, + 0.023729320615530014, + 0.02694208361208439, + 0.014187789522111416, + 0.022138802334666252, + 0.023044265806674957, + 0.02160162851214409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.008645249530673027 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot, shot from an upward angle, is imbued with the ethos of a martial arts film. In the frame, a warrior clad in pristine, dark iron armor is at the peak of his power. He maintains a firm horse stance, his center of gravity extremely low, his muscles bulging. His weathered armor, composed of plates of varying sizes strung together with leather cords, is covered in scratches and battle scars. The low, golden sunlight of dusk reflects a deep, metallic sheen. His long, black hair dances wildly back in the gust of wind, a few strands brushing against his stern, focused face, eyes wide and teeth clenched. With a silent roar, he thrusts his right fist forward with the force of thunder, the air around it seemingly warping slightly with the immense energy. His fist struck a massive bluestone stele before him, taller than a man, with precise and powerful force. At the moment of contact, a spiderweb of cracks centered on the center of his fist erupted across its surface. Then, with a resounding, muffled thud, the entire stele exploded, sending countless boulders and dust flying in all directions, clearly visible in the backlight. The entire scene took place on a desolate mountaintop, against a backdrop of sparse, withered trees and a sky stained orange-red by the sunset glow. The scene created an epic atmosphere of aloofness, power, and a destructive aesthetic.", + "seed": 24316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_6/index_1185_seed_24316_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": -0.04244688786566257, + "vlm_logprob": -0.04244688786566257, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04244688786566257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05214674770832062 + ], + "opposite_logprob_tokens": [ + -3.3021466732025146 + ], + "logprob_reward": -0.05214674770832062, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05214674770832062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1302218735218048, + 0.0581556111574173, + 0.13121989369392395, + 0.19948166608810425, + 0.1023373082280159, + 0.11815337836742401, + 0.04643482342362404, + 0.0505746565759182, + 0.11380679160356522, + 0.04961409047245979 + ], + "credit_map": [ + 0.019963426515460014, + 0.018567141145467758, + 0.01804329641163349, + 0.022804951295256615, + 0.01295522227883339, + 0.010744251310825348, + 0.01180104911327362, + 0.015342512167990208, + 0.006422463338822126, + 0.008997233584523201, + 0.0066370368003845215, + 0.010659855790436268, + 0.007217472419142723, + 0.004748400766402483, + 0.0058568562380969524, + 0.007616284303367138, + 0.014202931895852089, + 0.013788065873086452, + 0.017264382913708687, + 0.019482823088765144, + 0.017167309299111366, + 0.01790536195039749, + 0.013549395836889744, + 0.017859607934951782, + 0.026686836034059525, + 0.027108384296298027, + 0.025875259190797806, + 0.051524192094802856, + 0.01580061949789524, + 0.013491163961589336, + 0.017325492575764656, + 0.02166971005499363, + 0.011814659461379051, + 0.012522747740149498, + 0.021066168323159218, + 0.013923509046435356, + 0.01087313611060381, + 0.012490756809711456, + 0.008844933472573757, + 0.010801389813423157, + 0.035808853805065155, + 0.014162993989884853, + 0.010598750784993172, + 0.01129722036421299, + 0.009723646566271782, + 0.01571611501276493, + 0.009584286250174046, + 0.011261501349508762, + 0.007240889593958855, + 0.008700001984834671, + 0.004058687947690487, + 0.00620601000264287, + 0.005892082117497921, + 0.004237743094563484, + 0.00631801737472415, + 0.003781383391469717, + 0.006625032052397728, + 0.009854578413069248, + 0.005147768184542656, + 0.005146922077983618, + 0.0075592901557683945, + 0.004117457661777735, + 0.0067347935400903225, + 0.005388808902353048, + 0.024168387055397034, + 0.03790391981601715, + 0.00815972127020359, + 0.005934273824095726, + 0.009566017426550388, + 0.008535707369446754, + 0.008915544487535954, + 0.010623213835060596, + 0.005036620423197746, + 0.005863108206540346, + 0.005479997023940086, + 0.005184712819755077, + 0.0066344281658530235, + 0.003956224769353867, + 0.01068954635411501, + 0.0067694480530917645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.05214674770832062 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062090858817100525 + ], + "opposite_logprob_tokens": [ + -2.8120908737182617 + ], + "logprob_reward": -0.062090858817100525, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062090858817100525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14939701557159424, + 0.049447160214185715, + 0.36493462324142456, + 0.07365018874406815, + 0.07740963250398636, + 0.047834914177656174, + 0.06667660921812057, + 0.027187684550881386, + 0.09111696481704712, + 0.052345260977745056 + ], + "credit_map": [ + 0.019838927313685417, + 0.02309148572385311, + 0.020595336332917213, + 0.024449991062283516, + 0.01516713760793209, + 0.010356497950851917, + 0.016798749566078186, + 0.01909889467060566, + 0.0043967897072434425, + 0.005585658363997936, + 0.006167995743453503, + 0.007629160303622484, + 0.007182903587818146, + 0.006810082122683525, + 0.004566926043480635, + 0.007107642944902182, + 0.022732270881533623, + 0.030520601198077202, + 0.0902748852968216, + 0.05505990609526634, + 0.03825138136744499, + 0.026628943160176277, + 0.02933826856315136, + 0.07212837040424347, + 0.005760240368545055, + 0.009045446291565895, + 0.00794951245188713, + 0.023484328761696815, + 0.0036432372871786356, + 0.0035852137953042984, + 0.004312010016292334, + 0.01587020233273506, + 0.008208566345274448, + 0.010138215497136116, + 0.017820794135332108, + 0.008397961035370827, + 0.010084385983645916, + 0.01018418651074171, + 0.00586104579269886, + 0.006714479997754097, + 0.0151850376278162, + 0.004666259046643972, + 0.0052957977168262005, + 0.004780633375048637, + 0.003310002386569977, + 0.005523900501430035, + 0.005460378248244524, + 0.0036129015497863293, + 0.006560083478689194, + 0.012807253748178482, + 0.00491845328360796, + 0.011118537746369839, + 0.0125192916020751, + 0.006498515140265226, + 0.008791687898337841, + 0.003462784457951784, + 0.003072231775149703, + 0.003439672291278839, + 0.0037250970490276814, + 0.0030088622588664293, + 0.004097369033843279, + 0.0022916439920663834, + 0.004648212809115648, + 0.002904596272855997, + 0.021250411868095398, + 0.03329547122120857, + 0.004904656670987606, + 0.004031129647046328, + 0.0060301292687654495, + 0.006860490422695875, + 0.006974915042519569, + 0.007769760675728321, + 0.004529263358563185, + 0.006097997073084116, + 0.006014765240252018, + 0.0049432264640927315, + 0.006174825597554445, + 0.004032023251056671, + 0.012961845844984055, + 0.007591312751173973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.062090858817100525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002115249633789 + ], + "opposite_logprob_tokens": [ + -2.695021152496338 + ], + "logprob_reward": -0.07002115249633789, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002115249633789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11395207047462463, + 0.03699677065014839, + 0.1951882690191269, + 0.04624537378549576, + 0.2755888104438782, + 0.06421012431383133, + 0.12523521482944489, + 0.024882731959223747, + 0.08977620303630829, + 0.02792445383965969 + ], + "credit_map": [ + 0.01231272704899311, + 0.014049438759684563, + 0.014285532757639885, + 0.020058071240782738, + 0.012256340123713017, + 0.009458699263632298, + 0.014626274816691875, + 0.016904985532164574, + 0.002664038445800543, + 0.0045966594479978085, + 0.005139602348208427, + 0.005833953153342009, + 0.004731066059321165, + 0.005382738076150417, + 0.0034251504112035036, + 0.005223564337939024, + 0.013806924223899841, + 0.01641194149851799, + 0.06194670498371124, + 0.02998179756104946, + 0.012182238511741161, + 0.015204153954982758, + 0.009729450568556786, + 0.035925038158893585, + 0.0054037743248045444, + 0.005865101702511311, + 0.006243136245757341, + 0.012429501861333847, + 0.00350970309227705, + 0.003054453991353512, + 0.003962771035730839, + 0.005776925943791866, + 0.03614983707666397, + 0.031956739723682404, + 0.05657598376274109, + 0.03718337416648865, + 0.029141593724489212, + 0.0330144539475441, + 0.022745955735445023, + 0.02882087044417858, + 0.0169586930423975, + 0.007832110859453678, + 0.003696815110743046, + 0.008201704360544682, + 0.006946760229766369, + 0.006342883687466383, + 0.00984528660774231, + 0.004385869484394789, + 0.010434974916279316, + 0.02224714122712612, + 0.017356842756271362, + 0.01740150898694992, + 0.015000811778008938, + 0.012063778005540371, + 0.02075509913265705, + 0.009975053369998932, + 0.0025689091999083757, + 0.004018750507384539, + 0.004097824916243553, + 0.0032360306940972805, + 0.0031260012183338404, + 0.0020578759722411633, + 0.0037445540074259043, + 0.0020327854435890913, + 0.017558731138706207, + 0.03130359575152397, + 0.004696566611528397, + 0.0030576966237276793, + 0.005340035539120436, + 0.006293240003287792, + 0.007212902419269085, + 0.01431343611329794, + 0.004102854523807764, + 0.003917665686458349, + 0.002957177348434925, + 0.002315941732376814, + 0.003140706103295088, + 0.00349999126046896, + 0.004608433227986097, + 0.003381684422492981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.07002115249633789 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186835572123528 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018186835572123528, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186835572123528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11132333427667618, + 0.029066909104585648, + 0.12348122894763947, + 0.07655546069145203, + 0.07799594849348068, + 0.10618969798088074, + 0.1897529661655426, + 0.024686653167009354, + 0.0857105702161789, + 0.17523714900016785 + ], + "credit_map": [ + 0.01228850707411766, + 0.012690666131675243, + 0.01284763403236866, + 0.017140690237283707, + 0.013066401705145836, + 0.00703894067555666, + 0.018034717068076134, + 0.018215779215097427, + 0.0032691184896975756, + 0.002781270071864128, + 0.003300940152257681, + 0.004238945432007313, + 0.004148704465478659, + 0.0043834964744746685, + 0.0022459877654910088, + 0.004698450211435556, + 0.010634835809469223, + 0.014257204718887806, + 0.029276568442583084, + 0.01956802047789097, + 0.012690938077867031, + 0.01149891410022974, + 0.007871832698583603, + 0.01768292300403118, + 0.010958928614854813, + 0.010049131698906422, + 0.010602794587612152, + 0.009344369173049927, + 0.005321319680660963, + 0.006193187553435564, + 0.005109229125082493, + 0.01897650957107544, + 0.009795191697776318, + 0.009800095111131668, + 0.009170480072498322, + 0.010620282962918282, + 0.010550905019044876, + 0.012068260461091995, + 0.007843246683478355, + 0.00814750138670206, + 0.020655326545238495, + 0.00884498655796051, + 0.00718261394649744, + 0.01187234278768301, + 0.008058696053922176, + 0.012529375962913036, + 0.021760359406471252, + 0.015286009758710861, + 0.019641734659671783, + 0.02656164951622486, + 0.01905766688287258, + 0.02455129101872444, + 0.022420167922973633, + 0.036669038236141205, + 0.029501551762223244, + 0.011349876411259174, + 0.003362504532560706, + 0.0030132834799587727, + 0.0026141733396798372, + 0.0025599247310310602, + 0.00423298729583621, + 0.0017753953579813242, + 0.003877311944961548, + 0.003251074580475688, + 0.015346243977546692, + 0.023994674906134605, + 0.005643891170620918, + 0.0043894206173717976, + 0.011078111827373505, + 0.007762718480080366, + 0.010364748537540436, + 0.007130766753107309, + 0.014496738091111183, + 0.02593342214822769, + 0.01183073315769434, + 0.015382777899503708, + 0.027076026424765587, + 0.014930257573723793, + 0.04380575194954872, + 0.02178145945072174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.018186835572123528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 an upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing dark iron armor composed of plates strung with leather cords?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior in a horse stance with a low center of gravity and bulging muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the warrior thrust his right fist forward to strike a massive bluestone stele?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the stele explode into boulders and dust upon impact in a desolate mountaintop setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788844734430313 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.009788844734430313, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788844734430313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08416414260864258, + 0.02953735925257206, + 0.034505695104599, + 0.027113884687423706, + 0.03710928186774254, + 0.07105791568756104, + 0.06702131778001785, + 0.03825416415929794, + 0.3577980101108551, + 0.25343817472457886 + ], + "credit_map": [ + 0.008780200034379959, + 0.012494032271206379, + 0.00969769898802042, + 0.016125675290822983, + 0.007229363080114126, + 0.007456232327967882, + 0.009570242837071419, + 0.01281069964170456, + 0.0026379285845905542, + 0.0031633051112294197, + 0.006139656063169241, + 0.004972649272531271, + 0.002839126158505678, + 0.002844494301825762, + 0.00250824261456728, + 0.004431958310306072, + 0.0030813138000667095, + 0.005453038495033979, + 0.006084398366510868, + 0.0076075061224401, + 0.0030205987859517336, + 0.0034343518782407045, + 0.0023637826088815928, + 0.0034607085399329662, + 0.0018881993601098657, + 0.0038026024121791124, + 0.0035785811487585306, + 0.005872044246643782, + 0.003054175991564989, + 0.0019156737253069878, + 0.003198313992470503, + 0.0038042967207729816, + 0.004452827386558056, + 0.004031516145914793, + 0.008305882103741169, + 0.004966188687831163, + 0.004010696429759264, + 0.0048534986563026905, + 0.002807909157127142, + 0.003680761670693755, + 0.021692750975489616, + 0.006154279224574566, + 0.006681790109723806, + 0.006319816689938307, + 0.005066078156232834, + 0.009525571018457413, + 0.0071234991773962975, + 0.00849413312971592, + 0.007091491483151913, + 0.010969978757202625, + 0.005413149017840624, + 0.009389207698404789, + 0.013208400458097458, + 0.008195893839001656, + 0.008977781981229782, + 0.0037754138465970755, + 0.006476277019828558, + 0.00538629200309515, + 0.004277380183339119, + 0.004871933721005917, + 0.0052651152946054935, + 0.0032733341213315725, + 0.004825458396226168, + 0.0038783717900514603, + 0.052849240601062775, + 0.10211194306612015, + 0.021507173776626587, + 0.012098177336156368, + 0.019106676802039146, + 0.01967346854507923, + 0.04637128859758377, + 0.0840800479054451, + 0.02104397490620613, + 0.040149156004190445, + 0.023261303082108498, + 0.02119767665863037, + 0.03385559096932411, + 0.020068181678652763, + 0.06272492557764053, + 0.03113737888634205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.009788844734430313 + } + ], + "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/0000058/group_7/index_tensor([1219, 1219])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_tensor([1219, 1219])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..48b8dec1d6ebe49911ae626bfcf368fcb6f346ec --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_tensor([1219, 1219])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping shot begins at a very low angle, then smoothly pulls back and rises, ultimately transforming into a sweeping, wide-angle bird's-eye view. Against a dramatic dusk sky, the parched, cracked earth violently crumbles, and countless massive metal branches, gleaming with a cold, silver-gray sheen, burst forth with unstoppable force. These branches, bearing the marks of industrial welding and gently weathered textures, resemble living pythons, rapidly growing, interweaving, bending, and seamlessly welding together in a rhythm imbued with power and organic beauty. Low, golden sunlight paints sharp highlights on their curved surfaces and casts long, ever-shifting, dynamic shadows, creating a striking contrast against the sky, tinted in deep orange and indigo. As the camera zooms out, these seemingly disorganized metal structures gradually weave into a magnificent, precisely intricate grid, ultimately forming the iconic silhouette of the Beijing Bird's Nest Stadium across the vast plain. The entire process is imbued with a surreal atmosphere of creativity and awe-inspiring industrial aesthetic.", + "seed": 3083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3083_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": -0.00975643282290548, + "vlm_logprob": -0.00975643282290548, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.00975643282290548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.023377282544970512 + ], + "opposite_logprob_tokens": [ + -3.8983771800994873 + ], + "logprob_reward": -0.023377282544970512, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023377282544970512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3162333369255066, + 0.04967604577541351, + 0.1908753365278244, + 0.04631154611706734, + 0.13045042753219604, + 0.06486796587705612, + 0.09320424497127533, + 0.05828757584095001, + 0.024087779223918915, + 0.026005784049630165 + ], + "credit_map": [ + 0.01613728702068329, + 0.016384031623601913, + 0.12380635738372803, + 0.12380369752645493, + 0.01114464271813631, + 0.009090443141758442, + 0.009034738875925541, + 0.0068321144208312035, + 0.006711126770824194, + 0.005619879346340895, + 0.008822550997138023, + 0.0043789842166006565, + 0.011743949726223946, + 0.007586036343127489, + 0.0029229552019387484, + 0.0018905584001913667, + 0.012147621251642704, + 0.007359902840107679, + 0.06204361096024513, + 0.07317337393760681, + 0.006758725270628929, + 0.00866263173520565, + 0.011731011793017387, + 0.008998456411063671, + 0.004151861649006605, + 0.003943827003240585, + 0.00422836747020483, + 0.0033331571612507105, + 0.006865369156002998, + 0.00586013775318861, + 0.0030981351155787706, + 0.014830688945949078, + 0.028353767469525337, + 0.010786734521389008, + 0.0078832246363163, + 0.01278204657137394, + 0.015775851905345917, + 0.014588027261197567, + 0.004909212235361338, + 0.03537154942750931, + 0.01033345889300108, + 0.006371904164552689, + 0.004597979597747326, + 0.0071802022866904736, + 0.016246413812041283, + 0.009098954498767853, + 0.004416449926793575, + 0.0066225905902683735, + 0.02101682312786579, + 0.00998673401772976, + 0.005611752159893513, + 0.005749233067035675, + 0.01620527170598507, + 0.01095705758780241, + 0.005468865856528282, + 0.018208494409918785, + 0.006551799830049276, + 0.008073142729699612, + 0.006424691062420607, + 0.005798605270683765, + 0.011165530420839787, + 0.006827745586633682, + 0.006641137879341841, + 0.006804916076362133, + 0.004686169791966677, + 0.002392716705799103, + 0.0037883499171584845, + 0.0019562828820198774, + 0.002122825011610985, + 0.004449306055903435, + 0.0022704312577843666, + 0.002421693643555045, + 0.0045867501758039, + 0.004615481942892075, + 0.002369269961491227, + 0.0012185873929411173, + 0.0024494521785527468, + 0.0028333663940429688, + 0.003027235856279731, + 0.004905638284981251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.023377282544970512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.0017296605510637164 + ], + "opposite_logprob_tokens": [ + -6.376729488372803 + ], + "logprob_reward": -0.0017296605510637164, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017296605510637164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45756328105926514, + 0.027487456798553467, + 0.05643770098686218, + 0.08253899216651917, + 0.055122267454862595, + 0.05525142699480057, + 0.026813890784978867, + 0.07344034314155579, + 0.10512186586856842, + 0.060222797095775604 + ], + "credit_map": [ + 0.02142510749399662, + 0.015297931618988514, + 0.18983350694179535, + 0.18705129623413086, + 0.012259158305823803, + 0.011235134676098824, + 0.011588512919843197, + 0.00887265894562006, + 0.0023982960265129805, + 0.003190472489222884, + 0.003283134661614895, + 0.0031513951253145933, + 0.006140450481325388, + 0.005948235280811787, + 0.0017489545280113816, + 0.0016265189042314887, + 0.0036324241664260626, + 0.0030618112068623304, + 0.016230029985308647, + 0.02110747992992401, + 0.002917674370110035, + 0.0030523655004799366, + 0.0039926753379404545, + 0.0024432356003671885, + 0.016270112246274948, + 0.00697324937209487, + 0.008136662654578686, + 0.005543789826333523, + 0.010462966747581959, + 0.007448847405612469, + 0.005319798365235329, + 0.022383566945791245, + 0.009831845760345459, + 0.005108335055410862, + 0.004140304401516914, + 0.006660927087068558, + 0.004018299747258425, + 0.0060881273820996284, + 0.003637416986748576, + 0.01563701033592224, + 0.013559074141085148, + 0.004460408817976713, + 0.004682337399572134, + 0.004385706037282944, + 0.0055198948830366135, + 0.0036278623156249523, + 0.003677383530884981, + 0.015338760800659657, + 0.004674992058426142, + 0.00229022023268044, + 0.001701136352494359, + 0.002063007326796651, + 0.006306708790361881, + 0.003434665035456419, + 0.0019294635858386755, + 0.004413698799908161, + 0.007688960526138544, + 0.009157258085906506, + 0.009745577350258827, + 0.007917665876448154, + 0.010369238443672657, + 0.0092947818338871, + 0.00990605540573597, + 0.00936080887913704, + 0.017287112772464752, + 0.0136343352496624, + 0.015477180480957031, + 0.008670881390571594, + 0.01260832417756319, + 0.014168827794492245, + 0.010213780216872692, + 0.013061421923339367, + 0.007115927059203386, + 0.008775973692536354, + 0.007109635975211859, + 0.003935391549021006, + 0.00586300203576684, + 0.00538834510371089, + 0.006269469391554594, + 0.015765054151415825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0017296605510637164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.009792622178792953 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792622178792953, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792622178792953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3499594032764435, + 0.022289268672466278, + 0.09654281288385391, + 0.10426287353038788, + 0.09070857614278793, + 0.19384972751140594, + 0.019849883392453194, + 0.09339108318090439, + 0.010895010083913803, + 0.018251432105898857 + ], + "credit_map": [ + 0.01586288958787918, + 0.01402988564223051, + 0.14337730407714844, + 0.14042329788208008, + 0.009596058167517185, + 0.01045411080121994, + 0.008445612154901028, + 0.0077701956033706665, + 0.002179592614993453, + 0.0028756631072610617, + 0.0026203221641480923, + 0.0028649112209677696, + 0.004126328509300947, + 0.002743879798799753, + 0.0024709103163331747, + 0.0024076581466943026, + 0.008090526796877384, + 0.006123412400484085, + 0.0180138498544693, + 0.034204743802547455, + 0.008820469491183758, + 0.008284876123070717, + 0.00793782901018858, + 0.00506708724424243, + 0.017813844606280327, + 0.008086102083325386, + 0.010854378342628479, + 0.0065741329453885555, + 0.008209124207496643, + 0.01109148096293211, + 0.007203263230621815, + 0.03443052992224693, + 0.018812270835042, + 0.011158532463014126, + 0.005977510940283537, + 0.008117260411381721, + 0.01642392762005329, + 0.01085646916180849, + 0.0035504496190696955, + 0.01581212691962719, + 0.03953104466199875, + 0.017589863389730453, + 0.013953730463981628, + 0.011923952028155327, + 0.03777313232421875, + 0.025418471544981003, + 0.010346644558012486, + 0.03731287270784378, + 0.0017238641157746315, + 0.00215728348121047, + 0.003120650304481387, + 0.001904075383208692, + 0.002765126759186387, + 0.003561614314094186, + 0.0021943009924143553, + 0.0024229648988693953, + 0.011550740338861942, + 0.010070005431771278, + 0.011356114409863949, + 0.008171862922608852, + 0.012609077617526054, + 0.013625618070363998, + 0.014182512648403645, + 0.011825141496956348, + 0.001367699820548296, + 0.0014389383140951395, + 0.0012148594250902534, + 0.0011884626001119614, + 0.0011489889584481716, + 0.0016113321762531996, + 0.0016747209010645747, + 0.0012500062584877014, + 0.0022798320278525352, + 0.0028673268388956785, + 0.002251675818115473, + 0.0017784504452720284, + 0.0016312237130478024, + 0.001805253908969462, + 0.0028528801631182432, + 0.002784787444397807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.009792622178792953 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.005248339846730232 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248339846730232, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248339846730232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38608673214912415, + 0.025905271992087364, + 0.16665975749492645, + 0.06938307732343674, + 0.11752786487340927, + 0.04755950719118118, + 0.019021345302462578, + 0.10927467048168182, + 0.03194928914308548, + 0.026632510125637054 + ], + "credit_map": [ + 0.01806861348450184, + 0.012713245116174221, + 0.15888462960720062, + 0.15721170604228973, + 0.010369235649704933, + 0.009284128434956074, + 0.010421562008559704, + 0.009133604355156422, + 0.0030061202123761177, + 0.0027617933228611946, + 0.0036408593878149986, + 0.0024179667234420776, + 0.005508237052708864, + 0.0035133473575115204, + 0.002358432859182358, + 0.002698515309020877, + 0.010709269903600216, + 0.006731182336807251, + 0.05161099508404732, + 0.0665699914097786, + 0.005153143312782049, + 0.007328699808567762, + 0.011371255852282047, + 0.007185223046690226, + 0.011110849678516388, + 0.005837388336658478, + 0.0075356136076152325, + 0.003798393765464425, + 0.0061744386330246925, + 0.006725685205310583, + 0.004537293687462807, + 0.023663416504859924, + 0.02570643089711666, + 0.011909876018762589, + 0.007501035928726196, + 0.009236140176653862, + 0.01767740398645401, + 0.012952156364917755, + 0.004290757235139608, + 0.0282540712505579, + 0.0068261162377893925, + 0.005153345409780741, + 0.004195420071482658, + 0.0056263143196702, + 0.009335788898169994, + 0.00476338155567646, + 0.004473134409636259, + 0.0071860081516206264, + 0.0028368765488266945, + 0.0028780950233340263, + 0.0025336267426609993, + 0.0018276335904374719, + 0.0024374560452997684, + 0.0021711336448788643, + 0.0019226532895117998, + 0.0024138728622347116, + 0.012307509779930115, + 0.011366873048245907, + 0.014006130397319794, + 0.010829086415469646, + 0.012083562090992928, + 0.018178699538111687, + 0.016330545768141747, + 0.014172255992889404, + 0.004979928955435753, + 0.003749142400920391, + 0.005404375959187746, + 0.0028735625091940165, + 0.0034116357564926147, + 0.004236606881022453, + 0.0029692042153328657, + 0.00432483060285449, + 0.003156092017889023, + 0.00576160429045558, + 0.002429036423563957, + 0.0016641062684357166, + 0.002965907333418727, + 0.0017875536577776074, + 0.0030653339345008135, + 0.005802876781672239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.005248339846730232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.00863425899296999 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.00863425899296999, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863425899296999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2976619005203247, + 0.11812178045511246, + 0.07546961307525635, + 0.03636768460273743, + 0.12668034434318542, + 0.17969359457492828, + 0.032548997551202774, + 0.07097408920526505, + 0.02683323062956333, + 0.03564881905913353 + ], + "credit_map": [ + 0.015267852693796158, + 0.018386235460639, + 0.11194996535778046, + 0.11103694140911102, + 0.015357737429440022, + 0.010071220807731152, + 0.00823719147592783, + 0.007354728411883116, + 0.01158389262855053, + 0.01023934967815876, + 0.019443953409790993, + 0.013106856495141983, + 0.01680620014667511, + 0.025462524965405464, + 0.011357419192790985, + 0.010121582075953484, + 0.006636505946516991, + 0.005793929100036621, + 0.024771686643362045, + 0.02192353829741478, + 0.0038184639997780323, + 0.004185591824352741, + 0.0038543236441910267, + 0.004485569428652525, + 0.005656188819557428, + 0.0049062492325901985, + 0.0032316253054887056, + 0.002745457459241152, + 0.00287535204552114, + 0.00433763675391674, + 0.002336416393518448, + 0.010278753004968166, + 0.024916082620620728, + 0.012554731220006943, + 0.0073517197743058205, + 0.009853009134531021, + 0.02173573151230812, + 0.014237032271921635, + 0.0050236862152814865, + 0.03100835159420967, + 0.04890931770205498, + 0.014460437931120396, + 0.010250001214444637, + 0.00812726654112339, + 0.016855312511324883, + 0.019201312214136124, + 0.008535088039934635, + 0.05335483327507973, + 0.006011468358337879, + 0.003218915546312928, + 0.0024466966278851032, + 0.0022583885584026575, + 0.005920047406107187, + 0.004318443592637777, + 0.0021897335536777973, + 0.00618530111387372, + 0.01768779754638672, + 0.007062043063342571, + 0.005617148708552122, + 0.00508811604231596, + 0.005952155217528343, + 0.013474873267114162, + 0.006157777737826109, + 0.009934168308973312, + 0.004410686437040567, + 0.0032342025078833103, + 0.003299388103187084, + 0.0034397172275930643, + 0.0033636430744081736, + 0.0031892338301986456, + 0.0027308366261422634, + 0.003165521426126361, + 0.004440165590494871, + 0.00980463158339262, + 0.003245916683226824, + 0.0019797799177467823, + 0.0027527371421456337, + 0.003896185429766774, + 0.002775579923763871, + 0.006753821857273579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.00863425899296999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping shot begins at a very low angle, then smoothly pulls back and rises, ultimately transforming into a sweeping, wide-angle bird's-eye view. Against a dramatic dusk sky, the parched, cracked earth violently crumbles, and countless massive metal branches, gleaming with a cold, silver-gray sheen, burst forth with unstoppable force. These branches, bearing the marks of industrial welding and gently weathered textures, resemble living pythons, rapidly growing, interweaving, bending, and seamlessly welding together in a rhythm imbued with power and organic beauty. Low, golden sunlight paints sharp highlights on their curved surfaces and casts long, ever-shifting, dynamic shadows, creating a striking contrast against the sky, tinted in deep orange and indigo. As the camera zooms out, these seemingly disorganized metal structures gradually weave into a magnificent, precisely intricate grid, ultimately forming the iconic silhouette of the Beijing Bird's Nest Stadium across the vast plain. The entire process is imbued with a surreal atmosphere of creativity and awe-inspiring industrial aesthetic.", + "seed": 3084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3084_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": -0.01464073620736599, + "vlm_logprob": -0.01464073620736599, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01464073620736599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.02465827949345112 + ], + "opposite_logprob_tokens": [ + -3.899658203125 + ], + "logprob_reward": -0.02465827949345112, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02465827949345112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32993167638778687, + 0.28863218426704407, + 0.03493134677410126, + 0.041890572756528854, + 0.04949338734149933, + 0.06283477693796158, + 0.06221780925989151, + 0.02563016675412655, + 0.06975103169679642, + 0.03468701243400574 + ], + "credit_map": [ + 0.020600473508238792, + 0.023798249661922455, + 0.018372630700469017, + 0.09050045162439346, + 0.023085424676537514, + 0.10816241055727005, + 0.025260761380195618, + 0.020151278004050255, + 0.020478272810578346, + 0.02016001008450985, + 0.08551014959812164, + 0.015047905966639519, + 0.020214105024933815, + 0.09360964596271515, + 0.014698636718094349, + 0.01891346648335457, + 0.003625465091317892, + 0.0035792465787380934, + 0.003467423841357231, + 0.004595196805894375, + 0.003815330797806382, + 0.00422818586230278, + 0.003868680913001299, + 0.007751815486699343, + 0.004392474889755249, + 0.007144106086343527, + 0.006376378238201141, + 0.008308934979140759, + 0.003396378131583333, + 0.005516627803444862, + 0.002731889486312866, + 0.004023783374577761, + 0.005379526875913143, + 0.008610681630671024, + 0.010401539504528046, + 0.007160029374063015, + 0.0036170470993965864, + 0.005827769171446562, + 0.0032907607965171337, + 0.005206034518778324, + 0.013503899797797203, + 0.006823728326708078, + 0.0063960859552025795, + 0.005440550856292248, + 0.005868416279554367, + 0.013910307548940182, + 0.004758954979479313, + 0.006132832728326321, + 0.006282960996031761, + 0.009868738241493702, + 0.01653195545077324, + 0.007834501564502716, + 0.004729649517685175, + 0.006022968795150518, + 0.005222275387495756, + 0.00572475790977478, + 0.0022484634537249804, + 0.0026622856967151165, + 0.003045614343136549, + 0.0033954267855733633, + 0.004345875699073076, + 0.00428033946081996, + 0.0030163999181240797, + 0.0026357616297900677, + 0.006513714790344238, + 0.005019773729145527, + 0.006167492363601923, + 0.00756133021786809, + 0.008144332095980644, + 0.014596067368984222, + 0.01075028721243143, + 0.010998036712408066, + 0.0031209506560117006, + 0.004164388403296471, + 0.004973779898136854, + 0.004421512130647898, + 0.004148869309574366, + 0.003829994471743703, + 0.005254780873656273, + 0.0047727348282933235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02465827949345112 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.02636105753481388 + ], + "opposite_logprob_tokens": [ + -3.6513609886169434 + ], + "logprob_reward": -0.02636105753481388, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636105753481388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30123230814933777, + 0.2835760712623596, + 0.02974558435380459, + 0.03736855834722519, + 0.0525892935693264, + 0.05167793855071068, + 0.04666197672486305, + 0.05456792190670967, + 0.10330264270305634, + 0.03927768021821976 + ], + "credit_map": [ + 0.01604350656270981, + 0.015678655356168747, + 0.0107999537140131, + 0.09700474143028259, + 0.013652059249579906, + 0.11570306867361069, + 0.012863924726843834, + 0.01948641799390316, + 0.01662677340209484, + 0.022614050656557083, + 0.08562830835580826, + 0.013599571771919727, + 0.019213702529668808, + 0.09958450496196747, + 0.010856044478714466, + 0.015453125350177288, + 0.0035004448145627975, + 0.0029866353143006563, + 0.004936471581459045, + 0.0028575719334185123, + 0.0025017650332301855, + 0.0036970663350075483, + 0.002951335161924362, + 0.006314295344054699, + 0.00429078284651041, + 0.0046777268871665, + 0.008668697439134121, + 0.006889579351991415, + 0.0032789811957627535, + 0.0039033389184623957, + 0.002547054784372449, + 0.0031123925000429153, + 0.0070693655870854855, + 0.01107492670416832, + 0.004130535293370485, + 0.011647440493106842, + 0.004991021938621998, + 0.005608116742223501, + 0.003181807929649949, + 0.004886076785624027, + 0.0054097287356853485, + 0.004260135814547539, + 0.009193096309900284, + 0.0044309766963124275, + 0.006804690696299076, + 0.011194027960300446, + 0.004365224391222, + 0.006020058877766132, + 0.004302504938095808, + 0.006931767333298922, + 0.0113390376791358, + 0.005425563082098961, + 0.004834866151213646, + 0.0067069088108837605, + 0.0036226012744009495, + 0.003498729784041643, + 0.004924153909087181, + 0.007236515171825886, + 0.006282714661210775, + 0.007283366750925779, + 0.007285042200237513, + 0.007614362519234419, + 0.006512957625091076, + 0.007428806275129318, + 0.009432592429220676, + 0.009057669900357723, + 0.012696435675024986, + 0.012403825297951698, + 0.013745293021202087, + 0.013293500058352947, + 0.016851330175995827, + 0.015821995213627815, + 0.004300021566450596, + 0.004200482275336981, + 0.004086718428879976, + 0.005047537852078676, + 0.005495136138051748, + 0.0034878449514508247, + 0.00800033938139677, + 0.004659599624574184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.02636105753481388 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.0024946779012680054 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.0024946779012680054, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024946779012680054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3231271803379059, + 0.2429540455341339, + 0.07371094077825546, + 0.07422713935375214, + 0.03598891571164131, + 0.056207381188869476, + 0.05943822115659714, + 0.035358477383852005, + 0.055442363023757935, + 0.04354533180594444 + ], + "credit_map": [ + 0.015829358249902725, + 0.013803254812955856, + 0.012600351125001907, + 0.10795187205076218, + 0.0110059455037117, + 0.12651532888412476, + 0.016305549070239067, + 0.019115500152111053, + 0.014614112675189972, + 0.020218661054968834, + 0.07049736380577087, + 0.011957254260778427, + 0.015832792967557907, + 0.0871419683098793, + 0.010037851519882679, + 0.012654049322009087, + 0.0065611242316663265, + 0.008685235865414143, + 0.009858931414783001, + 0.009731501340866089, + 0.006409807130694389, + 0.007975716143846512, + 0.00815675500780344, + 0.016331862658262253, + 0.008287275210022926, + 0.008919812738895416, + 0.01108984649181366, + 0.0104903569445014, + 0.007976911030709743, + 0.008033816702663898, + 0.005700595676898956, + 0.013728521764278412, + 0.004522544797509909, + 0.005492659285664558, + 0.004214350134134293, + 0.005286055617034435, + 0.003498552832752466, + 0.006504096556454897, + 0.002520320937037468, + 0.003950337413698435, + 0.003946267534047365, + 0.005129927769303322, + 0.007118829060345888, + 0.006978990044444799, + 0.005591231863945723, + 0.011502617970108986, + 0.004759280011057854, + 0.011180240660905838, + 0.005469823256134987, + 0.00920167751610279, + 0.007301288656890392, + 0.007034446578472853, + 0.008535392582416534, + 0.008573651313781738, + 0.007524610497057438, + 0.005797333084046841, + 0.004165681544691324, + 0.004257225897163153, + 0.004103460349142551, + 0.005731978453695774, + 0.005379064008593559, + 0.0037202085368335247, + 0.004547515418380499, + 0.0034533434081822634, + 0.005649370141327381, + 0.004462990909814835, + 0.005973086692392826, + 0.006413196679204702, + 0.005986957810819149, + 0.01052976492792368, + 0.008820846676826477, + 0.007606154307723045, + 0.004417638294398785, + 0.005495517048984766, + 0.005606506485491991, + 0.007044212426990271, + 0.004172594286501408, + 0.0059225792065262794, + 0.005490418989211321, + 0.005395864136517048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0024946779012680054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.00151078961789608 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.00151078961789608, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00151078961789608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29284870624542236, + 0.06477522850036621, + 0.01649266667664051, + 0.03370243310928345, + 0.050193797796964645, + 0.21798363327980042, + 0.07939127087593079, + 0.0893952026963234, + 0.0861118882894516, + 0.06910518556833267 + ], + "credit_map": [ + 0.017383070662617683, + 0.018923092633485794, + 0.012165424413979053, + 0.08098104596138, + 0.018342399969697, + 0.10790836811065674, + 0.014377082698047161, + 0.022768229246139526, + 0.007038063369691372, + 0.007673982530832291, + 0.010409818030893803, + 0.005826372653245926, + 0.006279889959841967, + 0.014427678659558296, + 0.0073422836139798164, + 0.005777138751000166, + 0.0019626731518656015, + 0.0014188950881361961, + 0.0023119645193219185, + 0.0019058693433180451, + 0.0020947775337845087, + 0.0026339679025113583, + 0.0014809020794928074, + 0.0026836171746253967, + 0.0040788836777210236, + 0.00666443258523941, + 0.004603562410920858, + 0.003975418396294117, + 0.003550169290974736, + 0.0038896440528333187, + 0.002546314848586917, + 0.004394005984067917, + 0.006167721003293991, + 0.007283971179276705, + 0.01440919004380703, + 0.007576228119432926, + 0.0031354082748293877, + 0.004663882777094841, + 0.0024427957832813263, + 0.004514605738222599, + 0.043088894337415695, + 0.018387826159596443, + 0.012998008169233799, + 0.013942773453891277, + 0.02153494954109192, + 0.06334563344717026, + 0.015288186259567738, + 0.02939736098051071, + 0.007539866492152214, + 0.014510735869407654, + 0.009947291575372219, + 0.009716182015836239, + 0.010521496646106243, + 0.01080942153930664, + 0.008586334064602852, + 0.0077599408105015755, + 0.00620354013517499, + 0.009096124209463596, + 0.011713679879903793, + 0.009768424555659294, + 0.01622709073126316, + 0.01686786115169525, + 0.01029756385833025, + 0.009220918640494347, + 0.009030049666762352, + 0.009740903042256832, + 0.006827814504504204, + 0.008532334119081497, + 0.013158751651644707, + 0.01362190954387188, + 0.012270191684365273, + 0.012929936870932579, + 0.006293707061558962, + 0.0065728663466870785, + 0.00984048843383789, + 0.010398059152066708, + 0.006972063332796097, + 0.008210708387196064, + 0.011135187931358814, + 0.00968210306018591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.00151078961789608 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10055927187204361, + 0.27784475684165955, + 0.024984244257211685, + 0.032881081104278564, + 0.10448528826236725, + 0.17887826263904572, + 0.040594469755887985, + 0.05532677471637726, + 0.04552179574966431, + 0.1389240175485611 + ], + "credit_map": [ + 0.010311631485819817, + 0.016704771667718887, + 0.007383132819086313, + 0.01637181080877781, + 0.012865614145994186, + 0.012524565681815147, + 0.01117722224444151, + 0.013220516964793205, + 0.016076667234301567, + 0.02409149333834648, + 0.08044140785932541, + 0.013330680318176746, + 0.019584864377975464, + 0.09949108213186264, + 0.010650521144270897, + 0.014178040437400341, + 0.003606056096032262, + 0.0027050103526562452, + 0.0029483961407095194, + 0.0032373315189033747, + 0.0027419605758041143, + 0.0032257542479783297, + 0.0026124967262148857, + 0.003907238133251667, + 0.004066017922013998, + 0.0038866966497153044, + 0.005802778527140617, + 0.0036472671199589968, + 0.0044553703628480434, + 0.002724340185523033, + 0.003722423454746604, + 0.0045761847868561745, + 0.014618942514061928, + 0.013055972754955292, + 0.033202387392520905, + 0.018932342529296875, + 0.0054532201029360294, + 0.007242159452289343, + 0.005519968457520008, + 0.006460295524448156, + 0.01939406804740429, + 0.013898404315114021, + 0.06746367365121841, + 0.0186134185642004, + 0.011522471904754639, + 0.020766599103808403, + 0.011817428283393383, + 0.015402196906507015, + 0.005327507853507996, + 0.0062870136462152, + 0.007495904341340065, + 0.004233428742736578, + 0.0032199369743466377, + 0.005868809297680855, + 0.003755661193281412, + 0.00440620630979538, + 0.005372779909521341, + 0.005790163297206163, + 0.009056907147169113, + 0.00903568509966135, + 0.006704922299832106, + 0.006236975081264973, + 0.007662553805857897, + 0.005466788541525602, + 0.004763538483530283, + 0.005629475228488445, + 0.005983904469758272, + 0.00624585198238492, + 0.007058295886963606, + 0.004886001814156771, + 0.0063613480888307095, + 0.004593376535922289, + 0.017661990597844124, + 0.02083524875342846, + 0.018484050408005714, + 0.026198280975222588, + 0.009746675379574299, + 0.013191386125981808, + 0.016268467530608177, + 0.016537925228476524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.018178876489400864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping shot begins at a very low angle, then smoothly pulls back and rises, ultimately transforming into a sweeping, wide-angle bird's-eye view. Against a dramatic dusk sky, the parched, cracked earth violently crumbles, and countless massive metal branches, gleaming with a cold, silver-gray sheen, burst forth with unstoppable force. These branches, bearing the marks of industrial welding and gently weathered textures, resemble living pythons, rapidly growing, interweaving, bending, and seamlessly welding together in a rhythm imbued with power and organic beauty. Low, golden sunlight paints sharp highlights on their curved surfaces and casts long, ever-shifting, dynamic shadows, creating a striking contrast against the sky, tinted in deep orange and indigo. As the camera zooms out, these seemingly disorganized metal structures gradually weave into a magnificent, precisely intricate grid, ultimately forming the iconic silhouette of the Beijing Bird's Nest Stadium across the vast plain. The entire process is imbued with a surreal atmosphere of creativity and awe-inspiring industrial aesthetic.", + "seed": 3085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3085_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": -0.00949698556214571, + "vlm_logprob": -0.00949698556214571, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.00949698556214571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.019888972863554955 + ], + "opposite_logprob_tokens": [ + -4.019888877868652 + ], + "logprob_reward": -0.019888972863554955, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019888972863554955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06563471257686615, + 0.1305713802576065, + 0.19211120903491974, + 0.1523868441581726, + 0.052683427929878235, + 0.034421391785144806, + 0.15477131307125092, + 0.12432873994112015, + 0.07288292795419693, + 0.020208051428198814 + ], + "credit_map": [ + 0.008131562732160091, + 0.00792846642434597, + 0.006909811869263649, + 0.006769163999706507, + 0.006914188154041767, + 0.009938624687492847, + 0.009127901867032051, + 0.009914993308484554, + 0.013010069727897644, + 0.011072992347180843, + 0.031284086406230927, + 0.008153566159307957, + 0.016798965632915497, + 0.026347944512963295, + 0.009273627772927284, + 0.014630123041570187, + 0.011668561957776546, + 0.006698060315102339, + 0.06340859085321426, + 0.0098381657153368, + 0.009256464429199696, + 0.06176035851240158, + 0.012518638744950294, + 0.016962356865406036, + 0.020039383322000504, + 0.017980029806494713, + 0.022439848631620407, + 0.019063318148255348, + 0.020437125116586685, + 0.020685818046331406, + 0.015119212679564953, + 0.01662210375070572, + 0.0050612278282642365, + 0.004354897886514664, + 0.011737310327589512, + 0.004638425074517727, + 0.006493732798844576, + 0.012155815027654171, + 0.004832444246858358, + 0.0034095742739737034, + 0.0037299436517059803, + 0.0038564002607017756, + 0.004615382757037878, + 0.003994545433670282, + 0.0056307208724319935, + 0.00573031697422266, + 0.0037553736474364996, + 0.003108707955107093, + 0.00656426465138793, + 0.049426622688770294, + 0.005041018594056368, + 0.004619039595127106, + 0.062779001891613, + 0.012416251003742218, + 0.009218931198120117, + 0.004706194624304771, + 0.02757951058447361, + 0.01690128818154335, + 0.007809987757354975, + 0.012472099624574184, + 0.013580454513430595, + 0.02830597572028637, + 0.006917695514857769, + 0.01076172199100256, + 0.01086288969963789, + 0.011930827051401138, + 0.011437543667852879, + 0.00920300092548132, + 0.008059653453528881, + 0.008793103508651257, + 0.007260861806571484, + 0.005335050635039806, + 0.004321449436247349, + 0.0020118379034101963, + 0.002576307626441121, + 0.0019466738449409604, + 0.0028475176077336073, + 0.003394598374143243, + 0.0017771058483049273, + 0.0013325626496225595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.019888972863554955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.0024989587254822254 + ], + "opposite_logprob_tokens": [ + -6.002499103546143 + ], + "logprob_reward": -0.0024989587254822254, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024989587254822254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044907040894031525, + 0.07274837046861649, + 0.4859735667705536, + 0.05888188257813454, + 0.0745779350399971, + 0.038162097334861755, + 0.04286212846636772, + 0.07149048149585724, + 0.082410529255867, + 0.02798593044281006 + ], + "credit_map": [ + 0.00548000680282712, + 0.004722339101135731, + 0.003984556999057531, + 0.006225241348147392, + 0.0036865677684545517, + 0.006262920796871185, + 0.005641052033752203, + 0.008904355578124523, + 0.012371601536870003, + 0.009077368304133415, + 0.007852104492485523, + 0.005461241118609905, + 0.007739079650491476, + 0.01472367625683546, + 0.005930605344474316, + 0.00959269143640995, + 0.02823878638446331, + 0.01238963007926941, + 0.1899648755788803, + 0.02306198514997959, + 0.01643572375178337, + 0.18927165865898132, + 0.01439732313156128, + 0.012213579379022121, + 0.00628212746232748, + 0.009195096790790558, + 0.006393480580300093, + 0.005425718147307634, + 0.006741469260305166, + 0.01061240117996931, + 0.006676692981272936, + 0.007554891984909773, + 0.015043902210891247, + 0.00666213221848011, + 0.012144668027758598, + 0.007315243594348431, + 0.011388309299945831, + 0.009155218489468098, + 0.0052119046449661255, + 0.007656547240912914, + 0.0033145861234515905, + 0.0041359649039804935, + 0.004973259288817644, + 0.004058115184307098, + 0.009345894679427147, + 0.005729592405259609, + 0.004013382829725742, + 0.002591303549706936, + 0.004288386087864637, + 0.010706301778554916, + 0.00437668664380908, + 0.0033711108844727278, + 0.008828134275972843, + 0.004773966036736965, + 0.0032179835252463818, + 0.0032995573710650206, + 0.015112446621060371, + 0.01103740930557251, + 0.005931883119046688, + 0.009165257215499878, + 0.007139217108488083, + 0.013642309233546257, + 0.004352102056145668, + 0.005109855439513922, + 0.009245488792657852, + 0.012784394435584545, + 0.009352653287351131, + 0.008605766110122204, + 0.011539839208126068, + 0.01339514646679163, + 0.008932342752814293, + 0.008554904721677303, + 0.004157275892794132, + 0.0026785230729728937, + 0.0051610032096505165, + 0.0036638882011175156, + 0.0028063347563147545, + 0.003688907716423273, + 0.0030240719206631184, + 0.0028059238102287054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.0024989587254822254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.0031947072129696608 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031947072129696608, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031947072129696608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027772007510066032, + 0.03529389947652817, + 0.4957752227783203, + 0.08352596312761307, + 0.03674378991127014, + 0.09964903444051743, + 0.08081397414207458, + 0.03175010904669762, + 0.0831255167722702, + 0.025550412014126778 + ], + "credit_map": [ + 0.005063219927251339, + 0.004101232625544071, + 0.0022397232241928577, + 0.0030829838942736387, + 0.002923236694186926, + 0.003228268353268504, + 0.0027878438122570515, + 0.004345501307398081, + 0.0030196274165064096, + 0.0032742763869464397, + 0.005847023334354162, + 0.0025700596161186695, + 0.0038523508701473475, + 0.009701211005449295, + 0.003124107141047716, + 0.0039052427746355534, + 0.02943400852382183, + 0.01217349711805582, + 0.19330017268657684, + 0.022945204749703407, + 0.01706176996231079, + 0.192423477768898, + 0.015689486637711525, + 0.012747603468596935, + 0.011222682893276215, + 0.009658003225922585, + 0.011180344969034195, + 0.010680985637009144, + 0.008922086097300053, + 0.010532299987971783, + 0.010789874009788036, + 0.010539689101278782, + 0.005166794639080763, + 0.0040206024423241615, + 0.004212146159261465, + 0.005376663524657488, + 0.004566848278045654, + 0.0038694501854479313, + 0.005484366789460182, + 0.00404692068696022, + 0.00858375709503889, + 0.009585845284163952, + 0.011022940278053284, + 0.012093739584088326, + 0.017808042466640472, + 0.018992463126778603, + 0.012793513014912605, + 0.008768738247454166, + 0.0054322839714586735, + 0.022384073585271835, + 0.005748816765844822, + 0.0059180897660553455, + 0.02215573936700821, + 0.00828815158456564, + 0.004895116668194532, + 0.005991708487272263, + 0.0036019510589540005, + 0.004690270870923996, + 0.0031041805632412434, + 0.0045628976076841354, + 0.00295994128100574, + 0.006403845734894276, + 0.0029182368889451027, + 0.0035087852738797665, + 0.012814595364034176, + 0.01196092739701271, + 0.012013852596282959, + 0.011784439906477928, + 0.011228015646338463, + 0.006309466436505318, + 0.008355669677257538, + 0.00865855347365141, + 0.005004969425499439, + 0.002386429812759161, + 0.002908829366788268, + 0.002561843954026699, + 0.0034221012610942125, + 0.005347122438251972, + 0.002450596308335662, + 0.001468519913032651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.0031947072129696608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.0013329440262168646 + ], + "opposite_logprob_tokens": [ + -6.626332759857178 + ], + "logprob_reward": -0.0013329440262168646, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013329440262168646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029472563415765762, + 0.03503463417291641, + 0.5482447743415833, + 0.06531701982021332, + 0.02825886569917202, + 0.040450569242239, + 0.014461570419371128, + 0.04369216784834862, + 0.1320292204618454, + 0.06303855031728745 + ], + "credit_map": [ + 0.004611513111740351, + 0.0022596369963139296, + 0.003207557601854205, + 0.003052585991099477, + 0.0025216126814484596, + 0.004756866488605738, + 0.0036070419009774923, + 0.005455750040709972, + 0.00397648336365819, + 0.004215235821902752, + 0.004600065760314465, + 0.0029600937850773335, + 0.00515680480748415, + 0.005543910898268223, + 0.004061002749949694, + 0.004521040245890617, + 0.032284215092659, + 0.01421706285327673, + 0.2141251564025879, + 0.02647506259381771, + 0.018263572826981544, + 0.21401244401931763, + 0.01582436077296734, + 0.013042940758168697, + 0.007900652475655079, + 0.007041430566459894, + 0.009852257557213306, + 0.008646585047245026, + 0.008367474190890789, + 0.008928271010518074, + 0.007002652622759342, + 0.007577704731374979, + 0.003481421386823058, + 0.0021390540059655905, + 0.005789443384855986, + 0.004090374335646629, + 0.0029998652171343565, + 0.005406225100159645, + 0.0026022347155958414, + 0.0017502482514828444, + 0.003992572892457247, + 0.0046050287783145905, + 0.006035519298166037, + 0.006347723305225372, + 0.005511211697012186, + 0.00439688703045249, + 0.005934454035013914, + 0.003627172904089093, + 0.001680700108408928, + 0.0025526739191263914, + 0.0016272544162347913, + 0.00150757422670722, + 0.002289154566824436, + 0.0024051161017268896, + 0.0010593482293188572, + 0.0013397494331002235, + 0.006163861136883497, + 0.00771278515458107, + 0.004483829252421856, + 0.0077539780177176, + 0.004075506702065468, + 0.0057088457979261875, + 0.003931217361241579, + 0.0038621476851403713, + 0.017525658011436462, + 0.014932429417967796, + 0.022610902786254883, + 0.02074141800403595, + 0.012168949469923973, + 0.011869859881699085, + 0.015540202148258686, + 0.016639819368720055, + 0.018026404082775116, + 0.005646069068461657, + 0.007665074896067381, + 0.005566853564232588, + 0.00464343186467886, + 0.013490346260368824, + 0.004307860974222422, + 0.0036925170570611954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.0013329440262168646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.020569344982504845 + ], + "opposite_logprob_tokens": [ + -3.895569324493408 + ], + "logprob_reward": -0.020569344982504845, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020569344982504845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03433921933174133, + 0.062230803072452545, + 0.4292079210281372, + 0.08288755267858505, + 0.03487765043973923, + 0.04968715086579323, + 0.012083698995411396, + 0.09767656028270721, + 0.12716495990753174, + 0.06984452903270721 + ], + "credit_map": [ + 0.005783790722489357, + 0.0033065611496567726, + 0.002591828117147088, + 0.0040915957652032375, + 0.0029678221326321363, + 0.003893983317539096, + 0.005936982110142708, + 0.005766654387116432, + 0.005375765264034271, + 0.004451104439795017, + 0.01641005091369152, + 0.0028577123302966356, + 0.0046103945933282375, + 0.018095117062330246, + 0.004294595681130886, + 0.006136064883321524, + 0.02566566877067089, + 0.01159930694848299, + 0.16735222935676575, + 0.02062523551285267, + 0.014710364863276482, + 0.16674399375915527, + 0.012645351700484753, + 0.00986576545983553, + 0.011017709970474243, + 0.009517117403447628, + 0.012025834061205387, + 0.010935726575553417, + 0.010413041338324547, + 0.010487815365195274, + 0.009505140595138073, + 0.008985167369246483, + 0.007864694111049175, + 0.0034953842405229807, + 0.0063358694314956665, + 0.002817383036017418, + 0.0038081405218690634, + 0.0037861354649066925, + 0.003252152819186449, + 0.003517894074320793, + 0.004386434331536293, + 0.0058226268738508224, + 0.003435200545936823, + 0.004646490328013897, + 0.015328988432884216, + 0.007487007416784763, + 0.005652666557580233, + 0.002927737543359399, + 0.0012408894253894687, + 0.0025163854006677866, + 0.0009814505465328693, + 0.001020431867800653, + 0.0014270188985392451, + 0.002560851164162159, + 0.0008539025438949466, + 0.0014827693812549114, + 0.017284894362092018, + 0.0206285547465086, + 0.007904763333499432, + 0.012073180638253689, + 0.009980511851608753, + 0.01698215864598751, + 0.005768444389104843, + 0.007054048590362072, + 0.019921800121665, + 0.02342217043042183, + 0.014700095169246197, + 0.016156550496816635, + 0.0235956609249115, + 0.009965323843061924, + 0.010865297168493271, + 0.00853805523365736, + 0.015149841085076332, + 0.005453864112496376, + 0.007991131395101547, + 0.005048397928476334, + 0.008918749168515205, + 0.017741473391652107, + 0.004774278495460749, + 0.004766795784235001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.020569344982504845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping shot begins at a very low angle, then smoothly pulls back and rises, ultimately transforming into a sweeping, wide-angle bird's-eye view. Against a dramatic dusk sky, the parched, cracked earth violently crumbles, and countless massive metal branches, gleaming with a cold, silver-gray sheen, burst forth with unstoppable force. These branches, bearing the marks of industrial welding and gently weathered textures, resemble living pythons, rapidly growing, interweaving, bending, and seamlessly welding together in a rhythm imbued with power and organic beauty. Low, golden sunlight paints sharp highlights on their curved surfaces and casts long, ever-shifting, dynamic shadows, creating a striking contrast against the sky, tinted in deep orange and indigo. As the camera zooms out, these seemingly disorganized metal structures gradually weave into a magnificent, precisely intricate grid, ultimately forming the iconic silhouette of the Beijing Bird's Nest Stadium across the vast plain. The entire process is imbued with a surreal atmosphere of creativity and awe-inspiring industrial aesthetic.", + "seed": 3086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_7/index_1219_seed_3086_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": -0.043757249251939356, + "vlm_logprob": -0.043757249251939356, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.043757249251939356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.022806141525506973 + ], + "opposite_logprob_tokens": [ + -4.022806167602539 + ], + "logprob_reward": -0.022806141525506973, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022806141525506973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08530357480049133, + 0.02057480812072754, + 0.2790657877922058, + 0.08275340497493744, + 0.07889985293149948, + 0.2436269223690033, + 0.0885828286409378, + 0.07039512693881989, + 0.01245823036879301, + 0.03833947330713272 + ], + "credit_map": [ + 0.012172176502645016, + 0.016278794035315514, + 0.009655088186264038, + 0.006833348888903856, + 0.011004437692463398, + 0.0104141216725111, + 0.0075762271881103516, + 0.011369378305971622, + 0.0035760437604039907, + 0.0037657825741916895, + 0.002074094023555517, + 0.0014303862117230892, + 0.0021532345563173294, + 0.002510259859263897, + 0.0022197559010237455, + 0.002845252165570855, + 0.034914012998342514, + 0.03261604905128479, + 0.023265233263373375, + 0.04992406442761421, + 0.02385030873119831, + 0.024277757853269577, + 0.047442514449357986, + 0.04277585446834564, + 0.008937242440879345, + 0.011414283886551857, + 0.009477098472416401, + 0.011017944663763046, + 0.009296458214521408, + 0.00958174653351307, + 0.010777418501675129, + 0.012251211330294609, + 0.008965769782662392, + 0.014729739166796207, + 0.010160273872315884, + 0.009418105706572533, + 0.007918060757219791, + 0.0071297911927104, + 0.011129461228847504, + 0.009448652155697346, + 0.043635960668325424, + 0.030815256759524345, + 0.03377748280763626, + 0.025627991184592247, + 0.023370113223791122, + 0.024854106828570366, + 0.027195226401090622, + 0.034350767731666565, + 0.016562752425670624, + 0.0145422900095582, + 0.00953635573387146, + 0.008411144837737083, + 0.009121381677687168, + 0.009805574081838131, + 0.00900555495172739, + 0.01159777119755745, + 0.007073261309415102, + 0.014840623363852501, + 0.007773265242576599, + 0.012093638069927692, + 0.006717359647154808, + 0.007223215885460377, + 0.008794049732387066, + 0.005879716016352177, + 0.0016024884534999728, + 0.002531648613512516, + 0.000970665889326483, + 0.0016966122202575207, + 0.0012072913814336061, + 0.0013967369450256228, + 0.001634738058783114, + 0.001418047584593296, + 0.005720433313399553, + 0.005020714830607176, + 0.005189117975533009, + 0.005317172035574913, + 0.0032147567253559828, + 0.0036842983681708574, + 0.0048546697944402695, + 0.005338309332728386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.022806141525506973 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.0032086102291941643 + ], + "opposite_logprob_tokens": [ + -5.753208637237549 + ], + "logprob_reward": -0.0032086102291941643, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032086102291941643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1674443781375885, + 0.035296157002449036, + 0.0678781121969223, + 0.09906778484582901, + 0.07446184009313583, + 0.20607499778270721, + 0.13287755846977234, + 0.1450941115617752, + 0.03183466196060181, + 0.039970386773347855 + ], + "credit_map": [ + 0.02109149470925331, + 0.04112110286951065, + 0.013339654542505741, + 0.013395940884947777, + 0.022483373060822487, + 0.02287341095507145, + 0.013202082365751266, + 0.019937310367822647, + 0.0053143431432545185, + 0.004021175671368837, + 0.003687585238367319, + 0.0026461670640856028, + 0.0029531701002269983, + 0.003190315794199705, + 0.005875580944120884, + 0.007607819978147745, + 0.009555632248520851, + 0.009983736090362072, + 0.0053948285058140755, + 0.011582783423364162, + 0.0060695926658809185, + 0.006042906548827887, + 0.011709439568221569, + 0.007539188489317894, + 0.012385766953229904, + 0.011741477064788342, + 0.009897677227854729, + 0.014350579120218754, + 0.009208510629832745, + 0.010305109433829784, + 0.017435656860470772, + 0.01374301128089428, + 0.008864378556609154, + 0.012105470523238182, + 0.007803334388881922, + 0.013076473027467728, + 0.007657052483409643, + 0.006002902984619141, + 0.011374257504940033, + 0.007577971089631319, + 0.03644581884145737, + 0.028771474957466125, + 0.028970304876565933, + 0.021677343174815178, + 0.020014161244034767, + 0.020630747079849243, + 0.022052088752388954, + 0.0275130532681942, + 0.02469271421432495, + 0.01716996170580387, + 0.015329835005104542, + 0.012637902982532978, + 0.013029815629124641, + 0.015326965600252151, + 0.011912539601325989, + 0.022777825593948364, + 0.015098633244633675, + 0.028355129063129425, + 0.014206362888216972, + 0.02524382993578911, + 0.01252424530684948, + 0.014583094976842403, + 0.018831271678209305, + 0.016251541674137115, + 0.00657310476526618, + 0.0028902834746986628, + 0.003972978796809912, + 0.002423252910375595, + 0.0024353822227567434, + 0.0038640417624264956, + 0.004123049322515726, + 0.005552567075937986, + 0.007570880465209484, + 0.003936077002435923, + 0.007391524035483599, + 0.0038545734714716673, + 0.003186835441738367, + 0.0035347193479537964, + 0.0060183326713740826, + 0.004477447830140591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0032086102291941643 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.04866647347807884 + ], + "opposite_logprob_tokens": [ + -3.048666477203369 + ], + "logprob_reward": -0.04866647347807884, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866647347807884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17797549068927765, + 0.04824702814221382, + 0.40299928188323975, + 0.04500313848257065, + 0.07119167596101761, + 0.13923069834709167, + 0.05586247518658638, + 0.020862728357315063, + 0.0062284888699650764, + 0.03239903599023819 + ], + "credit_map": [ + 0.022948522120714188, + 0.041128333657979965, + 0.014212939888238907, + 0.01516204047948122, + 0.024140087887644768, + 0.02447649836540222, + 0.014561962336301804, + 0.021345095708966255, + 0.008861955255270004, + 0.010740815661847591, + 0.0029880397487431765, + 0.002756812609732151, + 0.005126600619405508, + 0.006700436118990183, + 0.00403624726459384, + 0.007036120165139437, + 0.046824775636196136, + 0.045992784202098846, + 0.034269779920578, + 0.07816258072853088, + 0.034148313105106354, + 0.03268561139702797, + 0.07296079397201538, + 0.05795465409755707, + 0.005472281016409397, + 0.0071629611775279045, + 0.00493873143568635, + 0.0063776494935154915, + 0.004462632816284895, + 0.004648658912628889, + 0.006633324548602104, + 0.005306899081915617, + 0.009248419664800167, + 0.011119878850877285, + 0.006878869142383337, + 0.0066038197837769985, + 0.01141760777682066, + 0.007700546178966761, + 0.008660669438540936, + 0.009561863727867603, + 0.01835630089044571, + 0.02463730238378048, + 0.018033422529697418, + 0.0205626729875803, + 0.01122710295021534, + 0.013119985349476337, + 0.015561826527118683, + 0.01773208938539028, + 0.012662326917052269, + 0.005708174780011177, + 0.006389897782355547, + 0.0030251890420913696, + 0.005434642545878887, + 0.006579909939318895, + 0.005453862715512514, + 0.010608474723994732, + 0.0025284839794039726, + 0.0032937952782958746, + 0.0024866524618119, + 0.0020023516844958067, + 0.003094147192314267, + 0.0019851799588650465, + 0.002675218740478158, + 0.002796898828819394, + 0.0011658591683954, + 0.0007615964277647436, + 0.0008586103795096278, + 0.0006285537383519113, + 0.0005507233436219394, + 0.0008351057767868042, + 0.0008096437086351216, + 0.0006183967925608158, + 0.004571280442178249, + 0.004344739019870758, + 0.0037224621046334505, + 0.004141598008573055, + 0.0029238408897072077, + 0.003429001895710826, + 0.00438610790297389, + 0.0048800078220665455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.04866647347807884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.0013340153964236379 + ], + "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": [ + 0.2931761145591736, + 0.019462786614894867, + 0.13534709811210632, + 0.06243029236793518, + 0.08951259404420853, + 0.172596737742424, + 0.07829754799604416, + 0.10993944853544235, + 0.014466124586760998, + 0.02477130852639675 + ], + "credit_map": [ + 0.03525960072875023, + 0.07273601740598679, + 0.024230701848864555, + 0.024328943341970444, + 0.04278891161084175, + 0.04179487004876137, + 0.021810663864016533, + 0.030226396396756172, + 0.004056629724800587, + 0.003631692146882415, + 0.0015284608816727996, + 0.001379326800815761, + 0.0018137417500838637, + 0.002252856269478798, + 0.0017912405310198665, + 0.0030088392086327076, + 0.01615460030734539, + 0.01624753139913082, + 0.01106882281601429, + 0.025251030921936035, + 0.012630099430680275, + 0.011386021040380001, + 0.02273566648364067, + 0.019873332232236862, + 0.006218269933015108, + 0.011275870725512505, + 0.008861047215759754, + 0.00666109099984169, + 0.007366794627159834, + 0.006455967202782631, + 0.008472424931824207, + 0.0071188281290233135, + 0.012882357463240623, + 0.018521137535572052, + 0.010969584807753563, + 0.007841642946004868, + 0.009274003095924854, + 0.00862591527402401, + 0.010029267519712448, + 0.011368685401976109, + 0.026041509583592415, + 0.02247260883450508, + 0.023870160803198814, + 0.025647785514593124, + 0.013456066139042377, + 0.01586444303393364, + 0.020311899483203888, + 0.0249322596937418, + 0.014495513401925564, + 0.010244050994515419, + 0.007427695207297802, + 0.0077922954224050045, + 0.006496873218566179, + 0.008149011060595512, + 0.007835615426301956, + 0.015856493264436722, + 0.021075721830129623, + 0.014685439877212048, + 0.011836702935397625, + 0.010448835790157318, + 0.010044928640127182, + 0.011078986339271069, + 0.011109868995845318, + 0.01965896598994732, + 0.0015589718241244555, + 0.001990199787542224, + 0.0019053282449021935, + 0.001527574029751122, + 0.0013265599263831973, + 0.0014432261232286692, + 0.002310844138264656, + 0.0024034215603023767, + 0.00455461023375392, + 0.0027694993186742067, + 0.004510150756686926, + 0.002687279600650072, + 0.0017384339589625597, + 0.002216637134552002, + 0.003038774710148573, + 0.0032559209503233433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0013340153964236379 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 very low camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera smoothly pull back and rise to transform into a wide-angle bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are massive metal branches visible that gleam with a cold, silver-gray sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show the metal branches weaving together to form the iconic silhouette of the Beijing Bird's Nest Stadium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the sky tinted in deep orange and indigo 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": [ + -0.14277100563049316 + ], + "opposite_logprob_tokens": [ + -2.017771005630493 + ], + "logprob_reward": -0.14277100563049316, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277100563049316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12825199961662292, + 0.02702747844159603, + 0.15033183991909027, + 0.038685090839862823, + 0.05521475523710251, + 0.3021184206008911, + 0.19074614346027374, + 0.07444644719362259, + 0.017785098403692245, + 0.015392794273793697 + ], + "credit_map": [ + 0.01671341434121132, + 0.0280087199062109, + 0.010632638819515705, + 0.011660417541861534, + 0.014795309863984585, + 0.01602357253432274, + 0.011915030889213085, + 0.01850287802517414, + 0.0025368048809468746, + 0.004371295217424631, + 0.003331046551465988, + 0.002117153024300933, + 0.002185339108109474, + 0.0026736357249319553, + 0.00427887262776494, + 0.0055333273485302925, + 0.016598088666796684, + 0.01617911458015442, + 0.01284053921699524, + 0.030459197238087654, + 0.012190772220492363, + 0.01198473758995533, + 0.02858138456940651, + 0.02149798721075058, + 0.0046940031461417675, + 0.006431709975004196, + 0.0037380401045084, + 0.00393800251185894, + 0.006857401691377163, + 0.005374337080866098, + 0.004650680813938379, + 0.0030009113252162933, + 0.005775246769189835, + 0.014070212841033936, + 0.006547754164785147, + 0.00566017534583807, + 0.005023283883929253, + 0.005693718325346708, + 0.006096578668802977, + 0.006347779184579849, + 0.04939957708120346, + 0.03293202444911003, + 0.041773341596126556, + 0.041122231632471085, + 0.025472022593021393, + 0.03021400049328804, + 0.038987550884485245, + 0.04221765324473381, + 0.05073614791035652, + 0.018413282930850983, + 0.01701471395790577, + 0.014678836800158024, + 0.019876671954989433, + 0.02113829366862774, + 0.01848878711462021, + 0.03039938025176525, + 0.009882807731628418, + 0.007278855890035629, + 0.008244425989687443, + 0.007518983446061611, + 0.006100671365857124, + 0.006063858978450298, + 0.010154632851481438, + 0.019202200695872307, + 0.0031264114659279585, + 0.0025458051823079586, + 0.0018115966813638806, + 0.0018661054782569408, + 0.0017228643409907818, + 0.0017183419549837708, + 0.002619207836687565, + 0.002374763134866953, + 0.0023693754337728024, + 0.0017694515408948064, + 0.0022463114000856876, + 0.0014493478229269385, + 0.0015926789492368698, + 0.0013543196255341172, + 0.0028186861891299486, + 0.001792621216736734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.14277100563049316 + } + ], + "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/0000058/group_8/index_tensor([6941, 6941])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_8/index_tensor([6941, 6941])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e5a629a8f094b083cb07f5c6f2f8fb53031010d1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_8/index_tensor([6941, 6941])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up shot from the passenger seat captures a figure driving a luxury car at night. The figure wears a meticulously tailored dark charcoal wool suit, its sharp lapels catching the soft light, paired with a crisp white shirt. Under dramatic, low-key lighting, their face is a study in black and white, with deep, enveloping shadows completely obscuring their eyes and upper features, revealing only a strong, determined jawline and taut lips in the intermittent glow of the dashboard. Their hands, focused on the leather-wrapped steering wheel, navigate the city with swift, precise movements. Through the side windows, the vibrant lights of the metropolis blur into long, dynamic streaks of neon red, electric blue, and warm yellow, painting fleeting, abstract patterns across the car's dark, polished interior and the sleeves of their suits. The scene, dominated by a cool, dark palette, punctuated by sharp, evocative highlights from the city outside, creates a tense, mysterious, and profound cinematic atmosphere, depicting a high-stakes journey through the heart of the night.", + "seed": 27109, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_8/index_6941_seed_27109_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": -0.05097449850291014, + "vlm_logprob": -0.05097449850291014, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.05097449850291014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07639268785715103 + ], + "opposite_logprob_tokens": [ + -3.451392650604248 + ], + "logprob_reward": -0.07639268785715103, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.07639268785715103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19917257130146027, + 0.08696260303258896, + 0.06680899858474731, + 0.07419104874134064, + 0.09065473079681396, + 0.10321702063083649, + 0.13525134325027466, + 0.08657214045524597, + 0.05402336269617081, + 0.1031462550163269 + ], + "credit_map": [ + 0.02741554193198681, + 0.026215890422463417, + 0.02640547975897789, + 0.02028072252869606, + 0.01872679591178894, + 0.0286562517285347, + 0.02277287468314171, + 0.028699003159999847, + 0.01319674402475357, + 0.008949981071054935, + 0.012908370234072208, + 0.009478277526795864, + 0.007583999074995518, + 0.00568140996620059, + 0.012303832918405533, + 0.016859987750649452, + 0.006429995410144329, + 0.007863357663154602, + 0.008593447506427765, + 0.010226213373243809, + 0.005185266956686974, + 0.005218283273279667, + 0.012186394073069096, + 0.011106039397418499, + 0.009807268157601357, + 0.005701073445379734, + 0.011320263147354126, + 0.006028673145920038, + 0.006034079007804394, + 0.009941140189766884, + 0.013939931988716125, + 0.011418621055781841, + 0.010343275032937527, + 0.008321111090481281, + 0.012555908411741257, + 0.010675424709916115, + 0.014717541635036469, + 0.00823661033064127, + 0.013069560751318932, + 0.012735297903418541, + 0.012242238968610764, + 0.013725150376558304, + 0.01429038681089878, + 0.009360319003462791, + 0.01047377660870552, + 0.006895736791193485, + 0.01823534443974495, + 0.01799406111240387, + 0.016352787613868713, + 0.018325241282582283, + 0.023616135120391846, + 0.013081393204629421, + 0.01810220628976822, + 0.011461363174021244, + 0.017047489061951637, + 0.017264725640416145, + 0.01162115577608347, + 0.011747046373784542, + 0.011752085760235786, + 0.007081866730004549, + 0.00566252414137125, + 0.007972907274961472, + 0.013759316876530647, + 0.01697523333132267, + 0.0051402258686721325, + 0.005488263908773661, + 0.007070740684866905, + 0.0083574578166008, + 0.005976507440209389, + 0.007498946040868759, + 0.006600421853363514, + 0.007890801876783371, + 0.013988500460982323, + 0.008966871537268162, + 0.015158651396632195, + 0.007010567933320999, + 0.009299359284341335, + 0.01379349920898676, + 0.019077397882938385, + 0.01585141383111477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.07639268785715103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14290885627269745 + ], + "opposite_logprob_tokens": [ + -2.017908811569214 + ], + "logprob_reward": -0.14290885627269745, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14290885627269745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3150550127029419, + 0.2022308111190796, + 0.03389871120452881, + 0.05390474200248718, + 0.07455252856016159, + 0.1471962332725525, + 0.1067836582660675, + 0.01670040935277939, + 0.02320285141468048, + 0.02647501789033413 + ], + "credit_map": [ + 0.03401505947113037, + 0.031599417328834534, + 0.04057041183114052, + 0.05411697179079056, + 0.04045837000012398, + 0.04609689861536026, + 0.033305034041404724, + 0.03489284962415695, + 0.02432401478290558, + 0.030154157429933548, + 0.03260762244462967, + 0.03709837421774864, + 0.014726183377206326, + 0.016955804079771042, + 0.026933778077363968, + 0.019430870190262794, + 0.00432655680924654, + 0.003337639616802335, + 0.004176254849880934, + 0.004360696766525507, + 0.003965227864682674, + 0.0032878220081329346, + 0.005597529001533985, + 0.004846981260925531, + 0.006797484587877989, + 0.00391036132350564, + 0.00791408121585846, + 0.004096693824976683, + 0.0054211001843214035, + 0.006370531860738993, + 0.012313113547861576, + 0.007081374526023865, + 0.009092417545616627, + 0.007310859393328428, + 0.01117477472871542, + 0.008755884133279324, + 0.008889921009540558, + 0.007817105390131474, + 0.010291233658790588, + 0.011220335029065609, + 0.02123856730759144, + 0.02022906020283699, + 0.02059825137257576, + 0.02670522965490818, + 0.013574421405792236, + 0.010031281039118767, + 0.015882305800914764, + 0.018937114626169205, + 0.009130130521953106, + 0.007577653042972088, + 0.0176306813955307, + 0.009994328022003174, + 0.009633088484406471, + 0.006515438202768564, + 0.021146204322576523, + 0.025156129151582718, + 0.003063915064558387, + 0.0021245782263576984, + 0.0022814087569713593, + 0.001970954705029726, + 0.0009753363556228578, + 0.0012767730513587594, + 0.0023481303360313177, + 0.002659313380718231, + 0.0031053361017256975, + 0.0028376709669828415, + 0.002895473735406995, + 0.003968951757997274, + 0.00211727200075984, + 0.003105706535279751, + 0.0021622946951538324, + 0.003010145388543606, + 0.0029985748697072268, + 0.002657847246155143, + 0.0032314362470060587, + 0.002767694415524602, + 0.002938962774351239, + 0.0033609759993851185, + 0.003330742008984089, + 0.005188784096390009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.14290885627269745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023312298581004143 + ], + "opposite_logprob_tokens": [ + -3.7733123302459717 + ], + "logprob_reward": -0.023312298581004143, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023312298581004143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21248623728752136, + 0.07654110342264175, + 0.051874011754989624, + 0.04622292518615723, + 0.14167381823062897, + 0.21902817487716675, + 0.10443832725286484, + 0.05078194662928581, + 0.033394910395145416, + 0.06355850398540497 + ], + "credit_map": [ + 0.0379590205848217, + 0.030781123787164688, + 0.03177886828780174, + 0.023603405803442, + 0.015231768600642681, + 0.030827119946479797, + 0.022645771503448486, + 0.019659150391817093, + 0.006988935638219118, + 0.010949468240141869, + 0.01152347307652235, + 0.01138277817517519, + 0.006708676926791668, + 0.006930553819984198, + 0.010646759532392025, + 0.011410454288125038, + 0.005757770501077175, + 0.006579109001904726, + 0.007924424484372139, + 0.00466904416680336, + 0.005159477237612009, + 0.004945436492562294, + 0.00840003415942192, + 0.008438716642558575, + 0.00596945034340024, + 0.008413786999881268, + 0.006832296960055828, + 0.00618117768317461, + 0.00335762114264071, + 0.004978520330041647, + 0.005503806751221418, + 0.004986263345927, + 0.01924283057451248, + 0.009045417420566082, + 0.01919524371623993, + 0.015640990808606148, + 0.0108256870880723, + 0.014429494738578796, + 0.026715131476521492, + 0.02657902240753174, + 0.03185625374317169, + 0.039139892905950546, + 0.02683994174003601, + 0.02448456920683384, + 0.028239956125617027, + 0.012200349010527134, + 0.02732575312256813, + 0.02894146740436554, + 0.011283698491752148, + 0.015467377379536629, + 0.017438100650906563, + 0.008632160723209381, + 0.009276878088712692, + 0.005571966990828514, + 0.019499504938721657, + 0.01726863905787468, + 0.005731169134378433, + 0.005726755131036043, + 0.007066452410072088, + 0.005109046120196581, + 0.0036245775409042835, + 0.003391845151782036, + 0.007949824444949627, + 0.01218227855861187, + 0.004383531399071217, + 0.004443638026714325, + 0.004524485673755407, + 0.006215136498212814, + 0.0019689088221639395, + 0.003311606589704752, + 0.004692137707024813, + 0.0038554666098207235, + 0.0072166272439062595, + 0.005806282162666321, + 0.009199514985084534, + 0.004098997917026281, + 0.00406549870967865, + 0.007511218078434467, + 0.012469151057302952, + 0.013191211968660355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.023312298581004143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2135116308927536, + 0.19101035594940186, + 0.07576422393321991, + 0.12266360223293304, + 0.10056443512439728, + 0.06418437510728836, + 0.06687464565038681, + 0.0551467128098011, + 0.02329266630113125, + 0.08698738366365433 + ], + "credit_map": [ + 0.023217802867293358, + 0.0232916921377182, + 0.03283417597413063, + 0.017347091808915138, + 0.02566581405699253, + 0.04575302451848984, + 0.021242819726467133, + 0.024159220978617668, + 0.019200976938009262, + 0.015054639428853989, + 0.02722162753343582, + 0.016557564958930016, + 0.021791210398077965, + 0.01901126652956009, + 0.027501258999109268, + 0.04467180743813515, + 0.009893655776977539, + 0.00660780118778348, + 0.009118218906223774, + 0.008199255913496017, + 0.008767122402787209, + 0.00845271348953247, + 0.013882409781217575, + 0.010843049734830856, + 0.013908213935792446, + 0.012730804271996021, + 0.017437392845749855, + 0.010127346031367779, + 0.01344175636768341, + 0.014564715325832367, + 0.02406187541782856, + 0.01639149896800518, + 0.0115448497235775, + 0.005742272827774286, + 0.015532085672020912, + 0.007402864750474691, + 0.029611296951770782, + 0.014162803068757057, + 0.009878849610686302, + 0.006689415313303471, + 0.007015362847596407, + 0.007467842195183039, + 0.008379965089261532, + 0.007460941560566425, + 0.008670450188219547, + 0.004804779775440693, + 0.010633829981088638, + 0.009751200675964355, + 0.006584972608834505, + 0.006465338170528412, + 0.009843407198786736, + 0.006957568693906069, + 0.008052753284573555, + 0.003823069855570793, + 0.012544077821075916, + 0.012603456154465675, + 0.007164571899920702, + 0.00463904207572341, + 0.0071802600286901, + 0.005743805784732103, + 0.00589166721329093, + 0.006423890125006437, + 0.009837700985372066, + 0.008265770971775055, + 0.0024356080684810877, + 0.0032669880893081427, + 0.0028673133347183466, + 0.0031840545125305653, + 0.0022542555816471577, + 0.0032489795703440905, + 0.0032364358194172382, + 0.0027990310918539762, + 0.011427734047174454, + 0.0064174416474998, + 0.014426734298467636, + 0.007225524168461561, + 0.00721686752513051, + 0.010187522508203983, + 0.016513431444764137, + 0.013572124764323235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.008639931678771973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036187181249260902 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036187181249260902, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036187181249260902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21223025023937225, + 0.08670379966497421, + 0.05347427725791931, + 0.04916692525148392, + 0.07252010703086853, + 0.16069307923316956, + 0.10649918019771576, + 0.049158792942762375, + 0.08583394438028336, + 0.12371967732906342 + ], + "credit_map": [ + 0.023693948984146118, + 0.030337246134877205, + 0.028204819187521935, + 0.022449204698204994, + 0.022972749546170235, + 0.02796105109155178, + 0.02495511993765831, + 0.03165613114833832, + 0.009627467952668667, + 0.011894996277987957, + 0.01346997357904911, + 0.010961046442389488, + 0.008599632419645786, + 0.007210347335785627, + 0.013493399135768414, + 0.011446942575275898, + 0.005548149812966585, + 0.00825498066842556, + 0.006482894532382488, + 0.005284806713461876, + 0.005815964192152023, + 0.0056777228601276875, + 0.009052462875843048, + 0.007357304450124502, + 0.005779698491096497, + 0.006368640810251236, + 0.007386667653918266, + 0.005447177682071924, + 0.004007372539490461, + 0.005996985360980034, + 0.008175287395715714, + 0.00600509624928236, + 0.007997014559805393, + 0.009098213165998459, + 0.009314760565757751, + 0.0066961864940822124, + 0.015154141932725906, + 0.005640495102852583, + 0.009260944090783596, + 0.009358357638120651, + 0.01937177963554859, + 0.02283257059752941, + 0.021560141816735268, + 0.02223086729645729, + 0.021862581372261047, + 0.01327161118388176, + 0.015923498198390007, + 0.023640045896172523, + 0.010371912270784378, + 0.010338954627513885, + 0.015288236550986767, + 0.009300721809267998, + 0.01251603476703167, + 0.005613451823592186, + 0.020045604556798935, + 0.02302427589893341, + 0.006095937918871641, + 0.0037235745694488287, + 0.00758745102211833, + 0.003619508119300008, + 0.0035651950165629387, + 0.005501187406480312, + 0.009254737757146358, + 0.00981120765209198, + 0.008600182831287384, + 0.008999629877507687, + 0.011158918030560017, + 0.011202756315469742, + 0.008201566524803638, + 0.012469630688428879, + 0.011536171659827232, + 0.0136650949716568, + 0.015486977063119411, + 0.00892389751970768, + 0.020564157515764236, + 0.011187057942152023, + 0.010290317237377167, + 0.015737751498818398, + 0.022217921912670135, + 0.019311605021357536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0036187181249260902 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up shot from the passenger seat captures a figure driving a luxury car at night. The figure wears a meticulously tailored dark charcoal wool suit, its sharp lapels catching the soft light, paired with a crisp white shirt. Under dramatic, low-key lighting, their face is a study in black and white, with deep, enveloping shadows completely obscuring their eyes and upper features, revealing only a strong, determined jawline and taut lips in the intermittent glow of the dashboard. Their hands, focused on the leather-wrapped steering wheel, navigate the city with swift, precise movements. Through the side windows, the vibrant lights of the metropolis blur into long, dynamic streaks of neon red, electric blue, and warm yellow, painting fleeting, abstract patterns across the car's dark, polished interior and the sleeves of their suits. The scene, dominated by a cool, dark palette, punctuated by sharp, evocative highlights from the city outside, creates a tense, mysterious, and profound cinematic atmosphere, depicting a high-stakes journey through the heart of the night.", + "seed": 27110, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_8/index_6941_seed_27110_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": -0.04744233367964625, + "vlm_logprob": -0.04744233367964625, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04744233367964625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09719972312450409 + ], + "opposite_logprob_tokens": [ + -2.9721996784210205 + ], + "logprob_reward": -0.09719972312450409, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.09719972312450409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14647436141967773, + 0.08489477634429932, + 0.08071134239435196, + 0.06130020692944527, + 0.08673835545778275, + 0.07883349806070328, + 0.09525129199028015, + 0.12571030855178833, + 0.18484148383140564, + 0.05524430796504021 + ], + "credit_map": [ + 0.01882149651646614, + 0.021846147254109383, + 0.018764495849609375, + 0.020860562101006508, + 0.015997618436813354, + 0.01442669052630663, + 0.015126102603971958, + 0.02063126675784588, + 0.01209540106356144, + 0.014133493416011333, + 0.011339914053678513, + 0.01178354024887085, + 0.005924377124756575, + 0.007704705931246281, + 0.008432022295892239, + 0.013481332920491695, + 0.014078759588301182, + 0.0169951394200325, + 0.014055460691452026, + 0.009650100953876972, + 0.005992078687995672, + 0.00602208124473691, + 0.006579791195690632, + 0.007337933871895075, + 0.007428336422890425, + 0.008195544593036175, + 0.007892892695963383, + 0.005720173008739948, + 0.006710351444780827, + 0.00657202024012804, + 0.009198951534926891, + 0.009581939317286015, + 0.010627448558807373, + 0.01163867861032486, + 0.011000420898199081, + 0.012569047510623932, + 0.01029353216290474, + 0.00868242233991623, + 0.009695879183709621, + 0.012230933643877506, + 0.008146665059030056, + 0.013978608883917332, + 0.012055585160851479, + 0.018351728096604347, + 0.008300022222101688, + 0.004827751312404871, + 0.007802463136613369, + 0.005370678845793009, + 0.008703058585524559, + 0.016380254179239273, + 0.014854952692985535, + 0.015252619981765747, + 0.006260282825678587, + 0.009274586103856564, + 0.013720051385462284, + 0.010805493220686913, + 0.020805073902010918, + 0.0172328669577837, + 0.01900695450603962, + 0.016121571883559227, + 0.008004477247595787, + 0.010038942098617554, + 0.016523586586117744, + 0.017976855859160423, + 0.019019436091184616, + 0.024583211168646812, + 0.01927538961172104, + 0.019633397459983826, + 0.02338751032948494, + 0.024044863879680634, + 0.026861760765314102, + 0.02803592011332512, + 0.00673058582469821, + 0.011603102087974548, + 0.006889836397022009, + 0.006815488450229168, + 0.004978544544428587, + 0.007717263884842396, + 0.005545458290725946, + 0.004964030813425779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.09719972312450409 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10042084008455276 + ], + "opposite_logprob_tokens": [ + -2.3504209518432617 + ], + "logprob_reward": -0.10042084008455276, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10042084008455276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17463597655296326, + 0.06787006556987762, + 0.03583988919854164, + 0.07427874207496643, + 0.0767165943980217, + 0.3543681204319, + 0.023148268461227417, + 0.09081452339887619, + 0.060419898480176926, + 0.04190802201628685 + ], + "credit_map": [ + 0.03236286714673042, + 0.027335701510310173, + 0.030826836824417114, + 0.03114275261759758, + 0.010113547556102276, + 0.008733321912586689, + 0.02101581171154976, + 0.013105113990604877, + 0.010657311417162418, + 0.010812308639287949, + 0.010645558126270771, + 0.010428877547383308, + 0.0044698952697217464, + 0.006480081472545862, + 0.006575869396328926, + 0.007800162769854069, + 0.0040602548979222775, + 0.005099508445709944, + 0.004419274628162384, + 0.004709278699010611, + 0.0034580270294100046, + 0.0056943087838590145, + 0.004183407872915268, + 0.004215827211737633, + 0.008946840651333332, + 0.01280848030000925, + 0.010956787504255772, + 0.010622244328260422, + 0.006479701027274132, + 0.0076588732190430164, + 0.00906551443040371, + 0.007740298751741648, + 0.006740462500602007, + 0.011564276181161404, + 0.009545153006911278, + 0.010469222441315651, + 0.014341413043439388, + 0.012162563391029835, + 0.006950340233743191, + 0.004943154752254486, + 0.06747248768806458, + 0.0704377144575119, + 0.06066565215587616, + 0.04254215955734253, + 0.03484880179166794, + 0.022522645071148872, + 0.033453647047281265, + 0.022424977272748947, + 0.0019470398547127843, + 0.005822988227009773, + 0.0040863254107534885, + 0.004402904771268368, + 0.0013737272238358855, + 0.0014425810659304261, + 0.00166843191254884, + 0.0024042692966759205, + 0.014520199038088322, + 0.012669192627072334, + 0.013681493699550629, + 0.00820881500840187, + 0.0058614714071154594, + 0.006853632628917694, + 0.012752027250826359, + 0.016267679631710052, + 0.007497583515942097, + 0.009108778089284897, + 0.008227032609283924, + 0.009877754375338554, + 0.0053680697456002235, + 0.007254447788000107, + 0.004202364012598991, + 0.008883866481482983, + 0.006087685935199261, + 0.007770885713398457, + 0.00475949514657259, + 0.006966809276491404, + 0.0028254955541342497, + 0.004792241845279932, + 0.005140021443367004, + 0.0035653829108923674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.10042084008455276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331276424229145 + ], + "opposite_logprob_tokens": [ + -3.77331280708313 + ], + "logprob_reward": -0.02331276424229145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331276424229145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15283040702342987, + 0.24098341166973114, + 0.055819232016801834, + 0.08234678953886032, + 0.0854705274105072, + 0.14116840064525604, + 0.05210842564702034, + 0.08202315121889114, + 0.04500998556613922, + 0.062239713966846466 + ], + "credit_map": [ + 0.02322746440768242, + 0.023879246786236763, + 0.0205320306122303, + 0.016899505630135536, + 0.008852849714457989, + 0.015249226242303848, + 0.023516252636909485, + 0.02067382074892521, + 0.045576710253953934, + 0.04446768760681152, + 0.036965712904930115, + 0.03355091065168381, + 0.015009026043117046, + 0.01503576710820198, + 0.02503819204866886, + 0.025339391082525253, + 0.0116218077018857, + 0.007254333235323429, + 0.008876927196979523, + 0.0048757451586425304, + 0.005156935658305883, + 0.005661115050315857, + 0.0059513733722269535, + 0.006420989520847797, + 0.008383397944271564, + 0.017786690965294838, + 0.011758483946323395, + 0.010592354461550713, + 0.007031997665762901, + 0.006440590135753155, + 0.010453426279127598, + 0.009899838827550411, + 0.01106085442006588, + 0.012474405579268932, + 0.011640549637377262, + 0.0076741185039281845, + 0.015234670601785183, + 0.011372397653758526, + 0.009281801991164684, + 0.006731717847287655, + 0.023410556837916374, + 0.024150418117642403, + 0.019485166296362877, + 0.02495989203453064, + 0.017358707264065742, + 0.01013697125017643, + 0.011814252473413944, + 0.00985241960734129, + 0.004553759936243296, + 0.009477864019572735, + 0.008121827617287636, + 0.007136233150959015, + 0.004493774380534887, + 0.005872225388884544, + 0.0066744666546583176, + 0.005778267979621887, + 0.008347980678081512, + 0.01957690715789795, + 0.012383541092276573, + 0.015120982192456722, + 0.0057833329774439335, + 0.00671526649966836, + 0.006788760423660278, + 0.007306376937776804, + 0.004006370902061462, + 0.0059811631217598915, + 0.005369716789573431, + 0.004161860793828964, + 0.004808763507753611, + 0.009454219602048397, + 0.004823830910027027, + 0.006404056213796139, + 0.008538241498172283, + 0.01139009092003107, + 0.008483141660690308, + 0.00897266250103712, + 0.005635594483464956, + 0.00712025398388505, + 0.007209000643342733, + 0.004890726413577795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.02331276424229145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643831126391888 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643831126391888, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643831126391888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1482486128807068, + 0.07968370616436005, + 0.06387212127447128, + 0.072671078145504, + 0.07020953297615051, + 0.0727144330739975, + 0.08212343603372574, + 0.286752849817276, + 0.07092418521642685, + 0.05280006676912308 + ], + "credit_map": [ + 0.02504146099090576, + 0.028028229251503944, + 0.021354462951421738, + 0.018814953044056892, + 0.012427245266735554, + 0.013119490817189217, + 0.01836545206606388, + 0.011097309179604053, + 0.013965310528874397, + 0.013667563907802105, + 0.010333471931517124, + 0.007548644207417965, + 0.006471685133874416, + 0.00958936382085085, + 0.007328169420361519, + 0.010779494419693947, + 0.007206758018583059, + 0.006667779758572578, + 0.007642494514584541, + 0.008510405197739601, + 0.005413278006017208, + 0.006569312419742346, + 0.010206500999629498, + 0.01165558397769928, + 0.0065326509065926075, + 0.009421785362064838, + 0.008950297720730305, + 0.006064625456929207, + 0.008403134532272816, + 0.008488167077302933, + 0.013162283226847649, + 0.01164812408387661, + 0.0077321031130850315, + 0.010008741170167923, + 0.009957154281437397, + 0.008723343722522259, + 0.009216942824423313, + 0.01329112984240055, + 0.004995820578187704, + 0.006284293718636036, + 0.009886208921670914, + 0.011146432720124722, + 0.010580038651823997, + 0.00754268653690815, + 0.010484817437827587, + 0.006911620497703552, + 0.009792963042855263, + 0.0063696629367768764, + 0.004950631409883499, + 0.010181333869695663, + 0.011355920694768429, + 0.007817289792001247, + 0.00613580783829093, + 0.010729944333434105, + 0.012370694428682327, + 0.018581809476017952, + 0.028649892657995224, + 0.02952428348362446, + 0.03517180681228638, + 0.02599349059164524, + 0.0331624336540699, + 0.0357065349817276, + 0.03406800329685211, + 0.0644763931632042, + 0.0075028385035693645, + 0.010762778110802174, + 0.008072890341281891, + 0.00780034763738513, + 0.011025982908904552, + 0.00809541903436184, + 0.009105910547077656, + 0.008558008819818497, + 0.006155161652714014, + 0.00656580226495862, + 0.005998772569000721, + 0.005471115466207266, + 0.005868425592780113, + 0.01129064243286848, + 0.004700160585343838, + 0.006749975960701704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.008643831126391888 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076345098204910755 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.0076345098204910755, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076345098204910755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1611630916595459, + 0.11229471117258072, + 0.04310571774840355, + 0.05554121360182762, + 0.23188941180706024, + 0.06607488542795181, + 0.05145144835114479, + 0.11494811624288559, + 0.08051794022321701, + 0.0830133855342865 + ], + "credit_map": [ + 0.026991086080670357, + 0.022481342777609825, + 0.02491476945579052, + 0.01774815283715725, + 0.01301228255033493, + 0.012995566241443157, + 0.02365185134112835, + 0.019368061795830727, + 0.02151576057076454, + 0.0182639267295599, + 0.01620360277593136, + 0.009963389486074448, + 0.009522955864667892, + 0.011040613055229187, + 0.011594993993639946, + 0.014189488254487514, + 0.005393549799919128, + 0.005743701476603746, + 0.005458960775285959, + 0.0042867413721978664, + 0.004140161443501711, + 0.005623765755444765, + 0.006217690650373697, + 0.006241153925657272, + 0.004316911567002535, + 0.007072398904711008, + 0.006775189656764269, + 0.006840160116553307, + 0.005740108899772167, + 0.00937837827950716, + 0.00669018505141139, + 0.00872789230197668, + 0.01883654296398163, + 0.029607700183987617, + 0.02908204309642315, + 0.013584905304014683, + 0.040691740810871124, + 0.051620930433273315, + 0.02084713615477085, + 0.027618443593382835, + 0.008497562259435654, + 0.011773734353482723, + 0.00877230055630207, + 0.011769542470574379, + 0.0077816215343773365, + 0.005611793138086796, + 0.006337426137179136, + 0.00553091149777174, + 0.004338585771620274, + 0.008228600025177002, + 0.008181406185030937, + 0.006976749747991562, + 0.00465570529922843, + 0.005667810793966055, + 0.0070007252506911755, + 0.00640187319368124, + 0.013061431236565113, + 0.010454920120537281, + 0.01520458236336708, + 0.010667200200259686, + 0.013246939517557621, + 0.011888894252479076, + 0.015880918130278587, + 0.02454325184226036, + 0.008786335587501526, + 0.009039506316184998, + 0.009120737202465534, + 0.006218902301043272, + 0.008970181457698345, + 0.014385060407221317, + 0.011163433082401752, + 0.012833797372877598, + 0.010659228079020977, + 0.011049683205783367, + 0.01025546807795763, + 0.007420243229717016, + 0.008358352817595005, + 0.018204977735877037, + 0.006456876173615456, + 0.010608567856252193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.0076345098204910755 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up shot from the passenger seat captures a figure driving a luxury car at night. The figure wears a meticulously tailored dark charcoal wool suit, its sharp lapels catching the soft light, paired with a crisp white shirt. Under dramatic, low-key lighting, their face is a study in black and white, with deep, enveloping shadows completely obscuring their eyes and upper features, revealing only a strong, determined jawline and taut lips in the intermittent glow of the dashboard. Their hands, focused on the leather-wrapped steering wheel, navigate the city with swift, precise movements. Through the side windows, the vibrant lights of the metropolis blur into long, dynamic streaks of neon red, electric blue, and warm yellow, painting fleeting, abstract patterns across the car's dark, polished interior and the sleeves of their suits. The scene, dominated by a cool, dark palette, punctuated by sharp, evocative highlights from the city outside, creates a tense, mysterious, and profound cinematic atmosphere, depicting a high-stakes journey through the heart of the night.", + "seed": 27111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_8/index_6941_seed_27111_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": -0.05058191046118736, + "vlm_logprob": -0.05058191046118736, + "vlm_yes_no_margin": 3.8499999761581423, + "vlm_token_credit": -0.05058191046118736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07650423049926758 + ], + "opposite_logprob_tokens": [ + -3.4515042304992676 + ], + "logprob_reward": -0.07650423049926758, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.07650423049926758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14133675396442413, + 0.08073454350233078, + 0.09729447215795517, + 0.10853023082017899, + 0.10363567620515823, + 0.09346029162406921, + 0.1310986429452896, + 0.05267520621418953, + 0.07582755386829376, + 0.11540666222572327 + ], + "credit_map": [ + 0.020576654002070427, + 0.021776745095849037, + 0.021961426362395287, + 0.01860293187201023, + 0.015109461732208729, + 0.012021977454423904, + 0.016115186735987663, + 0.015172365121543407, + 0.005813005845993757, + 0.014011242426931858, + 0.012629462406039238, + 0.010079653933644295, + 0.0066177742555737495, + 0.008205435238778591, + 0.009832901880145073, + 0.013545062392950058, + 0.01177271082997322, + 0.016080161556601524, + 0.011641710065305233, + 0.008812959305942059, + 0.013295144774019718, + 0.013121136464178562, + 0.009715709835290909, + 0.012854941189289093, + 0.009569531306624413, + 0.018929535523056984, + 0.01429701317101717, + 0.006101572886109352, + 0.015203372575342655, + 0.010095790028572083, + 0.022589262574911118, + 0.011744149960577488, + 0.01252303272485733, + 0.01854848489165306, + 0.01451744046062231, + 0.010806888341903687, + 0.01178835891187191, + 0.012455197982490063, + 0.01376987062394619, + 0.009226413443684578, + 0.015135126188397408, + 0.015796732157468796, + 0.015615749172866344, + 0.0076031992211937904, + 0.010025949217379093, + 0.009106128476560116, + 0.011426604352891445, + 0.008750805631279945, + 0.0112152686342597, + 0.020191263407468796, + 0.01932554505765438, + 0.012164723128080368, + 0.014285058714449406, + 0.02348097786307335, + 0.017653528600931168, + 0.012782267294824123, + 0.004594769328832626, + 0.007159051019698381, + 0.006501292809844017, + 0.0066012851893901825, + 0.005578391719609499, + 0.009281820617616177, + 0.006061671767383814, + 0.006896925158798695, + 0.005752106197178364, + 0.012129141949117184, + 0.013689243234694004, + 0.007136866915971041, + 0.01155826635658741, + 0.008551063947379589, + 0.007244111504405737, + 0.009766754694283009, + 0.011791899800300598, + 0.01828155852854252, + 0.015351560898125172, + 0.007903700694441795, + 0.015293119475245476, + 0.01823471486568451, + 0.016285834833979607, + 0.012264266610145569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.07650423049926758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14290916919708252 + ], + "opposite_logprob_tokens": [ + -2.017909049987793 + ], + "logprob_reward": -0.14290916919708252, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14290916919708252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25179585814476013, + 0.11748870462179184, + 0.15937650203704834, + 0.061651427298784256, + 0.13752812147140503, + 0.07959427684545517, + 0.040523361414670944, + 0.029832912608981133, + 0.06144373118877411, + 0.06076514720916748 + ], + "credit_map": [ + 0.036183297634124756, + 0.037752702832221985, + 0.03820418193936348, + 0.049839846789836884, + 0.02089763805270195, + 0.013944780454039574, + 0.03219874948263168, + 0.02277466282248497, + 0.01591663621366024, + 0.01828736625611782, + 0.01611150987446308, + 0.01690002717077732, + 0.011462853290140629, + 0.011952194385230541, + 0.015776261687278748, + 0.011081848293542862, + 0.032045233994722366, + 0.027519550174474716, + 0.02605772763490677, + 0.018671201542019844, + 0.013292391784489155, + 0.0129727553576231, + 0.013738158158957958, + 0.015079488977789879, + 0.007256639655679464, + 0.009483719244599342, + 0.00982044730335474, + 0.005583035293966532, + 0.006799309980124235, + 0.005953522864729166, + 0.008879516273736954, + 0.007875232957303524, + 0.016253920271992683, + 0.023942535743117332, + 0.016754858195781708, + 0.010716010816395283, + 0.011225415393710136, + 0.01962934248149395, + 0.0195084847509861, + 0.019497554749250412, + 0.008790618740022182, + 0.013077306561172009, + 0.0103036193177104, + 0.0070377555675804615, + 0.007825626991689205, + 0.011473504826426506, + 0.012887772172689438, + 0.008198064751923084, + 0.004078923724591732, + 0.006913180463016033, + 0.0050826105289161205, + 0.005371631123125553, + 0.004005731083452702, + 0.006641273852437735, + 0.00369667774066329, + 0.004733331501483917, + 0.004151659086346626, + 0.0052842432633042336, + 0.003586596343666315, + 0.002752982545644045, + 0.003148116171360016, + 0.004468580707907677, + 0.002867362694814801, + 0.003573368536308408, + 0.004761022049933672, + 0.009923802688717842, + 0.008512546308338642, + 0.006657236255705357, + 0.010975814424455166, + 0.007501604501157999, + 0.005756026599556208, + 0.007355674169957638, + 0.009056391194462776, + 0.009422017261385918, + 0.007075884845107794, + 0.012120690196752548, + 0.006606295704841614, + 0.005679085850715637, + 0.005490439012646675, + 0.005314341280609369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.14290916919708252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020610684528946877 + ], + "opposite_logprob_tokens": [ + -3.8956105709075928 + ], + "logprob_reward": -0.020610684528946877, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020610684528946877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16537952423095703, + 0.11704610288143158, + 0.18399320542812347, + 0.05099990963935852, + 0.14370916783809662, + 0.06859083473682404, + 0.06395884603261948, + 0.04575558006763458, + 0.0652499571442604, + 0.09531692415475845 + ], + "credit_map": [ + 0.0173158161342144, + 0.028680596500635147, + 0.02673736773431301, + 0.01760333962738514, + 0.018943527713418007, + 0.013876662589609623, + 0.028270484879612923, + 0.013951731845736504, + 0.01543921697884798, + 0.018662847578525543, + 0.019076047465205193, + 0.01306410413235426, + 0.010483664460480213, + 0.012753262184560299, + 0.014634830877184868, + 0.012932134792208672, + 0.036342889070510864, + 0.03128693252801895, + 0.029113803058862686, + 0.02224881574511528, + 0.0159712266176939, + 0.013909109868109226, + 0.017184648662805557, + 0.01793578453361988, + 0.006129380781203508, + 0.008139794692397118, + 0.007538781501352787, + 0.006204409524798393, + 0.00457058334723115, + 0.005113703198730946, + 0.007681507151573896, + 0.005621748510748148, + 0.01287190429866314, + 0.025890585035085678, + 0.019838374108076096, + 0.012847919017076492, + 0.01040545105934143, + 0.01946891099214554, + 0.02278991788625717, + 0.01959610916674137, + 0.007750443648546934, + 0.010945606045424938, + 0.007945714518427849, + 0.005667531862854958, + 0.006686302833259106, + 0.010744352824985981, + 0.00871436670422554, + 0.010136517696082592, + 0.005204361397773027, + 0.009899863041937351, + 0.007325184065848589, + 0.008046071045100689, + 0.007391282357275486, + 0.011434735730290413, + 0.007706325501203537, + 0.00695102009922266, + 0.008192088454961777, + 0.007759640924632549, + 0.00581559119746089, + 0.003100285306572914, + 0.0048842099495232105, + 0.006201271899044514, + 0.004709509667009115, + 0.005092985462397337, + 0.007579050958156586, + 0.011273128911852837, + 0.010009968653321266, + 0.006669813767075539, + 0.01110800914466381, + 0.007099982816725969, + 0.005604890175163746, + 0.0059051113203167915, + 0.01722436398267746, + 0.014928971417248249, + 0.012205422855913639, + 0.012347199022769928, + 0.00839695893228054, + 0.010651187039911747, + 0.008547193370759487, + 0.011015627533197403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.020610684528946877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11807361245155334, + 0.2741820514202118, + 0.053463224321603775, + 0.1832229495048523, + 0.08888797461986542, + 0.14541172981262207, + 0.02461823634803295, + 0.035153910517692566, + 0.046862225979566574, + 0.03012404777109623 + ], + "credit_map": [ + 0.012200599536299706, + 0.017779136076569557, + 0.019184943288564682, + 0.012302654795348644, + 0.015984348952770233, + 0.01188221201300621, + 0.01825980469584465, + 0.010479914024472237, + 0.022229116410017014, + 0.0440903976559639, + 0.0441378615796566, + 0.03552839905023575, + 0.02185542695224285, + 0.04107721522450447, + 0.03613358736038208, + 0.02913004904985428, + 0.007073031738400459, + 0.008839882910251617, + 0.006779724266380072, + 0.003707497613504529, + 0.007527712732553482, + 0.007859287783503532, + 0.005679334048181772, + 0.005996761377900839, + 0.015164529904723167, + 0.034889668226242065, + 0.021745754405856133, + 0.010729116387665272, + 0.02335982210934162, + 0.01756780408322811, + 0.039797328412532806, + 0.01996893249452114, + 0.013862165622413158, + 0.013858044520020485, + 0.014509147964417934, + 0.0061841304413974285, + 0.010662723332643509, + 0.011094555258750916, + 0.010923509486019611, + 0.007793700788170099, + 0.010150592774152756, + 0.022609027102589607, + 0.021272040903568268, + 0.01025319378823042, + 0.013525525107979774, + 0.015850476920604706, + 0.028782518580555916, + 0.022968366742134094, + 0.001971772639080882, + 0.003650569124147296, + 0.0027870391495525837, + 0.0028542415238916874, + 0.0025173479225486517, + 0.004560607485473156, + 0.0035943842958658934, + 0.002682275837287307, + 0.0030856661032885313, + 0.00496145011857152, + 0.0034642203245311975, + 0.004175035282969475, + 0.004459231626242399, + 0.0056280940771102905, + 0.004480808041989803, + 0.0048994095996022224, + 0.003978680353611708, + 0.008020113222301006, + 0.005911614745855331, + 0.004662311635911465, + 0.00836494006216526, + 0.006773003377020359, + 0.004522508941590786, + 0.004629055038094521, + 0.0028460074681788683, + 0.005242259241640568, + 0.004272582475095987, + 0.003362201852723956, + 0.004308825358748436, + 0.003937327768653631, + 0.003274683840572834, + 0.0028801627922803164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "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 shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17514102160930634, + 0.14970581233501434, + 0.06049694865942001, + 0.11364264786243439, + 0.17666998505592346, + 0.06833753734827042, + 0.035907451063394547, + 0.03629562258720398, + 0.08090565353631973, + 0.1028972864151001 + ], + "credit_map": [ + 0.02084549516439438, + 0.026566194370388985, + 0.030028311535716057, + 0.018903348594903946, + 0.02303265407681465, + 0.015992959961295128, + 0.025172222405672073, + 0.014599834568798542, + 0.014828735031187534, + 0.024771368131041527, + 0.024765465408563614, + 0.01690986193716526, + 0.01409859023988247, + 0.01858244650065899, + 0.015565862879157066, + 0.020183496177196503, + 0.009584330022335052, + 0.009408907033503056, + 0.008471670560538769, + 0.005473215598613024, + 0.007233743090182543, + 0.006188422907143831, + 0.007002063095569611, + 0.007134592160582542, + 0.011845138855278492, + 0.021354231983423233, + 0.01652645505964756, + 0.0063780322670936584, + 0.014838844537734985, + 0.010948770679533482, + 0.020732209086418152, + 0.011018958874046803, + 0.014227945357561111, + 0.03180928900837898, + 0.019884880632162094, + 0.017923368141055107, + 0.015543489716947079, + 0.024558570235967636, + 0.026926154270768166, + 0.02579628862440586, + 0.010976334102451801, + 0.011580360122025013, + 0.010564561933279037, + 0.00475018285214901, + 0.0061964476481080055, + 0.008780192583799362, + 0.008755598217248917, + 0.006733858957886696, + 0.0034721747506409883, + 0.0065663885325193405, + 0.003643113188445568, + 0.003790823509916663, + 0.0033190140966326, + 0.006945929490029812, + 0.003621203126385808, + 0.0045488011091947556, + 0.005517712794244289, + 0.004916261415928602, + 0.004718376789242029, + 0.005038960371166468, + 0.003011382184922695, + 0.004787416197359562, + 0.004810291342437267, + 0.0034952214919030666, + 0.0074379658326506615, + 0.014051228761672974, + 0.016028128564357758, + 0.005519051104784012, + 0.011851703748106956, + 0.0080645140260458, + 0.007799315731972456, + 0.010153745301067829, + 0.01221380289644003, + 0.018475545570254326, + 0.013990760780870914, + 0.011295373551547527, + 0.01488895621150732, + 0.013109304942190647, + 0.010776270180940628, + 0.008147268556058407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.005255098920315504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up shot from the passenger seat captures a figure driving a luxury car at night. The figure wears a meticulously tailored dark charcoal wool suit, its sharp lapels catching the soft light, paired with a crisp white shirt. Under dramatic, low-key lighting, their face is a study in black and white, with deep, enveloping shadows completely obscuring their eyes and upper features, revealing only a strong, determined jawline and taut lips in the intermittent glow of the dashboard. Their hands, focused on the leather-wrapped steering wheel, navigate the city with swift, precise movements. Through the side windows, the vibrant lights of the metropolis blur into long, dynamic streaks of neon red, electric blue, and warm yellow, painting fleeting, abstract patterns across the car's dark, polished interior and the sleeves of their suits. The scene, dominated by a cool, dark palette, punctuated by sharp, evocative highlights from the city outside, creates a tense, mysterious, and profound cinematic atmosphere, depicting a high-stakes journey through the heart of the night.", + "seed": 27112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_8/index_6941_seed_27112_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": -0.0484712403267622, + "vlm_logprob": -0.0484712403267622, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.0484712403267622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05981917306780815 + ], + "opposite_logprob_tokens": [ + -3.684819221496582 + ], + "logprob_reward": -0.05981917306780815, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05981917306780815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11185849457979202, + 0.07468506693840027, + 0.0855344757437706, + 0.10501968860626221, + 0.03668694943189621, + 0.05157328024506569, + 0.07487314939498901, + 0.030733725056052208, + 0.049176305532455444, + 0.3798588216304779 + ], + "credit_map": [ + 0.014528683386743069, + 0.012049581855535507, + 0.021398279815912247, + 0.009924962185323238, + 0.014918208122253418, + 0.01579871214926243, + 0.00854468997567892, + 0.014695392921566963, + 0.008799142204225063, + 0.00803480762988329, + 0.018894705921411514, + 0.005661974661052227, + 0.0071214912459254265, + 0.009971195831894875, + 0.007526658475399017, + 0.008675097487866879, + 0.017435133457183838, + 0.008611983619630337, + 0.008673322387039661, + 0.00722289364784956, + 0.009100307710468769, + 0.008574938401579857, + 0.008767432533204556, + 0.017148463055491447, + 0.01752178743481636, + 0.008720302022993565, + 0.010630464181303978, + 0.006997056305408478, + 0.03423008695244789, + 0.00657817954197526, + 0.012411621399223804, + 0.007930195890367031, + 0.0051835766062140465, + 0.004419959615916014, + 0.007120066788047552, + 0.003445833455771208, + 0.005237535573542118, + 0.0034791480284184217, + 0.0038369297981262207, + 0.003963903523981571, + 0.0063223522156476974, + 0.008631735108792782, + 0.007960284128785133, + 0.005112693645060062, + 0.007598052732646465, + 0.004528266377747059, + 0.005121881607919931, + 0.006298013497143984, + 0.015110878273844719, + 0.008413447998464108, + 0.016867822036147118, + 0.005634736735373735, + 0.007800622843205929, + 0.0068604047410190105, + 0.0060974913649261, + 0.008087744005024433, + 0.003137830877676606, + 0.004413594491779804, + 0.004888707306236029, + 0.002792161423712969, + 0.004382346756756306, + 0.002367571461945772, + 0.0038557820953428745, + 0.004895731806755066, + 0.0034534484148025513, + 0.005504722241312265, + 0.006199484225362539, + 0.003491050796583295, + 0.01081797108054161, + 0.005952322855591774, + 0.006439581047743559, + 0.007317729759961367, + 0.01844308339059353, + 0.009768419899046421, + 0.014602307230234146, + 0.008723950013518333, + 0.024994097650051117, + 0.02101476676762104, + 0.14456303417682648, + 0.13774916529655457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.05981917306780815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16043217480182648 + ], + "opposite_logprob_tokens": [ + -1.91043221950531 + ], + "logprob_reward": -0.16043217480182648, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16043217480182648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15426906943321228, + 0.04271307960152626, + 0.023771747946739197, + 0.11680183559656143, + 0.05472416803240776, + 0.028513547033071518, + 0.020117221400141716, + 0.024741536006331444, + 0.025944553315639496, + 0.5084033012390137 + ], + "credit_map": [ + 0.013982349075376987, + 0.031191108748316765, + 0.01338551938533783, + 0.01831892691552639, + 0.023105544969439507, + 0.02077523060142994, + 0.014525413513183594, + 0.01898498274385929, + 0.002389667322859168, + 0.0032387280371040106, + 0.0040110149420797825, + 0.00347713241353631, + 0.011601140722632408, + 0.00490791629999876, + 0.006887736730277538, + 0.0061997417360544205, + 0.0028180198278278112, + 0.0035411431454122066, + 0.002858631080016494, + 0.00204003369435668, + 0.0036687199026346207, + 0.0025251545011997223, + 0.0032242496963590384, + 0.00309579586610198, + 0.018776560202240944, + 0.010209772735834122, + 0.01520015113055706, + 0.009744503535330296, + 0.02890855260193348, + 0.007446283474564552, + 0.017565075308084488, + 0.008950941264629364, + 0.005930076353251934, + 0.01097237877547741, + 0.0061881933361291885, + 0.005652878433465958, + 0.008569915778934956, + 0.005230351816862822, + 0.006423411890864372, + 0.005756961647421122, + 0.005270438734441996, + 0.0034899860620498657, + 0.003726583905518055, + 0.0024197841994464397, + 0.003625729586929083, + 0.00202755955979228, + 0.004049631766974926, + 0.0039038320537656546, + 0.0018676979234442115, + 0.0023997114039957523, + 0.0030202888883650303, + 0.0019768066704273224, + 0.0028463148046284914, + 0.0031613847240805626, + 0.002147757215425372, + 0.0026972610503435135, + 0.0012333276681602001, + 0.0034149703569710255, + 0.0032942909747362137, + 0.0017662291647866368, + 0.003833578899502754, + 0.004959778394550085, + 0.0030336976051330566, + 0.003205664921551943, + 0.0029020782094448805, + 0.0034266035072505474, + 0.002283685840666294, + 0.0011958470568060875, + 0.006483409553766251, + 0.0037268989253789186, + 0.0023713300470262766, + 0.003554701805114746, + 0.02299175038933754, + 0.012569287791848183, + 0.019681068137288094, + 0.011740854009985924, + 0.03496674448251724, + 0.024310216307640076, + 0.2037586271762848, + 0.1783847212791443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.16043217480182648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016090890392661095 + ], + "opposite_logprob_tokens": [ + -4.1410908699035645 + ], + "logprob_reward": -0.016090890392661095, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016090890392661095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14739395678043365, + 0.08615870028734207, + 0.06117312237620354, + 0.10703527182340622, + 0.21156664192676544, + 0.06133592501282692, + 0.07969436049461365, + 0.04213929921388626, + 0.05671784654259682, + 0.14678482711315155 + ], + "credit_map": [ + 0.012964618392288685, + 0.015327762812376022, + 0.022598950192332268, + 0.016121624037623405, + 0.022514361888170242, + 0.02131749503314495, + 0.01664244569838047, + 0.01990671083331108, + 0.008992616087198257, + 0.008340290747582912, + 0.019512593746185303, + 0.007459426298737526, + 0.014018181711435318, + 0.010230915620923042, + 0.00858675129711628, + 0.009017935022711754, + 0.0130575867369771, + 0.006927164271473885, + 0.006891048979014158, + 0.004192865453660488, + 0.007790733594447374, + 0.00650448864325881, + 0.005596323870122433, + 0.010212916880846024, + 0.016265077516436577, + 0.00925180409103632, + 0.01538227777928114, + 0.00807010568678379, + 0.028073545545339584, + 0.00775342620909214, + 0.01344921812415123, + 0.008789828978478909, + 0.01998366415500641, + 0.04573990777134895, + 0.018547674641013145, + 0.017162805423140526, + 0.02830681763589382, + 0.030853236094117165, + 0.01934688538312912, + 0.0316256582736969, + 0.009812627919018269, + 0.009651255793869495, + 0.00768354581668973, + 0.004625848960131407, + 0.009085710160434246, + 0.005202791653573513, + 0.0067016370594501495, + 0.008572514168918133, + 0.01220749132335186, + 0.01065914612263441, + 0.012769337743520737, + 0.00567113421857357, + 0.00892731174826622, + 0.010252653621137142, + 0.009058702737092972, + 0.010148581117391586, + 0.00245730415917933, + 0.005126759875565767, + 0.005262564867734909, + 0.0028624830301851034, + 0.0058503225445747375, + 0.008786198683083057, + 0.005691489204764366, + 0.006102181039750576, + 0.003964757546782494, + 0.004625030793249607, + 0.007627465296536684, + 0.0030296912882477045, + 0.01103995181620121, + 0.00856996513903141, + 0.008595401421189308, + 0.009265592321753502, + 0.01057074498385191, + 0.006400846876204014, + 0.010930529795587063, + 0.004720116499811411, + 0.013715785928070545, + 0.010144638828933239, + 0.042108841240406036, + 0.04819333925843239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.016090890392661095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820206806063652 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820206806063652, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820206806063652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12929105758666992, + 0.0759628415107727, + 0.11027301847934723, + 0.0397040881216526, + 0.0522785447537899, + 0.03824247792363167, + 0.10468916594982147, + 0.05071446672081947, + 0.18086139857769012, + 0.2179829478263855 + ], + "credit_map": [ + 0.012399844825267792, + 0.019591405987739563, + 0.017870085313916206, + 0.016993394121527672, + 0.015601771883666515, + 0.01891498453915119, + 0.01301708072423935, + 0.014902491122484207, + 0.006755859591066837, + 0.006931594107300043, + 0.011979375034570694, + 0.010419256053864956, + 0.014443143270909786, + 0.009749538265168667, + 0.008687978610396385, + 0.006996100302785635, + 0.021048076450824738, + 0.014944263733923435, + 0.012465455569326878, + 0.008500723168253899, + 0.011821992695331573, + 0.009142253547906876, + 0.01145583763718605, + 0.020894411951303482, + 0.005317071918398142, + 0.004942604806274176, + 0.005842679645866156, + 0.005236878991127014, + 0.006913313176482916, + 0.003366433549672365, + 0.004423063714057207, + 0.0036620444152504206, + 0.006226334720849991, + 0.00548823457211256, + 0.007734952960163355, + 0.003537162207067013, + 0.007670464459806681, + 0.008558422327041626, + 0.005030139349400997, + 0.008032835088670254, + 0.005986114498227835, + 0.005386050324887037, + 0.004873698111623526, + 0.00339861330576241, + 0.005181347951292992, + 0.0041920505464077, + 0.005031917244195938, + 0.004192688502371311, + 0.014203415252268314, + 0.014926684089004993, + 0.016306262463331223, + 0.0067700110375881195, + 0.013464358635246754, + 0.01632031984627247, + 0.0092927860096097, + 0.013405319303274155, + 0.0047726561315357685, + 0.005608144216239452, + 0.005721289664506912, + 0.0039784773252904415, + 0.007012438960373402, + 0.008679578080773354, + 0.0065433369018137455, + 0.008398544043302536, + 0.020413639023900032, + 0.023100925609469414, + 0.022479627281427383, + 0.01444091647863388, + 0.03032601997256279, + 0.025314662605524063, + 0.02245406061410904, + 0.02233153022825718, + 0.012801090255379677, + 0.0070699965581297874, + 0.011258021928369999, + 0.004491396714001894, + 0.013620101846754551, + 0.011645277962088585, + 0.07766032218933105, + 0.07943674921989441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.002820206806063652 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium close-up from the passenger seat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark charcoal wool suit with a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the figure's eyes and upper face obscured by deep shadows, revealing only the jawline and lips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the city lights outside the windows appear as blurred streaks of neon red, electric blue, and warm yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene dominated by a cool, dark palette with sharp highlights from the dashboard and city lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17091408371925354, + 0.08482929319143295, + 0.03650322183966637, + 0.0860428512096405, + 0.04658699408173561, + 0.052215877920389175, + 0.019963108003139496, + 0.03174981474876404, + 0.02395681105554104, + 0.4472379684448242 + ], + "credit_map": [ + 0.01747545599937439, + 0.033879172056913376, + 0.016594145447015762, + 0.020234962925314903, + 0.023956909775733948, + 0.021433677524328232, + 0.017269503325223923, + 0.020070265978574753, + 0.010639458894729614, + 0.00958744715899229, + 0.02115682326257229, + 0.00678854389116168, + 0.008338903076946735, + 0.011142324656248093, + 0.008081424050033092, + 0.009094377048313618, + 0.008190383203327656, + 0.004664357751607895, + 0.0036194452550262213, + 0.0029361883644014597, + 0.004300049971789122, + 0.004043407738208771, + 0.0035369854886084795, + 0.005212408024817705, + 0.013061869889497757, + 0.009232518263161182, + 0.00831008143723011, + 0.009499751962721348, + 0.02273043431341648, + 0.005465720314532518, + 0.011882394552230835, + 0.005860085133463144, + 0.005794146563857794, + 0.00646328367292881, + 0.005326005630195141, + 0.003108937991783023, + 0.008487687446177006, + 0.006812460720539093, + 0.0051420629024505615, + 0.005452411714941263, + 0.007391584105789661, + 0.005061313975602388, + 0.010427857749164104, + 0.0041155884973704815, + 0.006083397194743156, + 0.0051600150763988495, + 0.006322833709418774, + 0.007653290871530771, + 0.0015651904977858067, + 0.0021828918252140284, + 0.0025763167068362236, + 0.0022128806449472904, + 0.0029002749361097813, + 0.002908080816268921, + 0.003065128345042467, + 0.0025523442309349775, + 0.0026020179502665997, + 0.004056065808981657, + 0.005026550497859716, + 0.0021057063713669777, + 0.0038757408037781715, + 0.004987721797078848, + 0.0042636580765247345, + 0.004832357633858919, + 0.001601200900040567, + 0.0023051369935274124, + 0.00249062804505229, + 0.0015706726117059588, + 0.0055732750333845615, + 0.0032889533322304487, + 0.0035219464916735888, + 0.0036049988120794296, + 0.0218222476541996, + 0.011985824443399906, + 0.016340425238013268, + 0.009463758207857609, + 0.029939355328679085, + 0.022949092090129852, + 0.17378395795822144, + 0.1609533131122589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.003193756565451622 + } + ], + "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/0000058/group_9/index_tensor([1913, 1913])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_9/index_tensor([1913, 1913])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2b70c692c1689832fb65cf93072af3db986787d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_9/index_tensor([1913, 1913])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, slightly low-angle wide shot statically depicts a police officer in his forties standing at the center of a bustling city intersection, a pillar of authority. The low, golden sunlight of late afternoon casts long shadows across the asphalt, while the surrounding fiberglass skyscrapers gleam. The officer, with short, dark hair and a focused, weathered expression, wears a crisp, dark blue police uniform partially covered by a high-visibility yellow traffic vest, making him a beacon amidst the chaos. His hands are covered with bare white gloves. He performs his duties with sharp, controlled precision; in polished black boots, he decisively raises his gloved hand to stop a line of yellow taxis and sedans, then with a fluid, sweeping motion of his other arm, he beckons a line of waiting cars to proceed. The city hums around him—the blurred movement of traffic, the changing colors of traffic lights, and the throngs of pedestrians on the sidewalk—creating an atmosphere of focused concentration, his controlled gestures imposing a rhythmic order on the city's pulse.", + "seed": 2699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_9/index_1913_seed_2699_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": -0.1002934331074357, + "vlm_logprob": -0.1002934331074357, + "vlm_yes_no_margin": 3.449999988079071, + "vlm_token_credit": -0.1002934331074357, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06662360578775406 + ], + "opposite_logprob_tokens": [ + -5.441623687744141 + ], + "logprob_reward": -0.06662360578775406, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.06662360578775406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11647792905569077, + 0.09173751622438431, + 0.2867968678474426, + 0.07355049252510071, + 0.17652621865272522, + 0.029558567330241203, + 0.07119744271039963, + 0.038842689245939255, + 0.057681258767843246, + 0.05763104930520058 + ], + "credit_map": [ + 0.013084919191896915, + 0.008845830336213112, + 0.011806050315499306, + 0.007861350663006306, + 0.011559685692191124, + 0.027284862473607063, + 0.013493712991476059, + 0.02254151739180088, + 0.006664632819592953, + 0.006011550780385733, + 0.011430993676185608, + 0.011634648777544498, + 0.007021330762654543, + 0.016741493716835976, + 0.009766282513737679, + 0.022466585040092468, + 0.018568972125649452, + 0.013171711005270481, + 0.014924057759344578, + 0.0849258154630661, + 0.02044958807528019, + 0.006362213287502527, + 0.10311079770326614, + 0.0252836924046278, + 0.00606938824057579, + 0.00582917733117938, + 0.009718415327370167, + 0.01424120832234621, + 0.010795513167977333, + 0.00897180661559105, + 0.007323557045310736, + 0.010601429268717766, + 0.012244166806340218, + 0.012679838575422764, + 0.01220470480620861, + 0.020951686426997185, + 0.032884642481803894, + 0.02867172844707966, + 0.038554780185222626, + 0.018334664404392242, + 0.0037794404197484255, + 0.003417662810534239, + 0.003055550856515765, + 0.005090503487735987, + 0.0043237479403615, + 0.0029429818969219923, + 0.0029660318978130817, + 0.003982646390795708, + 0.005197522230446339, + 0.005008508916944265, + 0.022820621728897095, + 0.013904210180044174, + 0.006189077626913786, + 0.005899510346353054, + 0.005557624157518148, + 0.006620361935347319, + 0.005383627489209175, + 0.004087038803845644, + 0.005919359624385834, + 0.005070464685559273, + 0.00627810787409544, + 0.003068909514695406, + 0.0052745528519153595, + 0.0037606256082654, + 0.006107754539698362, + 0.006085305009037256, + 0.01533924788236618, + 0.006844379473477602, + 0.00440233526751399, + 0.0037377350963652134, + 0.010155276395380497, + 0.005009226966649294, + 0.0054941726848483086, + 0.0042968327179551125, + 0.00865078903734684, + 0.015605571679770947, + 0.005471074488013983, + 0.003538970137014985, + 0.005820298567414284, + 0.008753339760005474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.06662360578775406 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869065806269646 + ], + "opposite_logprob_tokens": [ + -3.0486905574798584 + ], + "logprob_reward": -0.04869065806269646, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869065806269646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1487971842288971, + 0.04907771199941635, + 0.3646051585674286, + 0.11611056327819824, + 0.08511839807033539, + 0.03080231137573719, + 0.0386371836066246, + 0.05788196623325348, + 0.0780300498008728, + 0.030939452350139618 + ], + "credit_map": [ + 0.014661709778010845, + 0.009553307667374611, + 0.011334779672324657, + 0.01008398924022913, + 0.011226139962673187, + 0.0457121916115284, + 0.013711487874388695, + 0.032513577491045, + 0.0045159426517784595, + 0.004272273741662502, + 0.004090710077434778, + 0.009166049771010876, + 0.008354301564395428, + 0.0037538507021963596, + 0.005912618711590767, + 0.009011967107653618, + 0.023376459255814552, + 0.017941974103450775, + 0.0181390680372715, + 0.11485759913921356, + 0.0275439340621233, + 0.007560247089713812, + 0.12563523650169373, + 0.029550645500421524, + 0.009656666778028011, + 0.010474530048668385, + 0.009205532260239124, + 0.013033262453973293, + 0.01875176653265953, + 0.033679377287626266, + 0.008816518820822239, + 0.012492911890149117, + 0.006431835703551769, + 0.007157862652093172, + 0.006420934572815895, + 0.010802306234836578, + 0.015232034958899021, + 0.01627844013273716, + 0.012590578757226467, + 0.010204407386481762, + 0.003398426342755556, + 0.0036019778344780207, + 0.0038052089512348175, + 0.006907038390636444, + 0.003465804038569331, + 0.0030896260868757963, + 0.003317320952191949, + 0.00321690971031785, + 0.003079506801441312, + 0.0037947020027786493, + 0.0052367630414664745, + 0.008516141213476658, + 0.007522569969296455, + 0.002408604370430112, + 0.005188488867133856, + 0.0028904080390930176, + 0.005033599212765694, + 0.004755055997520685, + 0.00778349582105875, + 0.005758405663073063, + 0.013375373557209969, + 0.004052543081343174, + 0.01153682079166174, + 0.005586670711636543, + 0.005445961840450764, + 0.003307112492620945, + 0.011473900638520718, + 0.0053266119211912155, + 0.005100894719362259, + 0.0033465945161879063, + 0.02527988888323307, + 0.018749086186289787, + 0.0035127555020153522, + 0.003226432716473937, + 0.004196472000330687, + 0.00534750334918499, + 0.004080452024936676, + 0.002808128483593464, + 0.003242524340748787, + 0.004525184165686369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.04869065806269646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486693501472473 + ], + "opposite_logprob_tokens": [ + -1.2236692905426025 + ], + "logprob_reward": -0.3486693501472473, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486693501472473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09911539405584335, + 0.028114955872297287, + 0.3508201837539673, + 0.04241356626152992, + 0.04655900225043297, + 0.04134048894047737, + 0.25473061203956604, + 0.03532901778817177, + 0.059009913355112076, + 0.04256678372621536 + ], + "credit_map": [ + 0.010604956187307835, + 0.007105272728949785, + 0.006388247013092041, + 0.005590345244854689, + 0.01003540214151144, + 0.029333705082535744, + 0.007199001498520374, + 0.022858478128910065, + 0.0028463599737733603, + 0.004064679145812988, + 0.003515040734782815, + 0.004645497538149357, + 0.003346715122461319, + 0.0032253386452794075, + 0.00321879587136209, + 0.003252530237659812, + 0.02213536575436592, + 0.01748708076775074, + 0.01692056469619274, + 0.10746106505393982, + 0.02721821516752243, + 0.010069789364933968, + 0.12085065245628357, + 0.028677497059106827, + 0.0038003860972821712, + 0.003684332361444831, + 0.0044477274641394615, + 0.004970896057784557, + 0.007677838671952486, + 0.006458509247750044, + 0.004986229818314314, + 0.006387649103999138, + 0.00365589652210474, + 0.0036006812006235123, + 0.007197508588433266, + 0.005327103193849325, + 0.006907074712216854, + 0.005715705920010805, + 0.0073607852682471275, + 0.0067942519672214985, + 0.003923320211470127, + 0.004014764446765184, + 0.003339602379128337, + 0.006468287203460932, + 0.009486761875450611, + 0.004812433384358883, + 0.005292490590363741, + 0.004002831876277924, + 0.01408385206013918, + 0.008826270699501038, + 0.09973137080669403, + 0.09178560227155685, + 0.010927300900220871, + 0.007614574860781431, + 0.010946493595838547, + 0.010815168730914593, + 0.006167992949485779, + 0.0033504811581224203, + 0.005013140384107828, + 0.0031984534580260515, + 0.006069574039429426, + 0.0027033507358282804, + 0.005271547473967075, + 0.0035544789861887693, + 0.0033189391251653433, + 0.003160444786772132, + 0.007966461591422558, + 0.00396484462544322, + 0.005022217053920031, + 0.0029804836958646774, + 0.013719243928790092, + 0.018877284601330757, + 0.004847661592066288, + 0.00361573020927608, + 0.005586051382124424, + 0.004551547113806009, + 0.0067827822640538216, + 0.003668729215860367, + 0.005665753968060017, + 0.007848530076444149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.3486693501472473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294245824217796 + ], + "opposite_logprob_tokens": [ + -3.77329421043396 + ], + "logprob_reward": -0.023294245824217796, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294245824217796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.126678928732872, + 0.15644757449626923, + 0.3153858482837677, + 0.06856770813465118, + 0.06147194281220436, + 0.03210299834609032, + 0.07997528463602066, + 0.032595135271549225, + 0.10496865212917328, + 0.02180592156946659 + ], + "credit_map": [ + 0.009711853228509426, + 0.004453971516340971, + 0.009263250976800919, + 0.008917376399040222, + 0.008650572039186954, + 0.03759956359863281, + 0.008359388448297977, + 0.039722952991724014, + 0.013686712831258774, + 0.011246592737734318, + 0.019725758582353592, + 0.013753503561019897, + 0.02555846981704235, + 0.024882730096578598, + 0.01878756657242775, + 0.028806239366531372, + 0.02051103301346302, + 0.016006702557206154, + 0.015136286616325378, + 0.10060346126556396, + 0.02585218846797943, + 0.007169385906308889, + 0.10545284301042557, + 0.024653950706124306, + 0.005669298116117716, + 0.004605837631970644, + 0.006681830156594515, + 0.009496987797319889, + 0.009811708703637123, + 0.02198694832623005, + 0.005080503411591053, + 0.005234592128545046, + 0.0053512235172092915, + 0.00473592896014452, + 0.006398227997124195, + 0.004933618940412998, + 0.00814109481871128, + 0.01253106351941824, + 0.006734510883688927, + 0.012646276503801346, + 0.0035464991815388203, + 0.0036938744597136974, + 0.0028993990272283554, + 0.0045791552402079105, + 0.007930977270007133, + 0.0033442811109125614, + 0.0035439403727650642, + 0.002564873779192567, + 0.005382935982197523, + 0.0030263580847531557, + 0.03180162236094475, + 0.02573869563639164, + 0.0036852529738098383, + 0.003132461803033948, + 0.0032119567040354013, + 0.003996002487838268, + 0.0050678551197052, + 0.0028887928929179907, + 0.0037455882411450148, + 0.0033683781512081623, + 0.007128114812076092, + 0.00266597350127995, + 0.0043003857135772705, + 0.0034300456754863262, + 0.005580172874033451, + 0.0035014234017580748, + 0.03661748021841049, + 0.007332285866141319, + 0.007575757335871458, + 0.0047859619371593, + 0.018072720617055893, + 0.02150285430252552, + 0.001989595824852586, + 0.001847108593210578, + 0.002777428599074483, + 0.0036266993265599012, + 0.003290745196864009, + 0.002131636720150709, + 0.0025677212979644537, + 0.003574986010789871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.023294245824217796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418930571526289 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.01418930571526289, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418930571526289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13538016378879547, + 0.07384806126356125, + 0.5185494422912598, + 0.03330210968852043, + 0.05375368148088455, + 0.04725933074951172, + 0.05940176174044609, + 0.022713812068104744, + 0.03793497011065483, + 0.017856642603874207 + ], + "credit_map": [ + 0.017293373122811317, + 0.013903436250984669, + 0.019468771293759346, + 0.01259867288172245, + 0.012410957366228104, + 0.0254887156188488, + 0.009994988329708576, + 0.024221258237957954, + 0.005752997938543558, + 0.007633122615516186, + 0.006502491421997547, + 0.013599056750535965, + 0.008107196539640427, + 0.0072619556449353695, + 0.009360617958009243, + 0.015630623325705528, + 0.03356030210852623, + 0.027182232588529587, + 0.026361649855971336, + 0.16507819294929504, + 0.043534014374017715, + 0.01389629952609539, + 0.16955998539924622, + 0.03937680274248123, + 0.0026661399751901627, + 0.0029889377765357494, + 0.005348562262952328, + 0.006246458739042282, + 0.004086371511220932, + 0.0029365590307861567, + 0.004132012836635113, + 0.00489706639200449, + 0.004872648511081934, + 0.005043391603976488, + 0.004326355177909136, + 0.006404187995940447, + 0.01036430336534977, + 0.008965890854597092, + 0.007929676212370396, + 0.005847226828336716, + 0.005567991640418768, + 0.006459635682404041, + 0.005628594197332859, + 0.009673812426626682, + 0.0047921594232320786, + 0.006265817675739527, + 0.004865421913564205, + 0.0040058959275484085, + 0.0031270859763026237, + 0.0029673923272639513, + 0.02038351446390152, + 0.020522357895970345, + 0.003628534497693181, + 0.0023340480402112007, + 0.004112693015486002, + 0.002326137851923704, + 0.003238888457417488, + 0.0023179801646620035, + 0.0027498325798660517, + 0.0025847414508461952, + 0.002742934972047806, + 0.0032341505866497755, + 0.002308236202225089, + 0.003537046257406473, + 0.002993241185322404, + 0.0023173217196017504, + 0.011444908566772938, + 0.002732747234404087, + 0.0021026800386607647, + 0.0016062696231529117, + 0.010853649117052555, + 0.0038841518107801676, + 0.001513602677732706, + 0.0013410357059910893, + 0.002483271062374115, + 0.003428594907745719, + 0.0028420044109225273, + 0.0012633930891752243, + 0.0017772584687918425, + 0.0032074821647256613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.01418930571526289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, slightly low-angle wide shot statically depicts a police officer in his forties standing at the center of a bustling city intersection, a pillar of authority. The low, golden sunlight of late afternoon casts long shadows across the asphalt, while the surrounding fiberglass skyscrapers gleam. The officer, with short, dark hair and a focused, weathered expression, wears a crisp, dark blue police uniform partially covered by a high-visibility yellow traffic vest, making him a beacon amidst the chaos. His hands are covered with bare white gloves. He performs his duties with sharp, controlled precision; in polished black boots, he decisively raises his gloved hand to stop a line of yellow taxis and sedans, then with a fluid, sweeping motion of his other arm, he beckons a line of waiting cars to proceed. The city hums around him—the blurred movement of traffic, the changing colors of traffic lights, and the throngs of pedestrians on the sidewalk—creating an atmosphere of focused concentration, his controlled gestures imposing a rhythmic order on the city's pulse.", + "seed": 2700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_9/index_1913_seed_2700_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": -0.07638662941753864, + "vlm_logprob": -0.07638662941753864, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07638662941753864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07847808301448822 + ], + "opposite_logprob_tokens": [ + -4.82847785949707 + ], + "logprob_reward": -0.07847808301448822, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.07847808301448822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11648528277873993, + 0.07934597134590149, + 0.05732985958456993, + 0.1088317334651947, + 0.16018566489219666, + 0.10545564442873001, + 0.04654945433139801, + 0.08479104936122894, + 0.10746746510267258, + 0.13355791568756104 + ], + "credit_map": [ + 0.018799496814608574, + 0.011671729385852814, + 0.012672819197177887, + 0.013900212943553925, + 0.018133262172341347, + 0.01024909969419241, + 0.013731163926422596, + 0.01732749491930008, + 0.011272653006017208, + 0.006338045001029968, + 0.008650977164506912, + 0.008302909322082996, + 0.012657074257731438, + 0.007610812317579985, + 0.006858984008431435, + 0.017654508352279663, + 0.006338866893202066, + 0.004857719410210848, + 0.00808686576783657, + 0.00761920353397727, + 0.0077722640708088875, + 0.006894305814057589, + 0.009219036437571049, + 0.006541593000292778, + 0.01610223762691021, + 0.008714397437870502, + 0.017433827742934227, + 0.017776919528841972, + 0.014532684348523617, + 0.010043357498943806, + 0.015291758812963963, + 0.008936546742916107, + 0.00946319755166769, + 0.00993058830499649, + 0.04875550791621208, + 0.009216266684234142, + 0.01627030409872532, + 0.008289631456136703, + 0.04810450226068497, + 0.010155667550861835, + 0.01666819490492344, + 0.011751262471079826, + 0.01223570667207241, + 0.02075444720685482, + 0.009581614285707474, + 0.007774490397423506, + 0.01483522541821003, + 0.011854688636958599, + 0.0073549021035432816, + 0.004323246423155069, + 0.004779864102602005, + 0.007742541376501322, + 0.007324581034481525, + 0.004745478741824627, + 0.005735335871577263, + 0.004543500952422619, + 0.009711585007607937, + 0.006972223985940218, + 0.006055696867406368, + 0.012725835666060448, + 0.01279357634484768, + 0.008689370937645435, + 0.012024379335343838, + 0.01581837609410286, + 0.006412858609110117, + 0.005652170162647963, + 0.016024824231863022, + 0.03082563355565071, + 0.007564889267086983, + 0.00565768638625741, + 0.02363491617143154, + 0.011694478802382946, + 0.0120763648301363, + 0.012296517379581928, + 0.017149362713098526, + 0.01508315745741129, + 0.02454415336251259, + 0.015368770807981491, + 0.01803412288427353, + 0.019005458801984787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.07847808301448822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309457167983055 + ], + "opposite_logprob_tokens": [ + -3.1680946350097656 + ], + "logprob_reward": -0.04309457167983055, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309457167983055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05881728604435921, + 0.044629499316215515, + 0.09268206357955933, + 0.051154427230358124, + 0.040610697120428085, + 0.24564091861248016, + 0.03354426845908165, + 0.28084123134613037, + 0.0783100575208664, + 0.07376950979232788 + ], + "credit_map": [ + 0.007605627179145813, + 0.005481159780174494, + 0.008356014266610146, + 0.007456890773028135, + 0.008765286765992641, + 0.0068972292356193066, + 0.008129258640110493, + 0.006125814747065306, + 0.0045021953992545605, + 0.004460932221263647, + 0.0030912798829376698, + 0.0051318584010005, + 0.009010899811983109, + 0.0064231231808662415, + 0.0061265891417860985, + 0.005882617086172104, + 0.007550508249551058, + 0.005762844812124968, + 0.009231855161488056, + 0.01563139632344246, + 0.014196595177054405, + 0.011284120380878448, + 0.016808055341243744, + 0.012216686271131039, + 0.005578993819653988, + 0.005261932034045458, + 0.006706117186695337, + 0.008233548142015934, + 0.006093487609177828, + 0.007597779389470816, + 0.006838337052613497, + 0.004844233393669128, + 0.0038783864583820105, + 0.0024933358654379845, + 0.006483234465122223, + 0.004311695694923401, + 0.004749768879264593, + 0.0049921851605176926, + 0.006636823993176222, + 0.007065268699079752, + 0.02654082700610161, + 0.0185388270765543, + 0.04669434204697609, + 0.0491015687584877, + 0.01642330177128315, + 0.017591889947652817, + 0.03369074687361717, + 0.03705942630767822, + 0.0038187806494534016, + 0.0030809100717306137, + 0.003766026347875595, + 0.005114665720611811, + 0.00490228459239006, + 0.004790532868355513, + 0.004267602693289518, + 0.003803468309342861, + 0.042348701506853104, + 0.03667592257261276, + 0.01523407083004713, + 0.04153214022517204, + 0.05653342232108116, + 0.015969429165124893, + 0.035489704459905624, + 0.03705785423517227, + 0.0054448372684419155, + 0.0038070103619247675, + 0.020007062703371048, + 0.020951349288225174, + 0.00767676392570138, + 0.005222491454333067, + 0.006913373712450266, + 0.008287167176604271, + 0.008411726914346218, + 0.0066519021056592464, + 0.009988641366362572, + 0.008659868501126766, + 0.012584890238940716, + 0.0073722973465919495, + 0.009065644815564156, + 0.011034540832042694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.04309457167983055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255842685699463 + ], + "opposite_logprob_tokens": [ + -1.6005842685699463 + ], + "logprob_reward": -0.2255842685699463, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255842685699463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448034793138504, + 0.08261501044034958, + 0.1555737555027008, + 0.09129375219345093, + 0.047337163239717484, + 0.06566689908504486, + 0.04545118659734726, + 0.2023545503616333, + 0.07240772992372513, + 0.09249651432037354 + ], + "credit_map": [ + 0.017676036804914474, + 0.013771397061645985, + 0.013896024785935879, + 0.013537620194256306, + 0.022507918998599052, + 0.015585721470415592, + 0.029479507356882095, + 0.01834925077855587, + 0.01194676198065281, + 0.006523296236991882, + 0.007396886125206947, + 0.0077882008627057076, + 0.01325654610991478, + 0.008060489781200886, + 0.008096765726804733, + 0.019546058028936386, + 0.01396852731704712, + 0.013615126721560955, + 0.014376908540725708, + 0.023604214191436768, + 0.022389881312847137, + 0.01877043955028057, + 0.02321479097008705, + 0.02563387341797352, + 0.011423850432038307, + 0.008243927732110023, + 0.015621593222022057, + 0.01501310896128416, + 0.009532684460282326, + 0.009999802336096764, + 0.012184002436697483, + 0.009274786338210106, + 0.0034895495045930147, + 0.002722528064623475, + 0.01249642949551344, + 0.004873173777014017, + 0.003950088284909725, + 0.004593297373503447, + 0.0072406805120408535, + 0.007971419021487236, + 0.00897032581269741, + 0.005405274219810963, + 0.015531077980995178, + 0.004654584918171167, + 0.00425834720954299, + 0.003556920448318124, + 0.0063710398972034454, + 0.016919324174523354, + 0.006701831705868244, + 0.004335798788815737, + 0.006729891989380121, + 0.0066904546692967415, + 0.00437462143599987, + 0.004111019428819418, + 0.006436857394874096, + 0.0060707093216478825, + 0.021736567839980125, + 0.025745971128344536, + 0.010038151405751705, + 0.032216258347034454, + 0.049940094351768494, + 0.01247768010944128, + 0.025737373158335686, + 0.02446245402097702, + 0.006393512710928917, + 0.003806819673627615, + 0.017177492380142212, + 0.019292335957288742, + 0.006448866333812475, + 0.005656152963638306, + 0.0064317770302295685, + 0.007200775668025017, + 0.008220093324780464, + 0.00781333725899458, + 0.010693512856960297, + 0.005750036798417568, + 0.022717446088790894, + 0.011069291271269321, + 0.013514408841729164, + 0.012718386016786098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.2255842685699463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198120683431625 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198120683431625, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198120683431625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08953523635864258, + 0.10312706977128983, + 0.09518641233444214, + 0.16548913717269897, + 0.10721830278635025, + 0.06613253802061081, + 0.05583663657307625, + 0.06899119168519974, + 0.14437799155712128, + 0.10410554707050323 + ], + "credit_map": [ + 0.010252592153847218, + 0.006711641326546669, + 0.011013000272214413, + 0.009884063154459, + 0.014578812755644321, + 0.011119427159428596, + 0.015466100536286831, + 0.010509590618312359, + 0.01567801646888256, + 0.00848495215177536, + 0.01236607413738966, + 0.015256834216415882, + 0.012271063402295113, + 0.010706867091357708, + 0.009392162784934044, + 0.018971091136336327, + 0.008569132536649704, + 0.006542467977851629, + 0.009953154250979424, + 0.013399357907474041, + 0.010637274011969566, + 0.016295157372951508, + 0.013833630830049515, + 0.015956221148371696, + 0.020621325820684433, + 0.01845186948776245, + 0.025924092158675194, + 0.029361851513385773, + 0.017719652503728867, + 0.015835538506507874, + 0.021279558539390564, + 0.01629522256553173, + 0.0072604455053806305, + 0.006503745913505554, + 0.03481130301952362, + 0.007251870818436146, + 0.009223146364092827, + 0.005696855019778013, + 0.030486293137073517, + 0.0059846313670277596, + 0.00915489811450243, + 0.005606796592473984, + 0.007357665337622166, + 0.00935919675976038, + 0.014248448424041271, + 0.006053061690181494, + 0.007522065192461014, + 0.006830405443906784, + 0.008460760116577148, + 0.005383058451116085, + 0.006504329852759838, + 0.011538301594555378, + 0.004901804495602846, + 0.005290498957037926, + 0.007758375257253647, + 0.005999501328915358, + 0.0070894332602620125, + 0.005723990499973297, + 0.004412814509123564, + 0.016619767993688583, + 0.012120205909013748, + 0.0047451225109398365, + 0.008760517463088036, + 0.009519333951175213, + 0.009250747039914131, + 0.006884233560413122, + 0.026262007653713226, + 0.03350116312503815, + 0.016198351979255676, + 0.009685921482741833, + 0.024461869150400162, + 0.01813369430601597, + 0.008413489907979965, + 0.009030388668179512, + 0.010101941414177418, + 0.007048775441944599, + 0.026952775195240974, + 0.013215398415923119, + 0.01681056246161461, + 0.012532200664281845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.014198120683431625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057810313999653 + ], + "opposite_logprob_tokens": [ + -3.895578145980835 + ], + "logprob_reward": -0.02057810313999653, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057810313999653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1482176035642624, + 0.12357667088508606, + 0.17246367037296295, + 0.07906826585531235, + 0.13466180860996246, + 0.10467034578323364, + 0.0712466835975647, + 0.07601581513881683, + 0.029082486405968666, + 0.060996636748313904 + ], + "credit_map": [ + 0.020310204476118088, + 0.01507203746587038, + 0.01615198329091072, + 0.015316390432417393, + 0.02090262994170189, + 0.022959524765610695, + 0.019266778603196144, + 0.01823805645108223, + 0.01635882817208767, + 0.01065846998244524, + 0.013526550494134426, + 0.014020471833646297, + 0.02511209063231945, + 0.01673804223537445, + 0.014182024635374546, + 0.012980195693671703, + 0.017662500962615013, + 0.015517478808760643, + 0.019042963162064552, + 0.019939279183745384, + 0.033713024109601974, + 0.019387779757380486, + 0.027157390490174294, + 0.020043255761265755, + 0.007415901869535446, + 0.008234508335590363, + 0.012878783978521824, + 0.013485308736562729, + 0.011745509691536427, + 0.0071998159401118755, + 0.010139125399291515, + 0.007969310507178307, + 0.016311367973685265, + 0.0075288075022399426, + 0.014562230557203293, + 0.009531369432806969, + 0.017160335555672646, + 0.021269433200359344, + 0.024641720578074455, + 0.023656535893678665, + 0.014727242290973663, + 0.008253092877566814, + 0.01491722371429205, + 0.01454444881528616, + 0.011214973405003548, + 0.011989138089120388, + 0.014891505241394043, + 0.01413272600620985, + 0.006486608646810055, + 0.004953674040734768, + 0.005673751700669527, + 0.013659320771694183, + 0.01261581014841795, + 0.00854487158358097, + 0.010586146265268326, + 0.00872649997472763, + 0.009177332744002342, + 0.005533893592655659, + 0.0058244382962584496, + 0.017889130860567093, + 0.009672466665506363, + 0.006803356576710939, + 0.009900175966322422, + 0.011215025559067726, + 0.003888621460646391, + 0.0021137872245162725, + 0.003365206765010953, + 0.005001033656299114, + 0.0033074121456593275, + 0.0025025608483701944, + 0.004726655315607786, + 0.004177211318165064, + 0.00508699007332325, + 0.006103872787207365, + 0.007600829936563969, + 0.006990835536271334, + 0.009805768728256226, + 0.006452372297644615, + 0.009106415323913097, + 0.009849549271166325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02057810313999653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, slightly low-angle wide shot statically depicts a police officer in his forties standing at the center of a bustling city intersection, a pillar of authority. The low, golden sunlight of late afternoon casts long shadows across the asphalt, while the surrounding fiberglass skyscrapers gleam. The officer, with short, dark hair and a focused, weathered expression, wears a crisp, dark blue police uniform partially covered by a high-visibility yellow traffic vest, making him a beacon amidst the chaos. His hands are covered with bare white gloves. He performs his duties with sharp, controlled precision; in polished black boots, he decisively raises his gloved hand to stop a line of yellow taxis and sedans, then with a fluid, sweeping motion of his other arm, he beckons a line of waiting cars to proceed. The city hums around him—the blurred movement of traffic, the changing colors of traffic lights, and the throngs of pedestrians on the sidewalk—creating an atmosphere of focused concentration, his controlled gestures imposing a rhythmic order on the city's pulse.", + "seed": 2701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_9/index_1913_seed_2701_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": -0.09005031902343034, + "vlm_logprob": -0.09005031902343034, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.09005031902343034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0661081075668335 + ], + "opposite_logprob_tokens": [ + -5.566108226776123 + ], + "logprob_reward": -0.0661081075668335, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0661081075668335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23232068121433258, + 0.1447962373495102, + 0.06548602879047394, + 0.05614018440246582, + 0.06338855624198914, + 0.07980518043041229, + 0.103172168135643, + 0.08149383962154388, + 0.11128264665603638, + 0.06211439147591591 + ], + "credit_map": [ + 0.03471195697784424, + 0.026637893170118332, + 0.0209702979773283, + 0.02670573629438877, + 0.036860037595033646, + 0.03926556184887886, + 0.01619034819304943, + 0.030978865921497345, + 0.014167251996695995, + 0.014765923842787743, + 0.015397269278764725, + 0.02088910900056362, + 0.02910725772380829, + 0.016710322350263596, + 0.011281155981123447, + 0.022477956488728523, + 0.008084762841463089, + 0.00807309988886118, + 0.005282531026750803, + 0.010357340797781944, + 0.013964983634650707, + 0.006333256606012583, + 0.0071358694694936275, + 0.0062541957013309, + 0.0053970664739608765, + 0.0062232790514826775, + 0.00635064160451293, + 0.011712546460330486, + 0.008601154200732708, + 0.006310695316642523, + 0.0057808100245893, + 0.00576399639248848, + 0.007564019877463579, + 0.00804112944751978, + 0.006023467052727938, + 0.009031577967107296, + 0.009336419403553009, + 0.009813682176172733, + 0.006009745877236128, + 0.007568522356450558, + 0.013064605183899403, + 0.009509852156043053, + 0.007003869395703077, + 0.008582049049437046, + 0.013657128438353539, + 0.010691934265196323, + 0.008840668015182018, + 0.008455086499452591, + 0.013798261992633343, + 0.010678667575120926, + 0.012245645746588707, + 0.0110167870298028, + 0.017230428755283356, + 0.018972335383296013, + 0.00963009987026453, + 0.0095999576151371, + 0.012312893755733967, + 0.011327803134918213, + 0.0058424100279808044, + 0.007356638554483652, + 0.01111641712486744, + 0.016572989523410797, + 0.007631381042301655, + 0.009333315305411816, + 0.01824885420501232, + 0.012319468893110752, + 0.01307378988713026, + 0.014058848842978477, + 0.016748858615756035, + 0.016287444159388542, + 0.009991468861699104, + 0.010553929023444653, + 0.012742427177727222, + 0.009205201640725136, + 0.006138920783996582, + 0.005844898987561464, + 0.007263850886374712, + 0.009175463579595089, + 0.004974949173629284, + 0.0067686839029192924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0661081075668335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308018460869789 + ], + "opposite_logprob_tokens": [ + -3.1680800914764404 + ], + "logprob_reward": -0.04308018460869789, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308018460869789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38069286942481995, + 0.05721702426671982, + 0.096365787088871, + 0.03529868274927139, + 0.09951911866664886, + 0.03737315535545349, + 0.1301422268152237, + 0.07354946434497833, + 0.05468351021409035, + 0.0351581797003746 + ], + "credit_map": [ + 0.033180613070726395, + 0.04487760737538338, + 0.033443059772253036, + 0.04391853138804436, + 0.10958928614854813, + 0.04319904372096062, + 0.0309124905616045, + 0.041572216898202896, + 0.005395230837166309, + 0.004728477913886309, + 0.0065399035811424255, + 0.008897842839360237, + 0.012588730081915855, + 0.006766669452190399, + 0.005779011640697718, + 0.0065211583860218525, + 0.007500854786485434, + 0.010301562957465649, + 0.006731079891324043, + 0.010190866887569427, + 0.02888338267803192, + 0.01010138913989067, + 0.011027256958186626, + 0.011629403568804264, + 0.002698746742680669, + 0.0039435182698071, + 0.002899833954870701, + 0.004104856867343187, + 0.009269928559660912, + 0.00422787107527256, + 0.0053784907795488834, + 0.002775434637442231, + 0.012226431630551815, + 0.009283656254410744, + 0.009699085727334023, + 0.009141091257333755, + 0.027975574135780334, + 0.011190191842615604, + 0.010636748746037483, + 0.00936634186655283, + 0.004074987955391407, + 0.003806393127888441, + 0.0024023172445595264, + 0.0036021899431943893, + 0.00904945470392704, + 0.006910072639584541, + 0.0030567138455808163, + 0.004471024498343468, + 0.013729523867368698, + 0.013591757044196129, + 0.010641593486070633, + 0.013132107444107533, + 0.036039143800735474, + 0.020770180970430374, + 0.012182808481156826, + 0.0100551201030612, + 0.011287747882306576, + 0.011275172233581543, + 0.007854134775698185, + 0.007724895607680082, + 0.011148246005177498, + 0.01131792739033699, + 0.005672995001077652, + 0.0072683487087488174, + 0.0054811048321425915, + 0.005278440658003092, + 0.0068217189982533455, + 0.005787982139736414, + 0.007762952707707882, + 0.010526548139750957, + 0.005694021936506033, + 0.007330740336328745, + 0.0032842475920915604, + 0.0036964607425034046, + 0.00304266600869596, + 0.003747964045032859, + 0.009957212954759598, + 0.004574979189783335, + 0.004519881214946508, + 0.0023347686510533094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.04308018460869789 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813361883163452 + ], + "opposite_logprob_tokens": [ + -1.4063361883163452 + ], + "logprob_reward": -0.2813361883163452, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813361883163452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22865960001945496, + 0.06126724183559418, + 0.08120708167552948, + 0.03614674508571625, + 0.13378813862800598, + 0.08825162798166275, + 0.17694884538650513, + 0.08052422851324081, + 0.0751492902636528, + 0.03805721551179886 + ], + "credit_map": [ + 0.023054324090480804, + 0.03399874269962311, + 0.022137751802802086, + 0.026727618649601936, + 0.05674821883440018, + 0.026680095121264458, + 0.01593060791492462, + 0.02338222600519657, + 0.006348024122416973, + 0.005166369490325451, + 0.00601595314219594, + 0.007800798397511244, + 0.012356286868453026, + 0.008273512125015259, + 0.007952973246574402, + 0.007353322580456734, + 0.006832674145698547, + 0.00844233762472868, + 0.0043465010821819305, + 0.007958448491990566, + 0.02337457612156868, + 0.009763223119080067, + 0.009474526159465313, + 0.011014790274202824, + 0.004240992944687605, + 0.005216095596551895, + 0.0028911991976201534, + 0.004767592530697584, + 0.007183993700891733, + 0.003848080988973379, + 0.004187549464404583, + 0.00381123973056674, + 0.012850280851125717, + 0.01294529065489769, + 0.012669842690229416, + 0.016443375498056412, + 0.03353907912969589, + 0.014674192294478416, + 0.012286278419196606, + 0.01837979629635811, + 0.017925089225172997, + 0.014078319072723389, + 0.009373815730214119, + 0.015256681479513645, + 0.006765430327504873, + 0.010203558020293713, + 0.005220317747443914, + 0.009428411722183228, + 0.01879960112273693, + 0.018310265615582466, + 0.014328064396977425, + 0.018167300149798393, + 0.052306342869997025, + 0.023593947291374207, + 0.01795346848666668, + 0.01348986104130745, + 0.011031190864741802, + 0.011208265088498592, + 0.009019359946250916, + 0.011156491935253143, + 0.009658078663051128, + 0.008601246401667595, + 0.0071249972097575665, + 0.012724597938358784, + 0.009597753174602985, + 0.008606499992311, + 0.004685978405177593, + 0.00785782840102911, + 0.01340959221124649, + 0.012608957476913929, + 0.009202002547681332, + 0.009180680848658085, + 0.0046678576618433, + 0.0035602382849901915, + 0.004950294271111488, + 0.005117191467434168, + 0.0051850685849785805, + 0.004978892393410206, + 0.004866865463554859, + 0.004730804357677698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.2813361883163452 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048656027764081955 + ], + "opposite_logprob_tokens": [ + -3.0486559867858887 + ], + "logprob_reward": -0.048656027764081955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048656027764081955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1442156285047531, + 0.07609757035970688, + 0.07707502692937851, + 0.05046366900205612, + 0.06385507434606552, + 0.06409276276826859, + 0.1458914875984192, + 0.1065916046500206, + 0.2040945291519165, + 0.06762261688709259 + ], + "credit_map": [ + 0.015362044796347618, + 0.01460978388786316, + 0.017825061455368996, + 0.0214681513607502, + 0.027078846469521523, + 0.015711503103375435, + 0.013498449698090553, + 0.018661806359887123, + 0.006951433140784502, + 0.00674731470644474, + 0.008098771795630455, + 0.007828614674508572, + 0.018544161692261696, + 0.011385571211576462, + 0.007735290564596653, + 0.008806413970887661, + 0.007804654072970152, + 0.0100519685074687, + 0.006381964311003685, + 0.0072400979697704315, + 0.01993502862751484, + 0.00827872846275568, + 0.008729328401386738, + 0.008653257042169571, + 0.004736457951366901, + 0.006190631538629532, + 0.004111027345061302, + 0.005570787005126476, + 0.012739075347781181, + 0.007628172170370817, + 0.006186538841575384, + 0.003300981828942895, + 0.0066874646581709385, + 0.007591909728944302, + 0.005356654990464449, + 0.006674503907561302, + 0.015988251194357872, + 0.009346932172775269, + 0.005267346277832985, + 0.006942017935216427, + 0.011447890661656857, + 0.007772183045744896, + 0.006294664461165667, + 0.010867212899029255, + 0.008848275057971478, + 0.006777426227927208, + 0.0052595678716897964, + 0.006825548131018877, + 0.016486987471580505, + 0.01811368763446808, + 0.011424137279391289, + 0.016394618898630142, + 0.04081590473651886, + 0.01612839289009571, + 0.01440794114023447, + 0.01211981475353241, + 0.013313698582351208, + 0.011468017473816872, + 0.012268949300050735, + 0.012199368327856064, + 0.021949490532279015, + 0.012655641883611679, + 0.011025835759937763, + 0.011710602790117264, + 0.019986242055892944, + 0.020434608682990074, + 0.017795216292142868, + 0.02203143574297428, + 0.03530279919505119, + 0.03802924603223801, + 0.023772096261382103, + 0.02674289420247078, + 0.006587678100913763, + 0.006442468613386154, + 0.007957625202834606, + 0.006867456715553999, + 0.014222460798919201, + 0.010688110254704952, + 0.007009781431406736, + 0.007847033441066742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.048656027764081955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107108686119318 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.01107108686119318, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107108686119318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41393744945526123, + 0.08321277052164078, + 0.034929655492305756, + 0.06312112510204315, + 0.10224317759275436, + 0.04079911485314369, + 0.08367744833230972, + 0.09181664884090424, + 0.05436272174119949, + 0.03189988061785698 + ], + "credit_map": [ + 0.06242671236395836, + 0.042180292308330536, + 0.06596002727746964, + 0.05553784593939781, + 0.06157222390174866, + 0.047773756086826324, + 0.025174710899591446, + 0.053311899304389954, + 0.009499297477304935, + 0.008220016025006771, + 0.00858389027416706, + 0.01580839790403843, + 0.01261160522699356, + 0.009020213037729263, + 0.005171490367501974, + 0.014297865331172943, + 0.004501840565353632, + 0.004098458681255579, + 0.003310099244117737, + 0.005296149756759405, + 0.005439537577331066, + 0.0038846340030431747, + 0.0034776267129927874, + 0.004921309649944305, + 0.006856810301542282, + 0.005909958388656378, + 0.0074453214183449745, + 0.009996633976697922, + 0.01538834162056446, + 0.006051766686141491, + 0.003919712733477354, + 0.007552581373602152, + 0.010754012502729893, + 0.009886224754154682, + 0.01059118565171957, + 0.018725503236055374, + 0.01792702078819275, + 0.015601112507283688, + 0.007638920098543167, + 0.011119202710688114, + 0.0049334317445755005, + 0.004064700100570917, + 0.004539142828434706, + 0.0040801106952130795, + 0.008754723705351353, + 0.0058372351340949535, + 0.002951883478090167, + 0.00563788553699851, + 0.008902108296751976, + 0.010518462397158146, + 0.011831731535494328, + 0.006768098101019859, + 0.015023215673863888, + 0.015530012547969818, + 0.006941370666027069, + 0.008162455633282661, + 0.012061206623911858, + 0.01809140481054783, + 0.008363192901015282, + 0.008421028964221478, + 0.012852507643401623, + 0.015466602519154549, + 0.006325758993625641, + 0.010234944522380829, + 0.006491092499345541, + 0.005879546049982309, + 0.004624550696462393, + 0.005748666822910309, + 0.01193793211132288, + 0.007514732889831066, + 0.006752333138138056, + 0.0054138703271746635, + 0.005839141551405191, + 0.0036846508737653494, + 0.004283489193767309, + 0.0033021862618625164, + 0.0029502706602215767, + 0.005073146894574165, + 0.0023494751658290625, + 0.004417520947754383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.01107108686119318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, slightly low-angle wide shot statically depicts a police officer in his forties standing at the center of a bustling city intersection, a pillar of authority. The low, golden sunlight of late afternoon casts long shadows across the asphalt, while the surrounding fiberglass skyscrapers gleam. The officer, with short, dark hair and a focused, weathered expression, wears a crisp, dark blue police uniform partially covered by a high-visibility yellow traffic vest, making him a beacon amidst the chaos. His hands are covered with bare white gloves. He performs his duties with sharp, controlled precision; in polished black boots, he decisively raises his gloved hand to stop a line of yellow taxis and sedans, then with a fluid, sweeping motion of his other arm, he beckons a line of waiting cars to proceed. The city hums around him—the blurred movement of traffic, the changing colors of traffic lights, and the throngs of pedestrians on the sidewalk—creating an atmosphere of focused concentration, his controlled gestures imposing a rhythmic order on the city's pulse.", + "seed": 2702, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000058/group_9/index_1913_seed_2702_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": -0.08281668666750193, + "vlm_logprob": -0.08281668666750193, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.08281668666750193, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05714934691786766 + ], + "opposite_logprob_tokens": [ + -6.182149410247803 + ], + "logprob_reward": -0.05714934691786766, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.05714934691786766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056512318551540375, + 0.028354177251458168, + 0.03515344113111496, + 0.029167862609028816, + 0.029224956408143044, + 0.06386975198984146, + 0.02818797342479229, + 0.5919080376625061, + 0.05185782536864281, + 0.0857636034488678 + ], + "credit_map": [ + 0.00653856573626399, + 0.004432962276041508, + 0.009460747241973877, + 0.008250693790614605, + 0.006768835708498955, + 0.007325015962123871, + 0.004534485284239054, + 0.009201016277074814, + 0.004237814340740442, + 0.002943113911896944, + 0.004133312962949276, + 0.0043213521130383015, + 0.0027263176161795855, + 0.003660643007606268, + 0.002689559478312731, + 0.0036420642863959074, + 0.005365545861423016, + 0.00494727399200201, + 0.004909539595246315, + 0.0034578382037580013, + 0.003656785935163498, + 0.004120380152016878, + 0.0028338381089270115, + 0.005862238816916943, + 0.004526826087385416, + 0.004343830980360508, + 0.0037556816823780537, + 0.0035469166468828917, + 0.0031394425313919783, + 0.0037741370033472776, + 0.002134843496605754, + 0.003946186043322086, + 0.0029510254971683025, + 0.0030723672825843096, + 0.005520795937627554, + 0.004927084315568209, + 0.003191559575498104, + 0.0028195944614708424, + 0.002972666872665286, + 0.0037698622327297926, + 0.010006153956055641, + 0.007752318866550922, + 0.012179589830338955, + 0.00994332879781723, + 0.006065242923796177, + 0.004692292772233486, + 0.006311766337603331, + 0.006919058505445719, + 0.002647594315931201, + 0.003057164838537574, + 0.003487701527774334, + 0.003606579964980483, + 0.0037290528416633606, + 0.0026099286042153835, + 0.0028162114322185516, + 0.006233741529285908, + 0.05105205997824669, + 0.022312568500638008, + 0.22086787223815918, + 0.04491114988923073, + 0.02940438687801361, + 0.0260150246322155, + 0.0169206615537405, + 0.18042431771755219, + 0.007488724309951067, + 0.006808000151067972, + 0.01185031607747078, + 0.0056184385903179646, + 0.004576544743031263, + 0.005768067669123411, + 0.004614060278981924, + 0.005133672617375851, + 0.011060060933232307, + 0.0068690162152051926, + 0.02073477767407894, + 0.012499759905040264, + 0.003947230521589518, + 0.008643968030810356, + 0.005134674720466137, + 0.01687411405146122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.05714934691786766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811749815940857 + ], + "opposite_logprob_tokens": [ + -3.2881174087524414 + ], + "logprob_reward": -0.03811749815940857, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811749815940857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07524686306715012, + 0.0259962510317564, + 0.012324461713433266, + 0.021319057792425156, + 0.020792512223124504, + 0.011987295001745224, + 0.015516093000769615, + 0.7268686294555664, + 0.03178178146481514, + 0.058166999369859695 + ], + "credit_map": [ + 0.007990979589521885, + 0.00764872832223773, + 0.012990341521799564, + 0.009517163969576359, + 0.009740911424160004, + 0.010240447707474232, + 0.007051702123135328, + 0.010066593997180462, + 0.002815934829413891, + 0.0023327250964939594, + 0.0031083503272384405, + 0.0029929177835583687, + 0.0049469261430203915, + 0.0026698256842792034, + 0.0038447584956884384, + 0.0032848140690475702, + 0.0009408717742189765, + 0.0009377009700983763, + 0.0022610605228692293, + 0.0019657984375953674, + 0.001433099270798266, + 0.000976301496848464, + 0.0016117647755891085, + 0.0021978644654154778, + 0.0028668029699474573, + 0.002622021362185478, + 0.003192364703863859, + 0.0028692230116575956, + 0.002454182831570506, + 0.00221605459228158, + 0.0020479015074670315, + 0.0030505089089274406, + 0.002538990927860141, + 0.0018990375101566315, + 0.0037552493158727884, + 0.0031272065825760365, + 0.0030371842440217733, + 0.0017877380596473813, + 0.0023461368400603533, + 0.002300969557836652, + 0.0017071899492293596, + 0.001493463758379221, + 0.0021772708278149366, + 0.0014315496664494276, + 0.0016109258867800236, + 0.0008829732541926205, + 0.0015207676915451884, + 0.0011631542583927512, + 0.002412349684163928, + 0.0013288408517837524, + 0.0018348268931731582, + 0.0016856929287314415, + 0.001781654660589993, + 0.001385385636240244, + 0.0016630341997370124, + 0.003424308728426695, + 0.06208787485957146, + 0.02654712088406086, + 0.27236849069595337, + 0.04520785063505173, + 0.03650283440947533, + 0.031952135264873505, + 0.02153482474386692, + 0.23066751658916473, + 0.0029683115426450968, + 0.0025545628741383553, + 0.0039014595095068216, + 0.00349945155903697, + 0.007096554152667522, + 0.0035482337698340416, + 0.0037869911175221205, + 0.004426213912665844, + 0.00896135251969099, + 0.006617475766688585, + 0.008370231837034225, + 0.005858304910361767, + 0.00952264852821827, + 0.006756658665835857, + 0.005615307483822107, + 0.006465018726885319 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.03811749815940857 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133243322372437 + ], + "opposite_logprob_tokens": [ + -1.4063324928283691 + ], + "logprob_reward": -0.28133243322372437, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133243322372437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06371087580919266, + 0.026712743565440178, + 0.02784736268222332, + 0.021231938153505325, + 0.02094198204576969, + 0.03098936378955841, + 0.01952051930129528, + 0.7161365747451782, + 0.04775628447532654, + 0.025152333080768585 + ], + "credit_map": [ + 0.0062706805765628815, + 0.004843403585255146, + 0.015582230873405933, + 0.009665487334132195, + 0.004977615084499121, + 0.007305505685508251, + 0.0058195688761770725, + 0.009246381931006908, + 0.0032126116566359997, + 0.0023945998400449753, + 0.005060585215687752, + 0.002826116746291518, + 0.0037711006589233875, + 0.002828828990459442, + 0.002431312808766961, + 0.004187587182968855, + 0.003810354508459568, + 0.003906718920916319, + 0.003937598317861557, + 0.0031687563750892878, + 0.0029944684356451035, + 0.002928550820797682, + 0.002946827095001936, + 0.004154088441282511, + 0.0019268475007265806, + 0.002211527433246374, + 0.0039205062203109264, + 0.003083248157054186, + 0.0026372773572802544, + 0.002719251438975334, + 0.0016007621306926012, + 0.003132516285404563, + 0.0022740464191883802, + 0.0023350557312369347, + 0.0031526638194918633, + 0.003595398971810937, + 0.0020207837224006653, + 0.0024295111652463675, + 0.0017468795413151383, + 0.0033876418601721525, + 0.003124234266579151, + 0.002969274530187249, + 0.006209303624927998, + 0.0037515871226787567, + 0.005291094072163105, + 0.0027433098293840885, + 0.003784323111176491, + 0.0031162367668002844, + 0.002332946052774787, + 0.0020422623492777348, + 0.0026888044085353613, + 0.003447402734309435, + 0.0027512891683727503, + 0.0018313332693651319, + 0.0022451351396739483, + 0.0021813460625708103, + 0.06160154566168785, + 0.02389145828783512, + 0.27720531821250916, + 0.056191857904195786, + 0.03419126570224762, + 0.029917001724243164, + 0.02052069641649723, + 0.212617427110672, + 0.004554429557174444, + 0.004237668123096228, + 0.011991044506430626, + 0.005037826020270586, + 0.005256877280771732, + 0.004859886132180691, + 0.0045802900567650795, + 0.007238262798637152, + 0.0031114122830331326, + 0.0026921492535620928, + 0.004905660171061754, + 0.002786838449537754, + 0.002971275709569454, + 0.002667934400960803, + 0.002128229709342122, + 0.003888832638040185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.28133243322372437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329261600971222 + ], + "opposite_logprob_tokens": [ + -3.7732925415039062 + ], + "logprob_reward": -0.02329261600971222, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329261600971222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061311811208724976, + 0.03373650088906288, + 0.01587330363690853, + 0.07620133459568024, + 0.04499945789575577, + 0.10207110643386841, + 0.036921896040439606, + 0.5569671392440796, + 0.04490312188863754, + 0.02701425738632679 + ], + "credit_map": [ + 0.006861047353595495, + 0.004947812762111425, + 0.010811340063810349, + 0.008425753563642502, + 0.00821369793266058, + 0.006463627330958843, + 0.006210040301084518, + 0.00937849935144186, + 0.005193780176341534, + 0.003384449053555727, + 0.006151868961751461, + 0.004345650319010019, + 0.005634341388940811, + 0.0028694155625998974, + 0.0028217181097716093, + 0.0033352768514305353, + 0.0016646049916744232, + 0.001667100004851818, + 0.002718757139518857, + 0.0015767950098961592, + 0.0023488819133490324, + 0.0018522210884839296, + 0.0017769953701645136, + 0.002267950214445591, + 0.007687798701226711, + 0.005960838403552771, + 0.013203473761677742, + 0.009915966540575027, + 0.010088363662362099, + 0.007526493165642023, + 0.010199823416769505, + 0.011618583463132381, + 0.005936755333095789, + 0.0034868393559008837, + 0.007973306812345982, + 0.00851544737815857, + 0.003679566318169236, + 0.004556318279355764, + 0.0039222645573318005, + 0.006928964052349329, + 0.00897844135761261, + 0.009766804054379463, + 0.012997916899621487, + 0.015644153580069542, + 0.012671885080635548, + 0.0074573419988155365, + 0.022584866732358932, + 0.011969707906246185, + 0.004598267842084169, + 0.003026080783456564, + 0.004629166331142187, + 0.004537360277026892, + 0.0044599901884794235, + 0.003288229228928685, + 0.005075096618384123, + 0.007307711523026228, + 0.045274436473846436, + 0.01621200516819954, + 0.2121351659297943, + 0.08499907702207565, + 0.023402955383062363, + 0.023936446756124496, + 0.017649274319410324, + 0.13335782289505005, + 0.004835678730159998, + 0.003871320281177759, + 0.011008542962372303, + 0.006506851874291897, + 0.002862926572561264, + 0.00465110270306468, + 0.0028465702198445797, + 0.008320138789713383, + 0.004160377196967602, + 0.0030644433572888374, + 0.005528899375349283, + 0.0027921267319470644, + 0.00270132371224463, + 0.002701780293136835, + 0.0017579203704372048, + 0.004307388793677092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.02329261600971222 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 static, slightly low-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a police officer in his forties wearing a dark blue uniform and a high-visibility yellow traffic vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the officer's hands covered with bare white gloves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the officer perform the action of raising one gloved hand to stop yellow taxis and sedans while beckoning other cars to proceed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a bustling city intersection with fiberglass skyscrapers, long shadows from golden sunlight, and blurred traffic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191539026796818 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191539026796818, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191539026796818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10793223977088928, + 0.04328358918428421, + 0.06751257926225662, + 0.020280582830309868, + 0.013046718202531338, + 0.024038219824433327, + 0.01538468711078167, + 0.5932868123054504, + 0.041080523282289505, + 0.07415404170751572 + ], + "credit_map": [ + 0.012724407017230988, + 0.011618280783295631, + 0.024307873100042343, + 0.011816776357591152, + 0.010452254675328732, + 0.011184250004589558, + 0.010240799747407436, + 0.015587597154080868, + 0.007107837591320276, + 0.0036258185282349586, + 0.009392852894961834, + 0.006445787847042084, + 0.0034284659195691347, + 0.004257328808307648, + 0.0027613090351223946, + 0.006264193449169397, + 0.007384431082755327, + 0.007689299061894417, + 0.013701380230486393, + 0.009416920132935047, + 0.005525731015950441, + 0.007993062026798725, + 0.005112193059176207, + 0.010689567774534225, + 0.0032538280356675386, + 0.0023228107020258904, + 0.0030696752946823835, + 0.002288956195116043, + 0.0027989817317575216, + 0.002064532833173871, + 0.0020860263612121344, + 0.00239577260799706, + 0.0013597970828413963, + 0.0010173412738367915, + 0.0028768975753337145, + 0.0019649553578346968, + 0.001299245865084231, + 0.001032513566315174, + 0.001583741744980216, + 0.001912226201966405, + 0.003530853660777211, + 0.0032002145890146494, + 0.004076613578945398, + 0.002767007565125823, + 0.0026039741933345795, + 0.002031872281804681, + 0.0033471831120550632, + 0.002480502240359783, + 0.0022353532258421183, + 0.0020825816318392754, + 0.002500380389392376, + 0.0021210620179772377, + 0.001839119358919561, + 0.001411250326782465, + 0.0016030509723350406, + 0.0015918903518468142, + 0.0509645901620388, + 0.019658226519823074, + 0.230953186750412, + 0.05422929674386978, + 0.027820667251944542, + 0.02461806684732437, + 0.016391290351748466, + 0.1686515063047409, + 0.005074033048003912, + 0.004501572344452143, + 0.006712272297590971, + 0.00418383814394474, + 0.002463790588080883, + 0.005148170981556177, + 0.0028336187824606895, + 0.010163228958845139, + 0.0072361440397799015, + 0.006984270177781582, + 0.01592935435473919, + 0.012743537314236164, + 0.0036381343379616737, + 0.007760175038129091, + 0.0041914572939276695, + 0.01567097008228302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.014191539026796818 + } + ], + "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/0000122/group_0/index_tensor([6626, 6626])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_0/index_tensor([6626, 6626])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..88246a88cc5e8cabf2f0939c318f3d91120a371f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_0/index_tensor([6626, 6626])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins at an extremely low angle, against a backdrop of pure, endless, minimalist black void. At the very bottom of the frame, a thin, irregular horizontal crack suddenly tears through the darkness, bursting forth a blade-sharp, pure white beam of light. This beam pierces the silent void with a resolute, upward force, while the camera tilts upward in a steady, deliberate motion, gracefully following its vertical trajectory. The beam itself is incredibly intense, with crisp edges, creating extreme contrast against the pure black background. A nearly invisible, subtle, volumetric haze of light permeates the image, imbuing it with a tangible quality. The entire image is rendered in stark black and white, devoid of any color. This minimalist visual language and the slow, solemn camera movements create an atmosphere of mystery, solemnity, and dramatic tension, as if witnessing a moment of creation or revelation.", + "seed": 32833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_0/index_6626_seed_32833_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": -0.11323703296948225, + "vlm_logprob": -0.11323703296948225, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.11323703296948225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22765274345874786 + ], + "opposite_logprob_tokens": [ + -1.6026527881622314 + ], + "logprob_reward": -0.22765274345874786, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22765274345874786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3472253680229187, + 0.03139965981245041, + 0.23128925263881683, + 0.13905787467956543, + 0.04028454050421715, + 0.024929111823439598, + 0.04109025001525879, + 0.0884738564491272, + 0.027841687202453613, + 0.028408391401171684 + ], + "credit_map": [ + 0.04091415926814079, + 0.02281738817691803, + 0.0632811188697815, + 0.0340103916823864, + 0.03453243896365166, + 0.05322861298918724, + 0.025290289893746376, + 0.07315097004175186, + 0.004855447914451361, + 0.002012991113588214, + 0.004680928774178028, + 0.003103477880358696, + 0.003119996516034007, + 0.005046425852924585, + 0.001770100207068026, + 0.00681029399856925, + 0.05005225911736488, + 0.01453587505966425, + 0.04859914630651474, + 0.011200525797903538, + 0.018328165635466576, + 0.038024336099624634, + 0.013243435882031918, + 0.03730550408363342, + 0.02455783262848854, + 0.006431138142943382, + 0.039109572768211365, + 0.0056283376179635525, + 0.013643302023410797, + 0.021921653300523758, + 0.009164491668343544, + 0.018601544201374054, + 0.005109915044158697, + 0.0020402518566697836, + 0.012265371158719063, + 0.0029048360884189606, + 0.0021703133825212717, + 0.008192051202058792, + 0.0020698055159300566, + 0.00553199602290988, + 0.0029299112502485514, + 0.0012191036948934197, + 0.007448991294950247, + 0.0017395822796970606, + 0.0013845390640199184, + 0.0048843491822481155, + 0.0018764544511213899, + 0.0034461806062608957, + 0.006586494389921427, + 0.0029904444236308336, + 0.008106770925223827, + 0.003249842906370759, + 0.00309602334164083, + 0.006578974425792694, + 0.0037244439590722322, + 0.0067572565749287605, + 0.015694061294198036, + 0.004656870383769274, + 0.024152299389243126, + 0.0032849751878529787, + 0.009319464676082134, + 0.014603677205741405, + 0.00774456188082695, + 0.009017943404614925, + 0.004485892131924629, + 0.0019118856871500611, + 0.005428000818938017, + 0.0026528150774538517, + 0.0023413768503814936, + 0.004388012923300266, + 0.002986281644552946, + 0.003647421021014452, + 0.00385528476908803, + 0.0014353542355820537, + 0.010078901425004005, + 0.001749167568050325, + 0.0013414503773674369, + 0.005885169841349125, + 0.0013219470856711268, + 0.002741116564720869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.22765274345874786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307057097554207 + ], + "opposite_logprob_tokens": [ + -3.7733070850372314 + ], + "logprob_reward": -0.023307057097554207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307057097554207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5080148577690125, + 0.03366900607943535, + 0.03956672176718712, + 0.1585981845855713, + 0.05034135654568672, + 0.01910867542028427, + 0.062305569648742676, + 0.06945989280939102, + 0.03861024230718613, + 0.02032552845776081 + ], + "credit_map": [ + 0.06721494346857071, + 0.038173649460077286, + 0.08384089916944504, + 0.03939558565616608, + 0.053423866629600525, + 0.07328596711158752, + 0.03375488892197609, + 0.11892500519752502, + 0.00430505583062768, + 0.0034645989071577787, + 0.00497361458837986, + 0.0043943473137915134, + 0.004098689183592796, + 0.005272678565233946, + 0.0025658539962023497, + 0.004594163037836552, + 0.006769297178834677, + 0.002734086010605097, + 0.008020053617656231, + 0.004080069251358509, + 0.004266028758138418, + 0.005570664070546627, + 0.0025704731233417988, + 0.005556046962738037, + 0.029005032032728195, + 0.007197307422757149, + 0.044918857514858246, + 0.005807145964354277, + 0.01585232838988304, + 0.025385411456227303, + 0.010521654970943928, + 0.019910424947738647, + 0.007028593216091394, + 0.0037532662972807884, + 0.010257263667881489, + 0.004094731993973255, + 0.004327279981225729, + 0.009385132230818272, + 0.0045512220822274685, + 0.006943862419575453, + 0.0026946659199893475, + 0.0017237861175090075, + 0.003970499616116285, + 0.0020379628986120224, + 0.0014505404978990555, + 0.0029118526726961136, + 0.0018863774603232741, + 0.0024329889565706253, + 0.010250275023281574, + 0.003751640673726797, + 0.01512299757450819, + 0.0038516607601195574, + 0.005065164994448423, + 0.009893158450722694, + 0.005763660650700331, + 0.008607003837823868, + 0.010479561984539032, + 0.003447916591539979, + 0.01859380304813385, + 0.0033489400520920753, + 0.007695693057030439, + 0.011846426874399185, + 0.00617709755897522, + 0.007870444096624851, + 0.006743448320776224, + 0.0024168759118765593, + 0.007447793148458004, + 0.0027324496768414974, + 0.004397343844175339, + 0.005725450813770294, + 0.004463331773877144, + 0.00468354532495141, + 0.003029086859896779, + 0.0008131968206726015, + 0.007097622379660606, + 0.0009676323970779777, + 0.0009330614120699465, + 0.0039855255745351315, + 0.0011220704764127731, + 0.002377330558374524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.023307057097554207 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005648924270644784 + ], + "opposite_logprob_tokens": [ + -7.500565052032471 + ], + "logprob_reward": -0.0005648924270644784, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005648924270644784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2676190137863159, + 0.05368202179670334, + 0.15200673043727875, + 0.0899868905544281, + 0.06506399810314178, + 0.053221914917230606, + 0.08512917160987854, + 0.10887611657381058, + 0.050131019204854965, + 0.0742831602692604 + ], + "credit_map": [ + 0.0333850122988224, + 0.018793955445289612, + 0.043486587703228, + 0.021905096247792244, + 0.024296754971146584, + 0.041298218071460724, + 0.01966593600809574, + 0.06478746980428696, + 0.007770350202918053, + 0.005145006813108921, + 0.007419350557029247, + 0.006200478877872229, + 0.008482029661536217, + 0.008197237737476826, + 0.0052176546305418015, + 0.0052499123848974705, + 0.0360708124935627, + 0.008157258853316307, + 0.0291048064827919, + 0.008639851585030556, + 0.012407385744154453, + 0.02491034008562565, + 0.009001093916594982, + 0.0237151887267828, + 0.014025822281837463, + 0.003782865358516574, + 0.029872631654143333, + 0.0034144569654017687, + 0.0069908201694488525, + 0.01663650944828987, + 0.004777679685503244, + 0.01048610545694828, + 0.013236659578979015, + 0.002987859072163701, + 0.016193177551031113, + 0.0029347355011850595, + 0.005122657399624586, + 0.010940108448266983, + 0.0032142368145287037, + 0.010434565134346485, + 0.008269820362329483, + 0.0031244256533682346, + 0.010713574476540089, + 0.005178829655051231, + 0.0032429127022624016, + 0.010377260856330395, + 0.0036091029178351164, + 0.008705984801054, + 0.013254234567284584, + 0.005758083891123533, + 0.019854163751006126, + 0.005988067481666803, + 0.006688717752695084, + 0.013487969525158405, + 0.008616313338279724, + 0.011481618508696556, + 0.015157431364059448, + 0.005690542049705982, + 0.030743112787604332, + 0.007419184781610966, + 0.008844085037708282, + 0.019622802734375, + 0.009180096909403801, + 0.012218858115375042, + 0.008687463589012623, + 0.003044435055926442, + 0.00985485129058361, + 0.003603181801736355, + 0.004988803993910551, + 0.008093141950666904, + 0.005534673109650612, + 0.006324467249214649, + 0.010513774119317532, + 0.0039992062374949455, + 0.014636973850429058, + 0.009602845646440983, + 0.004017912317067385, + 0.013042925857007504, + 0.006451623514294624, + 0.012017905712127686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0005648924270644784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011856677010655403 + ], + "opposite_logprob_tokens": [ + -6.751185894012451 + ], + "logprob_reward": -0.0011856677010655403, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011856677010655403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12881335616111755, + 0.03935818001627922, + 0.14093221724033356, + 0.04789328575134277, + 0.022783247753977776, + 0.012664983980357647, + 0.031841784715652466, + 0.23901315033435822, + 0.09351592510938644, + 0.24318386614322662 + ], + "credit_map": [ + 0.013455267064273357, + 0.01178457960486412, + 0.024155544117093086, + 0.015700295567512512, + 0.01678934507071972, + 0.019476808607578278, + 0.01110995002090931, + 0.016341563314199448, + 0.007864249870181084, + 0.0035052653402090073, + 0.006200746167451143, + 0.002912045456469059, + 0.00404542637988925, + 0.007624004036188126, + 0.002251253928989172, + 0.004955185577273369, + 0.028179816901683807, + 0.01030068937689066, + 0.026460200548171997, + 0.008168470114469528, + 0.012532207183539867, + 0.026332950219511986, + 0.009861915372312069, + 0.019095974043011665, + 0.0072570196352899075, + 0.001693203579634428, + 0.01636982336640358, + 0.0017925284337252378, + 0.002926927525550127, + 0.00918585155159235, + 0.0020506412256509066, + 0.006617290433496237, + 0.002855283673852682, + 0.0008558775880374014, + 0.008792154490947723, + 0.0011349129490554333, + 0.0010016064625233412, + 0.0049135722219944, + 0.0008692855481058359, + 0.002360555110499263, + 0.001481462735682726, + 0.0005851924070157111, + 0.004130878020077944, + 0.0007742319139651954, + 0.0007568977307528257, + 0.002868515206500888, + 0.0007659726543352008, + 0.0013018329627811909, + 0.004347268026322126, + 0.0018735663034021854, + 0.00893635954707861, + 0.0019659181125462055, + 0.0031203965190798044, + 0.005194824188947678, + 0.002647324465215206, + 0.0037561289500445127, + 0.03755815327167511, + 0.012507185339927673, + 0.06457218527793884, + 0.012016864493489265, + 0.023010019212961197, + 0.040880847722291946, + 0.020999526605010033, + 0.027468372136354446, + 0.01671614870429039, + 0.004852785263210535, + 0.021639829501509666, + 0.003650261089205742, + 0.013077839277684689, + 0.013050341047346592, + 0.00979014951735735, + 0.010738572105765343, + 0.037338320165872574, + 0.012696944177150726, + 0.04714101552963257, + 0.030087294057011604, + 0.013881569728255272, + 0.04119131714105606, + 0.020803486928343773, + 0.040043916553258896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.0011856677010655403 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134748041629791 + ], + "opposite_logprob_tokens": [ + -1.3134747743606567 + ], + "logprob_reward": -0.3134748041629791, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134748041629791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09849011898040771, + 0.07473716139793396, + 0.27513638138771057, + 0.1518123596906662, + 0.0844789370894432, + 0.09314535558223724, + 0.041779935359954834, + 0.07766620069742203, + 0.07972240447998047, + 0.02303117886185646 + ], + "credit_map": [ + 0.009761866182088852, + 0.009043493308126926, + 0.016935566440224648, + 0.01252775639295578, + 0.009693165309727192, + 0.013485933654010296, + 0.011231101118028164, + 0.01581123284995556, + 0.009204218164086342, + 0.0060961563140153885, + 0.009226907975971699, + 0.00799141637980938, + 0.012725546956062317, + 0.011491614393889904, + 0.007041112519800663, + 0.01096019335091114, + 0.03328055515885353, + 0.012060326524078846, + 0.09257679432630539, + 0.014637868851423264, + 0.013686339370906353, + 0.05289661884307861, + 0.010325955227017403, + 0.04567192494869232, + 0.025832103565335274, + 0.007164597511291504, + 0.047100648283958435, + 0.005789883900433779, + 0.014688605442643166, + 0.025716479867696762, + 0.008661245927214622, + 0.016858790069818497, + 0.01134069450199604, + 0.005727818701416254, + 0.017598213627934456, + 0.006848666351288557, + 0.006450805813074112, + 0.015605133958160877, + 0.0075838458724319935, + 0.013323762454092503, + 0.013116689398884773, + 0.005121156573295593, + 0.019962415099143982, + 0.009019331075251102, + 0.005402224138379097, + 0.01783912256360054, + 0.007043052930384874, + 0.015641363337635994, + 0.007322043180465698, + 0.0028286322485655546, + 0.008426526561379433, + 0.0029167113825678825, + 0.003727254457771778, + 0.006208863575011492, + 0.004392331000417471, + 0.0059575727209448814, + 0.012713432312011719, + 0.00452874880284071, + 0.01847546175122261, + 0.004007354844361544, + 0.008688953705132008, + 0.012212414294481277, + 0.008252727799117565, + 0.008787105791270733, + 0.013720710761845112, + 0.004883668851107359, + 0.014588909223675728, + 0.005937166046351194, + 0.009470529854297638, + 0.011296888813376427, + 0.0096760718151927, + 0.010148459114134312, + 0.002937633777037263, + 0.001129743061028421, + 0.007412779610604048, + 0.0020607460755854845, + 0.0010434547439217567, + 0.004303494933992624, + 0.0015629726694896817, + 0.0025803539901971817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.3134748041629791 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins at an extremely low angle, against a backdrop of pure, endless, minimalist black void. At the very bottom of the frame, a thin, irregular horizontal crack suddenly tears through the darkness, bursting forth a blade-sharp, pure white beam of light. This beam pierces the silent void with a resolute, upward force, while the camera tilts upward in a steady, deliberate motion, gracefully following its vertical trajectory. The beam itself is incredibly intense, with crisp edges, creating extreme contrast against the pure black background. A nearly invisible, subtle, volumetric haze of light permeates the image, imbuing it with a tangible quality. The entire image is rendered in stark black and white, devoid of any color. This minimalist visual language and the slow, solemn camera movements create an atmosphere of mystery, solemnity, and dramatic tension, as if witnessing a moment of creation or revelation.", + "seed": 32834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_0/index_6626_seed_32834_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": -0.12868359104031696, + "vlm_logprob": -0.12868359104031696, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.12868359104031696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2544514238834381 + ], + "opposite_logprob_tokens": [ + -1.5044513940811157 + ], + "logprob_reward": -0.2544514238834381, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2544514238834381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18999448418617249, + 0.024051962420344353, + 0.05549990013241768, + 0.036251865327358246, + 0.07211314886808395, + 0.23074978590011597, + 0.03438832610845566, + 0.057245176285505295, + 0.05808524787425995, + 0.24162015318870544 + ], + "credit_map": [ + 0.029640458524227142, + 0.01016948465257883, + 0.033800531178712845, + 0.01271112821996212, + 0.013174213469028473, + 0.026493653655052185, + 0.01818963512778282, + 0.045815370976924896, + 0.0031867274083197117, + 0.0028507686220109463, + 0.0031273835338652134, + 0.0029172173235565424, + 0.0031200989615172148, + 0.003306564874947071, + 0.0029557948000729084, + 0.002587402705103159, + 0.00828020554035902, + 0.005061498377472162, + 0.005981663707643747, + 0.007813258096575737, + 0.005503359716385603, + 0.010277988389134407, + 0.006250610109418631, + 0.006331313401460648, + 0.0040419381111860275, + 0.0024184267967939377, + 0.005159514024853706, + 0.004070596769452095, + 0.0049657318741083145, + 0.0050847893580794334, + 0.006249948870390654, + 0.004260913468897343, + 0.013270094990730286, + 0.0029775563161820173, + 0.00878959521651268, + 0.00250286515802145, + 0.006211318075656891, + 0.015019772574305534, + 0.00421510636806488, + 0.019126836210489273, + 0.021350828930735588, + 0.01963147707283497, + 0.03943757340312004, + 0.0280904620885849, + 0.0439583882689476, + 0.027345340698957443, + 0.03307278826832771, + 0.017862912267446518, + 0.0028928311076015234, + 0.0029907263815402985, + 0.005668023135513067, + 0.0038266924675554037, + 0.005157774779945612, + 0.006561798043549061, + 0.003606335958465934, + 0.003684140043333173, + 0.007186905946582556, + 0.00625400897115469, + 0.012108721770346165, + 0.004162227269262075, + 0.008743183687329292, + 0.005857264157384634, + 0.007290740497410297, + 0.00564211793243885, + 0.005450609605759382, + 0.005914933513849974, + 0.020980726927518845, + 0.004853182006627321, + 0.007431528996676207, + 0.005446250084787607, + 0.004801302216947079, + 0.0032067117281258106, + 0.013374564237892628, + 0.023829910904169083, + 0.05115041509270668, + 0.04450750723481178, + 0.028558116406202316, + 0.025845464318990707, + 0.039457082748413086, + 0.014897088520228863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.2544514238834381 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812220320105553 + ], + "opposite_logprob_tokens": [ + -3.2881221771240234 + ], + "logprob_reward": -0.03812220320105553, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812220320105553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5383256673812866, + 0.06871573626995087, + 0.02658178098499775, + 0.02074856124818325, + 0.12855900824069977, + 0.05603659898042679, + 0.05496169626712799, + 0.015482250601053238, + 0.054922789335250854, + 0.03566594049334526 + ], + "credit_map": [ + 0.08368857949972153, + 0.03548968583345413, + 0.09064833074808121, + 0.037629250437021255, + 0.039523202925920486, + 0.07957136631011963, + 0.05720851197838783, + 0.11456674337387085, + 0.014291453175246716, + 0.005572622176259756, + 0.005959492642432451, + 0.005216959398239851, + 0.006360020022839308, + 0.016468172892928123, + 0.004847926087677479, + 0.009999092668294907, + 0.003948528785258532, + 0.0022128575947135687, + 0.003837637137621641, + 0.003097333014011383, + 0.002518570050597191, + 0.004022292327135801, + 0.003287774743512273, + 0.0036567889619618654, + 0.0031250424217432737, + 0.0009246928966604173, + 0.0024514158722013235, + 0.0016179149970412254, + 0.002519601257517934, + 0.004410455469042063, + 0.002122282050549984, + 0.0035771569237113, + 0.014104672707617283, + 0.0034141954965889454, + 0.015534613281488419, + 0.005128560122102499, + 0.01267445832490921, + 0.03398427367210388, + 0.007566644344478846, + 0.036151591688394547, + 0.004575982224196196, + 0.004082309082150459, + 0.011335326358675957, + 0.004040419589728117, + 0.007475389167666435, + 0.009933870285749435, + 0.005828964989632368, + 0.008764337748289108, + 0.0068103172816336155, + 0.006028297357261181, + 0.00790688768029213, + 0.004827707074582577, + 0.009110416285693645, + 0.006350474897772074, + 0.007800014689564705, + 0.006127579137682915, + 0.0012376665836200118, + 0.0020639945287257433, + 0.0021883943118155003, + 0.0017588644986972213, + 0.0024582259356975555, + 0.001933504594489932, + 0.0024317458737641573, + 0.0014098542742431164, + 0.005225020460784435, + 0.005818504840135574, + 0.01792706921696663, + 0.0061431340873241425, + 0.006525930017232895, + 0.00500575453042984, + 0.004957784432917833, + 0.0033195889554917812, + 0.0021507777273654938, + 0.003736268263310194, + 0.008025716058909893, + 0.006429512053728104, + 0.004439431242644787, + 0.00355389597825706, + 0.005082000978291035, + 0.0022483374923467636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03812220320105553 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006403064471669495 + ], + "opposite_logprob_tokens": [ + -7.375640392303467 + ], + "logprob_reward": -0.0006403064471669495, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006403064471669495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10566362738609314, + 0.045178454369306564, + 0.11025162786245346, + 0.05376797914505005, + 0.16808569431304932, + 0.16097937524318695, + 0.1071370467543602, + 0.07835777848958969, + 0.09831856191158295, + 0.07225985080003738 + ], + "credit_map": [ + 0.010037007741630077, + 0.016881050541996956, + 0.01357994694262743, + 0.009001316502690315, + 0.015002463944256306, + 0.014011536724865437, + 0.012267162092030048, + 0.014883141033351421, + 0.0055963629856705666, + 0.006074519362300634, + 0.0053054229356348515, + 0.007377843372523785, + 0.004314114339649677, + 0.004583681467920542, + 0.00652554165571928, + 0.00540096964687109, + 0.0150354178622365, + 0.006113892886787653, + 0.011141885071992874, + 0.009057809598743916, + 0.01030820608139038, + 0.02819180302321911, + 0.01015841867774725, + 0.02024419791996479, + 0.005963361356407404, + 0.00358310854062438, + 0.005675738677382469, + 0.004650590941309929, + 0.007471555378288031, + 0.00883126724511385, + 0.010086978785693645, + 0.007505378220230341, + 0.022519951686263084, + 0.004723831079900265, + 0.045161355286836624, + 0.006862264592200518, + 0.015297116711735725, + 0.035546381026506424, + 0.012985734269022942, + 0.024989061057567596, + 0.015493843704462051, + 0.01258936244994402, + 0.034440722316503525, + 0.016914276406168938, + 0.016652530059218407, + 0.026195617392659187, + 0.014694594778120518, + 0.02399841696023941, + 0.012803589925169945, + 0.00879413727670908, + 0.022983061149716377, + 0.009695579297840595, + 0.011462152004241943, + 0.01522595714777708, + 0.014599445275962353, + 0.011573129333555698, + 0.007854054681956768, + 0.008712428621947765, + 0.014554843306541443, + 0.00730766961351037, + 0.011868253350257874, + 0.010103050619363785, + 0.011834625154733658, + 0.006122851278632879, + 0.009169301018118858, + 0.011670847423374653, + 0.030646255239844322, + 0.008375266566872597, + 0.01379967387765646, + 0.008495974354445934, + 0.009482257068157196, + 0.006678982637822628, + 0.004757194314152002, + 0.007456093095242977, + 0.017939109355211258, + 0.014570357277989388, + 0.008378040976822376, + 0.006951627321541309, + 0.008127320557832718, + 0.004080102313309908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.0006403064471669495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015236446633934975 + ], + "opposite_logprob_tokens": [ + -6.501523494720459 + ], + "logprob_reward": -0.0015236446633934975, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015236446633934975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08960891515016556, + 0.05241912603378296, + 0.027733689174056053, + 0.05209367722272873, + 0.078452467918396, + 0.26371344923973083, + 0.14263059198856354, + 0.045870330184698105, + 0.06167088821530342, + 0.18580684065818787 + ], + "credit_map": [ + 0.010998053476214409, + 0.008689717389643192, + 0.013087757863104343, + 0.007930015213787556, + 0.00804597232490778, + 0.014051097445189953, + 0.009823702275753021, + 0.01698259636759758, + 0.008601732552051544, + 0.006904342211782932, + 0.005465199239552021, + 0.0064862859435379505, + 0.005243979394435883, + 0.01003735326230526, + 0.004788530524820089, + 0.004891700576990843, + 0.0030497699044644833, + 0.0027941532898694277, + 0.005231079645454884, + 0.003938727080821991, + 0.0031374460086226463, + 0.0037489947862923145, + 0.0026468550786376, + 0.003186663379892707, + 0.006481590215116739, + 0.0027872512582689524, + 0.006761839613318443, + 0.004977370146661997, + 0.005490671377629042, + 0.01083820965141058, + 0.007231149356812239, + 0.007525595370680094, + 0.013440464623272419, + 0.0037140846252441406, + 0.009232130832970142, + 0.004175012465566397, + 0.007044072262942791, + 0.015217596665024757, + 0.005699772387742996, + 0.019929341971874237, + 0.022363867610692978, + 0.02081136777997017, + 0.04047265648841858, + 0.026927312836050987, + 0.06321553140878677, + 0.028687341138720512, + 0.04469817131757736, + 0.01653718389570713, + 0.01947091333568096, + 0.014368616044521332, + 0.0245502982288599, + 0.011020053178071976, + 0.024379553273320198, + 0.015717243775725365, + 0.018578264862298965, + 0.014545640908181667, + 0.00477022398263216, + 0.0049119447357952595, + 0.009365665726363659, + 0.004843367729336023, + 0.004614559467881918, + 0.005623192992061377, + 0.006384019274264574, + 0.005357355810701847, + 0.005911706481128931, + 0.006951115559786558, + 0.016260067000985146, + 0.007140816189348698, + 0.008520369417965412, + 0.00534829730167985, + 0.006601936649531126, + 0.004936577286571264, + 0.01139499619603157, + 0.019073273986577988, + 0.035293761640787125, + 0.033923014998435974, + 0.02193673513829708, + 0.020996564999222755, + 0.03130468353629112, + 0.011883820407092571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.0015236446633934975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34868037700653076 + ], + "opposite_logprob_tokens": [ + -1.2236803770065308 + ], + "logprob_reward": -0.34868037700653076, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34868037700653076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18210861086845398, + 0.14624141156673431, + 0.06002924218773842, + 0.08220222592353821, + 0.07294673472642899, + 0.0758366659283638, + 0.06838627904653549, + 0.21246479451656342, + 0.058418817818164825, + 0.04136531800031662 + ], + "credit_map": [ + 0.022150136530399323, + 0.015832029283046722, + 0.030654704198241234, + 0.014639649540185928, + 0.015410183928906918, + 0.023017656058073044, + 0.020176006481051445, + 0.04022820666432381, + 0.02495238557457924, + 0.018736451864242554, + 0.016490863636136055, + 0.018529776483774185, + 0.014603178948163986, + 0.024823253974318504, + 0.013500343076884747, + 0.014605146832764149, + 0.009970786049962044, + 0.0031151161529123783, + 0.005930114537477493, + 0.0033846006263047457, + 0.005193792283535004, + 0.01863119751214981, + 0.0032021068036556244, + 0.01060151495039463, + 0.009783518500626087, + 0.003951490391045809, + 0.009996545501053333, + 0.007900161668658257, + 0.008452525362372398, + 0.017422493547201157, + 0.012810594402253628, + 0.011884881183505058, + 0.009945104829967022, + 0.0020605497993528843, + 0.019983265548944473, + 0.002891052281484008, + 0.006061061751097441, + 0.01088386308401823, + 0.005329503677785397, + 0.01579231768846512, + 0.005630378145724535, + 0.005225378088653088, + 0.010320194065570831, + 0.005101121496409178, + 0.016678201034665108, + 0.011190854012966156, + 0.011734076775610447, + 0.009956441819667816, + 0.006778402253985405, + 0.006286310497671366, + 0.012559178285300732, + 0.006624285597354174, + 0.00804054457694292, + 0.008991634473204613, + 0.011600792407989502, + 0.007505119778215885, + 0.025648275390267372, + 0.02408454939723015, + 0.032665640115737915, + 0.01584131456911564, + 0.04511701688170433, + 0.021430307999253273, + 0.03344292938709259, + 0.014234730042517185, + 0.0052043478935956955, + 0.007840071804821491, + 0.014495455659925938, + 0.006435989867895842, + 0.007775337900966406, + 0.005851720925420523, + 0.0068408409133553505, + 0.0039750393480062485, + 0.0020970781333744526, + 0.00504394993185997, + 0.010557244531810284, + 0.007094386033713818, + 0.0049634333699941635, + 0.003550398163497448, + 0.00591669837012887, + 0.0021421201527118683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.34868037700653076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins at an extremely low angle, against a backdrop of pure, endless, minimalist black void. At the very bottom of the frame, a thin, irregular horizontal crack suddenly tears through the darkness, bursting forth a blade-sharp, pure white beam of light. This beam pierces the silent void with a resolute, upward force, while the camera tilts upward in a steady, deliberate motion, gracefully following its vertical trajectory. The beam itself is incredibly intense, with crisp edges, creating extreme contrast against the pure black background. A nearly invisible, subtle, volumetric haze of light permeates the image, imbuing it with a tangible quality. The entire image is rendered in stark black and white, devoid of any color. This minimalist visual language and the slow, solemn camera movements create an atmosphere of mystery, solemnity, and dramatic tension, as if witnessing a moment of creation or revelation.", + "seed": 32835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_0/index_6626_seed_32835_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": -0.18258516711648554, + "vlm_logprob": -0.18258516711648554, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.18258516711648554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6356212496757507 + ], + "opposite_logprob_tokens": [ + -0.7606212496757507 + ], + "logprob_reward": -0.6356212496757507, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6356212496757507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07578454166650772, + 0.05553552880883217, + 0.053945042192935944, + 0.006180332973599434, + 0.04710794985294342, + 0.2714076042175293, + 0.23431678116321564, + 0.018100151792168617, + 0.228720024228096, + 0.008902059867978096 + ], + "credit_map": [ + 0.006719390861690044, + 0.015062007121741772, + 0.008097515441477299, + 0.004868560004979372, + 0.015849877148866653, + 0.013770140707492828, + 0.0023534938227385283, + 0.00906355120241642, + 0.004897673614323139, + 0.015662280842661858, + 0.004522595088928938, + 0.004384366329759359, + 0.0030065125320106745, + 0.016636541113257408, + 0.0025368211790919304, + 0.0038887308910489082, + 0.005245306063443422, + 0.003356492379680276, + 0.023244168609380722, + 0.00131420255638659, + 0.0016699293628335, + 0.015409538522362709, + 0.0011078451061621308, + 0.002597557846456766, + 0.0009053617832250893, + 0.0007703255396336317, + 0.0014198233839124441, + 0.0004974325420334935, + 0.0006343388231471181, + 0.0009421373251825571, + 0.00043154554441571236, + 0.0005793676245957613, + 0.005095505155622959, + 0.006685222033411264, + 0.019727224484086037, + 0.0018087959615513682, + 0.0019389253575354815, + 0.0057229758240282536, + 0.0015176351880654693, + 0.004611663054674864, + 0.040526315569877625, + 0.0975482165813446, + 0.030705779790878296, + 0.010330941528081894, + 0.03961176052689552, + 0.027287645265460014, + 0.005582612007856369, + 0.01981431432068348, + 0.038987450301647186, + 0.024703485891222954, + 0.02868555672466755, + 0.01813848875463009, + 0.026553766801953316, + 0.038713183254003525, + 0.010231543332338333, + 0.04830331355333328, + 0.0011254593264311552, + 0.0021351242903620005, + 0.002827426651492715, + 0.0025249728932976723, + 0.0032491059973835945, + 0.002236616099253297, + 0.002505415352061391, + 0.0014960292028263211, + 0.01364844385534525, + 0.022239549085497856, + 0.06669113785028458, + 0.05055972561240196, + 0.02390158176422119, + 0.01993207074701786, + 0.020075565204024315, + 0.011671924032270908, + 0.0006466611521318555, + 0.0009537470177747309, + 0.0012323788832873106, + 0.0014394735917448997, + 0.0016788390930742025, + 0.0012880105059593916, + 0.0011081958655267954, + 0.0005547534092329443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.6356212496757507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060566283762455 + ], + "opposite_logprob_tokens": [ + -3.8956055641174316 + ], + "logprob_reward": -0.02060566283762455, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060566283762455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4691406786441803, + 0.08023128658533096, + 0.07015973329544067, + 0.01077944878488779, + 0.03303094208240509, + 0.14224611222743988, + 0.16547653079032898, + 0.009985385462641716, + 0.008247732184827328, + 0.010702195577323437 + ], + "credit_map": [ + 0.019059451296925545, + 0.19673991203308105, + 0.022940441966056824, + 0.014693059027194977, + 0.055592723190784454, + 0.13536818325519562, + 0.005322277080267668, + 0.01942458190023899, + 0.013566173613071442, + 0.014768721535801888, + 0.00989297591149807, + 0.008601325564086437, + 0.004499086178839207, + 0.014167859219014645, + 0.005396363325417042, + 0.009338773787021637, + 0.006642499007284641, + 0.004745809826999903, + 0.029586244374513626, + 0.0022088573314249516, + 0.00313938083127141, + 0.016262412071228027, + 0.0016751625807955861, + 0.005899364594370127, + 0.0015001259744167328, + 0.0008729805122129619, + 0.004140863660722971, + 0.0004627900489140302, + 0.0005597504787147045, + 0.0021847744937986135, + 0.00042389376903884113, + 0.0006342699052765965, + 0.005142321810126305, + 0.00437308382242918, + 0.010839381255209446, + 0.0012988941743969917, + 0.0014061548281461, + 0.005452810786664486, + 0.0010677353711798787, + 0.003450558055192232, + 0.027348820120096207, + 0.03300083801150322, + 0.016639921814203262, + 0.003958674147725105, + 0.014750111848115921, + 0.017081204801797867, + 0.0041315290145576, + 0.025335008278489113, + 0.026502642780542374, + 0.01770716719329357, + 0.02044445462524891, + 0.013480273075401783, + 0.01909457892179489, + 0.026563478633761406, + 0.006800525356084108, + 0.03488340601325035, + 0.000661779020447284, + 0.001341077033430338, + 0.002723029814660549, + 0.0008645775960758328, + 0.0015360059915110469, + 0.0009683320531621575, + 0.0010662056738510728, + 0.0008243781630881131, + 0.0006227281410247087, + 0.000774565152823925, + 0.0015216137981042266, + 0.001462796819396317, + 0.0009132820996455848, + 0.0010367889190092683, + 0.0010057422332465649, + 0.0009102149051614106, + 0.00046270963503047824, + 0.001214498421177268, + 0.001120600150898099, + 0.001912302104756236, + 0.002168419770896435, + 0.0017295252764597535, + 0.0012857337715104222, + 0.0008084061555564404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.02060566283762455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000925112864933908 + ], + "opposite_logprob_tokens": [ + -7.000925064086914 + ], + "logprob_reward": -0.000925112864933908, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000925112864933908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11517732590436935, + 0.02188774384558201, + 0.05390004813671112, + 0.07265540212392807, + 0.013424969278275967, + 0.2750438153743744, + 0.3901962637901306, + 0.013805199414491653, + 0.021023880690336227, + 0.022885344922542572 + ], + "credit_map": [ + 0.008045589551329613, + 0.04343628138303757, + 0.009158414788544178, + 0.00616740295663476, + 0.023947089910507202, + 0.011551864445209503, + 0.004506457597017288, + 0.008364228531718254, + 0.002320690779015422, + 0.003032147651538253, + 0.0036088544875383377, + 0.0025776494294404984, + 0.0019350209040567279, + 0.0034793862141668797, + 0.0017535508377477527, + 0.0031804435420781374, + 0.005028571933507919, + 0.004452097229659557, + 0.021363548934459686, + 0.0018887658370658755, + 0.0037067660596221685, + 0.01105646975338459, + 0.001376293133944273, + 0.005027531646192074, + 0.008724022656679153, + 0.005209849216043949, + 0.03127974271774292, + 0.0022476266603916883, + 0.0014116587117314339, + 0.011170915327966213, + 0.0016988341230899096, + 0.010912753641605377, + 0.0016491981223225594, + 0.0015874847304075956, + 0.003095165127888322, + 0.0008735014125704765, + 0.0012208359548822045, + 0.0015244991518557072, + 0.0010119974613189697, + 0.0024622874334454536, + 0.052734002470970154, + 0.06656178086996078, + 0.031142735853791237, + 0.00791366957128048, + 0.02963016740977764, + 0.032413601875305176, + 0.007749108597636223, + 0.0468987412750721, + 0.06592737883329391, + 0.04119366407394409, + 0.0459647998213768, + 0.02867160178720951, + 0.04452330246567726, + 0.06562797725200653, + 0.015772037208080292, + 0.08251548558473587, + 0.0009662575903348625, + 0.0016631381586194038, + 0.0021537612192332745, + 0.0018756208010017872, + 0.002067006891593337, + 0.002167297061532736, + 0.0017333969008177519, + 0.0011787201510742307, + 0.0011745209340006113, + 0.0020655475091189146, + 0.0047437883913517, + 0.00518135167658329, + 0.002202190924435854, + 0.002071620663627982, + 0.0022596274502575397, + 0.001325234305113554, + 0.001172605436295271, + 0.002467399463057518, + 0.002857564715668559, + 0.0042936489917337894, + 0.0040361639112234116, + 0.004015293437987566, + 0.002622700296342373, + 0.0014199696015566587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.000925112864933908 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036278641782701015 + ], + "opposite_logprob_tokens": [ + -5.628627777099609 + ], + "logprob_reward": -0.0036278641782701015, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036278641782701015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1663201004266739, + 0.01640475168824196, + 0.019386347383260727, + 0.004478417802602053, + 0.016474461182951927, + 0.5269541144371033, + 0.14268839359283447, + 0.009025707840919495, + 0.08440229296684265, + 0.013865494169294834 + ], + "credit_map": [ + 0.00852256640791893, + 0.07087434083223343, + 0.008599751628935337, + 0.0049043335020542145, + 0.024854136630892754, + 0.035387590527534485, + 0.0031130535062402487, + 0.0100643215700984, + 0.0022926663514226675, + 0.003805824089795351, + 0.001706223818473518, + 0.0015054421965032816, + 0.0010046594543382525, + 0.0035096965730190277, + 0.0010322992457076907, + 0.0015479408903047442, + 0.0030172874685376883, + 0.0015149684622883797, + 0.0034192600287497044, + 0.0006666279514320195, + 0.0015636923490092158, + 0.003982709255069494, + 0.000712656241375953, + 0.0045091453939676285, + 0.0005846490967087448, + 0.0005382475210353732, + 0.0010767570929601789, + 0.00044812128180637956, + 0.0002931742637883872, + 0.0006935272831469774, + 0.00030646781669929624, + 0.0005374736501835287, + 0.0018337896326556802, + 0.0024038604460656643, + 0.006294237449765205, + 0.0008862415561452508, + 0.0007947690901346505, + 0.0018407407915219665, + 0.0006651062867604196, + 0.0017557175597175956, + 0.09073848277330399, + 0.15773916244506836, + 0.06023988127708435, + 0.016674093902111053, + 0.06550490111112595, + 0.05705432966351509, + 0.012431378476321697, + 0.06657188385725021, + 0.02327873930335045, + 0.014709969982504845, + 0.017938129603862762, + 0.010837605223059654, + 0.016424531117081642, + 0.024555029347538948, + 0.0058433241210877895, + 0.02910107746720314, + 0.0005890452885068953, + 0.0009864564053714275, + 0.0009755970677360892, + 0.0011675246059894562, + 0.0013614759081974626, + 0.0013589875306934118, + 0.0015524326590821147, + 0.0010341887827962637, + 0.005156013183295727, + 0.008108360692858696, + 0.023975973948836327, + 0.018777374178171158, + 0.008748721331357956, + 0.007379394955933094, + 0.007581833750009537, + 0.004674619995057583, + 0.0008026223513297737, + 0.001588007085956633, + 0.0024649344850331545, + 0.00245578377507627, + 0.0022050910629332066, + 0.0021270622964948416, + 0.0014139655977487564, + 0.0008080264087766409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0036278641782701015 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521459460258484 + ], + "opposite_logprob_tokens": [ + -1.5021460056304932 + ], + "logprob_reward": -0.2521459460258484, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521459460258484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3122096657752991, + 0.04626567289233208, + 0.09080423414707184, + 0.17588379979133606, + 0.018764449283480644, + 0.03334464132785797, + 0.2590796649456024, + 0.017553169280290604, + 0.03224644809961319, + 0.013848238624632359 + ], + "credit_map": [ + 0.015599406324326992, + 0.13060429692268372, + 0.019630681723356247, + 0.01222052052617073, + 0.04723150655627251, + 0.0576631985604763, + 0.005778029095381498, + 0.023482024669647217, + 0.005467353854328394, + 0.006645040586590767, + 0.007786338217556477, + 0.0061318823136389256, + 0.0038853809237480164, + 0.008572106249630451, + 0.002942920196801424, + 0.004834651481360197, + 0.010002302937209606, + 0.007396162021905184, + 0.02835657075047493, + 0.004521483089774847, + 0.006185044534504414, + 0.020599495619535446, + 0.004535260144621134, + 0.009207922033965588, + 0.02097410522401333, + 0.012596716172993183, + 0.07599276304244995, + 0.005299355834722519, + 0.003145887516438961, + 0.025724289938807487, + 0.003944145515561104, + 0.028206540271639824, + 0.003673233324661851, + 0.0020528773311525583, + 0.002034513046965003, + 0.0009054493275471032, + 0.001731141353957355, + 0.00276420870795846, + 0.0012482444290071726, + 0.004354782402515411, + 0.007079365663230419, + 0.003490753471851349, + 0.003617437556385994, + 0.0013154818443581462, + 0.0027442839927971363, + 0.004791734274476767, + 0.0013363112229853868, + 0.008969269692897797, + 0.04110410064458847, + 0.027609240263700485, + 0.04127828776836395, + 0.016793951392173767, + 0.025599399581551552, + 0.03566578030586243, + 0.012773530557751656, + 0.058255359530448914, + 0.0012356371153146029, + 0.002034887671470642, + 0.0026554532814770937, + 0.0022248951718211174, + 0.0019014852587133646, + 0.002438455121591687, + 0.003045568708330393, + 0.002016786951571703, + 0.0018326743738725781, + 0.003272687317803502, + 0.006639726925641298, + 0.009417813271284103, + 0.0028386441990733147, + 0.0032375003211200237, + 0.003152393037453294, + 0.0018550071399658918, + 0.000901438994333148, + 0.0013823366025462747, + 0.0025830669328570366, + 0.002527301199734211, + 0.0024604550562798977, + 0.001568466774187982, + 0.001355886459350586, + 0.001069286954589188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.2521459460258484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins at an extremely low angle, against a backdrop of pure, endless, minimalist black void. At the very bottom of the frame, a thin, irregular horizontal crack suddenly tears through the darkness, bursting forth a blade-sharp, pure white beam of light. This beam pierces the silent void with a resolute, upward force, while the camera tilts upward in a steady, deliberate motion, gracefully following its vertical trajectory. The beam itself is incredibly intense, with crisp edges, creating extreme contrast against the pure black background. A nearly invisible, subtle, volumetric haze of light permeates the image, imbuing it with a tangible quality. The entire image is rendered in stark black and white, devoid of any color. This minimalist visual language and the slow, solemn camera movements create an atmosphere of mystery, solemnity, and dramatic tension, as if witnessing a moment of creation or revelation.", + "seed": 32836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_0/index_6626_seed_32836_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": -0.2789917239686474, + "vlm_logprob": -0.2789917239686474, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.2789917239686474, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6943718791007996 + ], + "opposite_logprob_tokens": [ + -0.6943718791007996 + ], + "logprob_reward": -0.6943718791007996, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6943718791007996, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06628556549549103, + 0.11335033178329468, + 0.16668780148029327, + 0.13507096469402313, + 0.03833615779876709, + 0.26328375935554504, + 0.06463783979415894, + 0.08089139312505722, + 0.03011869080364704, + 0.041337527334690094 + ], + "credit_map": [ + 0.007517690304666758, + 0.008584525436162949, + 0.01494929101318121, + 0.00793508067727089, + 0.006274393759667873, + 0.010093800723552704, + 0.004687188193202019, + 0.006243598181754351, + 0.01293924916535616, + 0.016150355339050293, + 0.023854220286011696, + 0.02038920298218727, + 0.003799746511504054, + 0.017530718818306923, + 0.005184653680771589, + 0.013502191752195358, + 0.02387569285929203, + 0.021151790395379066, + 0.043334126472473145, + 0.022758711129426956, + 0.008467485196888447, + 0.022329702973365784, + 0.01214633509516716, + 0.012623950839042664, + 0.01879231631755829, + 0.019073355942964554, + 0.027895363047719002, + 0.022515136748552322, + 0.00852466281503439, + 0.015172748826444149, + 0.009962514042854309, + 0.013134859502315521, + 0.0029985690489411354, + 0.005905116908252239, + 0.005059840157628059, + 0.005698568653315306, + 0.0024431031197309494, + 0.006633225828409195, + 0.004899473395198584, + 0.004698260687291622, + 0.02292463183403015, + 0.037875574082136154, + 0.05348414555191994, + 0.05795537307858467, + 0.014264795929193497, + 0.04086664691567421, + 0.015082288533449173, + 0.020830320194363594, + 0.006736763287335634, + 0.010122464038431644, + 0.011967004276812077, + 0.011188595555722713, + 0.005139015149325132, + 0.007662503514438868, + 0.005908927880227566, + 0.005912563763558865, + 0.00757675850763917, + 0.007782313507050276, + 0.019684752449393272, + 0.011690910905599594, + 0.005127264652401209, + 0.015047085471451283, + 0.007007874082773924, + 0.006974435411393642, + 0.002011673292145133, + 0.0036562816239893436, + 0.005103070754557848, + 0.005320474971085787, + 0.004096257034689188, + 0.0033285736571997404, + 0.004683225881308317, + 0.0019191334722563624, + 0.0032629843335598707, + 0.005577444564551115, + 0.006915280595421791, + 0.007925362326204777, + 0.0031126404646784067, + 0.0057461936958134174, + 0.004331361036747694, + 0.004466257989406586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.6943718791007996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309103265404701 + ], + "opposite_logprob_tokens": [ + -3.168091058731079 + ], + "logprob_reward": -0.04309103265404701, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309103265404701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28036272525787354, + 0.05987751483917236, + 0.058270975947380066, + 0.07007566839456558, + 0.1948804408311844, + 0.25852465629577637, + 0.01762511394917965, + 0.014427741058170795, + 0.018484173342585564, + 0.02747098170220852 + ], + "credit_map": [ + 0.02686324343085289, + 0.05266490951180458, + 0.03673289343714714, + 0.027217334136366844, + 0.019820038229227066, + 0.09030639380216599, + 0.011615864001214504, + 0.015142053365707397, + 0.0076369233429431915, + 0.005016916431486607, + 0.022151082754135132, + 0.004395093768835068, + 0.0019522783113643527, + 0.010429291985929012, + 0.003026110352948308, + 0.005269814748317003, + 0.007390368264168501, + 0.008237662725150585, + 0.009984551928937435, + 0.006834131665527821, + 0.0037456986028701067, + 0.008198156952857971, + 0.006945743225514889, + 0.006934661418199539, + 0.007197821978479624, + 0.006843967828899622, + 0.018286140635609627, + 0.008758988231420517, + 0.003297988558188081, + 0.0120840510353446, + 0.008787351660430431, + 0.004819363821297884, + 0.016273146495223045, + 0.011430829763412476, + 0.0675295889377594, + 0.014589918777346611, + 0.005362063646316528, + 0.05476115271449089, + 0.00830763392150402, + 0.016626087948679924, + 0.021547354757785797, + 0.02866952307522297, + 0.07542109489440918, + 0.04467833414673805, + 0.009316964074969292, + 0.04790282994508743, + 0.011841642670333385, + 0.01914691925048828, + 0.00154905638191849, + 0.003325180383399129, + 0.003346869256347418, + 0.002541378140449524, + 0.0013426629593595862, + 0.0032262937165796757, + 0.0011564481537789106, + 0.001137225772254169, + 0.0014642068417742848, + 0.002052565570920706, + 0.002691637258976698, + 0.0029110547620803118, + 0.0008929615141823888, + 0.002140939701348543, + 0.0012397037353366613, + 0.0010346714407205582, + 0.0015313869807869196, + 0.001543155638501048, + 0.004259820096194744, + 0.002829545410349965, + 0.0017010780284181237, + 0.0030311758164316416, + 0.0020484495908021927, + 0.001539561664685607, + 0.002165252808481455, + 0.004240000154823065, + 0.004632370546460152, + 0.004377832170575857, + 0.0025218662340193987, + 0.0034181650262326, + 0.0032258813735097647, + 0.0028896157164126635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.04309103265404701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013404440833255649 + ], + "opposite_logprob_tokens": [ + -6.626340389251709 + ], + "logprob_reward": -0.0013404440833255649, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013404440833255649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12743277847766876, + 0.1286056786775589, + 0.1519543081521988, + 0.09044663608074188, + 0.11420348286628723, + 0.1459641456604004, + 0.07300357520580292, + 0.08045052736997604, + 0.05894890055060387, + 0.02898990735411644 + ], + "credit_map": [ + 0.013318704441189766, + 0.021135497838258743, + 0.02218703180551529, + 0.010434032417833805, + 0.011156938038766384, + 0.03382696211338043, + 0.005460401996970177, + 0.009913219138979912, + 0.018721483647823334, + 0.008502117358148098, + 0.051058996468782425, + 0.008232553489506245, + 0.002994528040289879, + 0.023907650262117386, + 0.003949453588575125, + 0.01123889908194542, + 0.015915684401988983, + 0.019991686567664146, + 0.04019035026431084, + 0.016459215432405472, + 0.006605489645153284, + 0.02884511463344097, + 0.008499750867486, + 0.01544701587408781, + 0.011309782974421978, + 0.010622872970998287, + 0.018502093851566315, + 0.014720231294631958, + 0.008183257654309273, + 0.010152584873139858, + 0.00823719147592783, + 0.008718617260456085, + 0.012197460979223251, + 0.008467421866953373, + 0.03879464045166969, + 0.010443729348480701, + 0.004609243478626013, + 0.02242918126285076, + 0.008369578048586845, + 0.008892233483493328, + 0.012207637540996075, + 0.01944749802350998, + 0.03650568798184395, + 0.028820017352700233, + 0.006326800212264061, + 0.021058237180113792, + 0.008108772337436676, + 0.013489490374922752, + 0.007431821431964636, + 0.015455292537808418, + 0.010772000066936016, + 0.013764924369752407, + 0.004971187096089125, + 0.009094742126762867, + 0.006407638546079397, + 0.005105969030410051, + 0.007250808645039797, + 0.007161919493228197, + 0.02031509205698967, + 0.012210722081363201, + 0.006117295939475298, + 0.013895622454583645, + 0.008133060298860073, + 0.005366005934774876, + 0.005007945466786623, + 0.006456519011408091, + 0.014777514152228832, + 0.010433935560286045, + 0.0033072985243052244, + 0.010007812641561031, + 0.004322193097323179, + 0.004635680466890335, + 0.0023607041221112013, + 0.004579917993396521, + 0.004416955169290304, + 0.004527723416686058, + 0.003189150243997574, + 0.003202341962605715, + 0.004253721795976162, + 0.002459392650052905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.0013404440833255649 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023311948403716087 + ], + "opposite_logprob_tokens": [ + -3.7733118534088135 + ], + "logprob_reward": -0.023311948403716087, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023311948403716087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09233852475881577, + 0.1746491938829422, + 0.3063898980617523, + 0.20322121679782867, + 0.036032285541296005, + 0.09570551663637161, + 0.03992703557014465, + 0.010080629959702492, + 0.02192891575396061, + 0.019726762548089027 + ], + "credit_map": [ + 0.008280781097710133, + 0.0159923043102026, + 0.01729494519531727, + 0.008157809264957905, + 0.007287084124982357, + 0.02467511221766472, + 0.0039054108783602715, + 0.006745081394910812, + 0.024148283526301384, + 0.013860778883099556, + 0.06680979579687119, + 0.018451737239956856, + 0.003140204818919301, + 0.0315629243850708, + 0.004975009709596634, + 0.011700457893311977, + 0.033068910241127014, + 0.03784620761871338, + 0.08541812002658844, + 0.03289058431982994, + 0.013599351979792118, + 0.054927583783864975, + 0.017575286328792572, + 0.031063860282301903, + 0.02927173115313053, + 0.021567776799201965, + 0.050035834312438965, + 0.029428234323859215, + 0.014043863862752914, + 0.02448795922100544, + 0.014695397578179836, + 0.01969042420387268, + 0.002922297455370426, + 0.004039237275719643, + 0.009005368687212467, + 0.004138594958931208, + 0.002525143790990114, + 0.007648848928511143, + 0.0025808566715568304, + 0.003171936608850956, + 0.007644983008503914, + 0.012039829976856709, + 0.025853730738162994, + 0.017734946683049202, + 0.004295862279832363, + 0.014908741228282452, + 0.005145115777850151, + 0.00808231346309185, + 0.0033616675063967705, + 0.007620947901159525, + 0.007674729451537132, + 0.007614762522280216, + 0.002999468008056283, + 0.004238943103700876, + 0.0035630010534077883, + 0.002853516722097993, + 0.0012894818792119622, + 0.0011396416230127215, + 0.0025881698820739985, + 0.0011292413109913468, + 0.0006835749372839928, + 0.0014487863518297672, + 0.0010603194823488593, + 0.0007414140272885561, + 0.0021309135481715202, + 0.0023974161595106125, + 0.005753856152296066, + 0.004550647456198931, + 0.0010950510622933507, + 0.0032193860970437527, + 0.0012825507437810302, + 0.0014990950003266335, + 0.0014395033940672874, + 0.002715781796723604, + 0.0033412950579077005, + 0.003462262451648712, + 0.0021917319390922785, + 0.0023409666027873755, + 0.002418055897578597, + 0.0018171642441302538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.023311948403716087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pure, endless black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, irregular horizontal crack visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a sharp, pure white beam of light burst from the crack and pierce upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in stark black and white without any color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward in a steady motion following the light beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328433156013489 + ], + "opposite_logprob_tokens": [ + -0.7578433156013489 + ], + "logprob_reward": -0.6328433156013489, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328433156013489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1546497792005539, + 0.08956556022167206, + 0.06669445335865021, + 0.047721076756715775, + 0.3550227880477905, + 0.0548245832324028, + 0.04107851907610893, + 0.04633845388889313, + 0.07767543196678162, + 0.06642938405275345 + ], + "credit_map": [ + 0.01533965952694416, + 0.021330634132027626, + 0.0331818051636219, + 0.020306391641497612, + 0.013507318682968616, + 0.028849931433796883, + 0.009431084617972374, + 0.012702951207756996, + 0.008335157297551632, + 0.009834582917392254, + 0.021039729937911034, + 0.014210482127964497, + 0.005032614339143038, + 0.014350302517414093, + 0.006879605818539858, + 0.00988308060914278, + 0.009184647351503372, + 0.008366676978766918, + 0.013133193366229534, + 0.008340480737388134, + 0.004483037628233433, + 0.008266568183898926, + 0.010144272819161415, + 0.004775573965162039, + 0.00569903664290905, + 0.0037446727510541677, + 0.01440515462309122, + 0.005482579581439495, + 0.002114867325872183, + 0.008584135212004185, + 0.003802857594564557, + 0.0038877748884260654, + 0.02871546521782875, + 0.04013839364051819, + 0.07828012108802795, + 0.049339376389980316, + 0.021040646359324455, + 0.07939170300960541, + 0.024963732808828354, + 0.033153340220451355, + 0.00548858055844903, + 0.005459324922412634, + 0.014890991151332855, + 0.00899598654359579, + 0.0031143762171268463, + 0.009101866744458675, + 0.004262966103851795, + 0.0035104926209896803, + 0.00470634363591671, + 0.0036226578522473574, + 0.013982532545924187, + 0.003126957220956683, + 0.0027128253132104874, + 0.007795287296175957, + 0.0017688011284917593, + 0.0033631178084760904, + 0.003336123889312148, + 0.005354308057576418, + 0.009706133976578712, + 0.008526432327926159, + 0.0029750100802630186, + 0.00844320748001337, + 0.0034133538138121367, + 0.004583884496241808, + 0.0088550616055727, + 0.009343666024506092, + 0.011481582187116146, + 0.020882170647382736, + 0.005138518754392862, + 0.009861732833087444, + 0.006475737784057856, + 0.005636959336698055, + 0.0054892427287995815, + 0.006870033219456673, + 0.018187347799539566, + 0.01042165420949459, + 0.004094331990927458, + 0.012930541299283504, + 0.0032965687569230795, + 0.005139664281159639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.6328433156013489 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_tensor([2102, 2102])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_tensor([2102, 2102])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0ac6f36537cc854f03505c4fa05f824e85ab1778 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_tensor([2102, 2102])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, static, eye-level medium shot captures a mesmerizing underwater scene, rendered in a surreal anime aesthetic reminiscent of Makoto Shinkai's style. Perfectly centered in the frame, two barracudas engage in a captivating courtship dance. Their sleek, elongated bodies, marked by distinct black vertical bars, glide through the crystal-clear water in a slow, synchronized ballet. As they circle each other, their movements are fluid and deliberate, their translucent fins barely disturbing their surroundings. Sunlight filters through the surface above, creating shifting, ethereal volumetric rays that penetrate the deep blue depths and illuminate the scene with a divine glow. This dramatic lighting, combined with high-contrast shadows, causes the fish's iridescent silver scales to shimmer with a magical intensity. Countless tiny, jagged bubbles rise from their gentle movements, catching the light like scattered diamonds. In the background, the flooded riverbank, depicted in soft-focus, painted detail and featuring moss-covered stones and delicate aquatic plants, creates a stark, beautiful contrast with the crisp rendering of the fish, creating an atmosphere of profound tranquility and enchantment.", + "seed": 9820, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_2102_seed_9820_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": -0.06708244178444148, + "vlm_logprob": -0.06708244178444148, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.06708244178444148, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2164473682641983 + ], + "opposite_logprob_tokens": [ + -1.716447353363037 + ], + "logprob_reward": -0.2164473682641983, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2164473682641983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1059252992272377, + 0.13429124653339386, + 0.12421482056379318, + 0.08298258483409882, + 0.04808683320879936, + 0.127157524228096, + 0.0896514281630516, + 0.2113281935453415, + 0.030513200908899307, + 0.045848820358514786 + ], + "credit_map": [ + 0.008208579383790493, + 0.0072381761856377125, + 0.01340728159993887, + 0.012383897788822651, + 0.010935095138847828, + 0.016189876943826675, + 0.02219167910516262, + 0.015370718203485012, + 0.01891379803419113, + 0.009871388785541058, + 0.013479604385793209, + 0.016318127512931824, + 0.010509975254535675, + 0.018349874764680862, + 0.028585392981767654, + 0.018263088539242744, + 0.017710397019982338, + 0.010080458596348763, + 0.01464766077697277, + 0.015630396082997322, + 0.008999153971672058, + 0.02981039695441723, + 0.012391463853418827, + 0.014944896101951599, + 0.009978954680263996, + 0.00796743668615818, + 0.008551537990570068, + 0.011484276503324509, + 0.005248574540019035, + 0.01976378634572029, + 0.009942153468728065, + 0.010045861825346947, + 0.0029855251777917147, + 0.0042477622628211975, + 0.005111122503876686, + 0.0071854619309306145, + 0.010360018350183964, + 0.0040072607807815075, + 0.007843461818993092, + 0.006346222013235092, + 0.021463926881551743, + 0.015161712653934956, + 0.017526449635624886, + 0.02049461007118225, + 0.009447568096220493, + 0.013362741097807884, + 0.014513798989355564, + 0.015186706557869911, + 0.0108760055154562, + 0.008665204979479313, + 0.008384563960134983, + 0.01080094464123249, + 0.009457237087190151, + 0.022651521489024162, + 0.008114463649690151, + 0.010701491497457027, + 0.0201212577521801, + 0.031060999259352684, + 0.026822641491889954, + 0.03358002379536629, + 0.02137533761560917, + 0.024406850337982178, + 0.02524235099554062, + 0.028718750923871994, + 0.002721691271290183, + 0.003128087380900979, + 0.0047885216772556305, + 0.005246123764663935, + 0.003203880274668336, + 0.0040271044708788395, + 0.003202726598829031, + 0.004195067100226879, + 0.004739938769489527, + 0.007004177197813988, + 0.006062461994588375, + 0.007940351031720638, + 0.003650113707408309, + 0.004137576557695866, + 0.005646811798214912, + 0.006667385809123516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.2164473682641983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496840178966522 + ], + "opposite_logprob_tokens": [ + -2.9299683570861816 + ], + "logprob_reward": -0.05496840178966522, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496840178966522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13959571719169617, + 0.0687498152256012, + 0.2370884269475937, + 0.12072386592626572, + 0.04349098354578018, + 0.05803573876619339, + 0.07538830488920212, + 0.1188938245177269, + 0.09299352020025253, + 0.04503978416323662 + ], + "credit_map": [ + 0.014711447060108185, + 0.01494316104799509, + 0.01330811157822609, + 0.01672254130244255, + 0.011970673687756062, + 0.015110119245946407, + 0.030311066657304764, + 0.022518616169691086, + 0.006405004765838385, + 0.00730634480714798, + 0.006871742662042379, + 0.010409840382635593, + 0.0064324201084673405, + 0.009695494547486305, + 0.012758883647620678, + 0.008870086632668972, + 0.0208743829280138, + 0.019180778414011, + 0.025612784549593925, + 0.038471005856990814, + 0.021486753597855568, + 0.03226511925458908, + 0.04658835753798485, + 0.0326092503964901, + 0.013998531736433506, + 0.00762080866843462, + 0.014974585734307766, + 0.016587644815444946, + 0.010830727405846119, + 0.02759397402405739, + 0.014388787560164928, + 0.01472879946231842, + 0.003411781508475542, + 0.002971872454509139, + 0.004885828588157892, + 0.005716732237488031, + 0.007701286114752293, + 0.004507636651396751, + 0.008924366906285286, + 0.005371477454900742, + 0.006940677762031555, + 0.004571550991386175, + 0.004809264559298754, + 0.0073456293903291225, + 0.010352603159844875, + 0.006143789738416672, + 0.010601199232041836, + 0.0072710211388766766, + 0.007675355300307274, + 0.00928755383938551, + 0.0065303402952849865, + 0.00867934338748455, + 0.010672745294868946, + 0.013682037591934204, + 0.009344262070953846, + 0.009516672231256962, + 0.009231439791619778, + 0.02227155677974224, + 0.014298447407782078, + 0.017413096502423286, + 0.011950370855629444, + 0.013363910838961601, + 0.015681222081184387, + 0.014683783054351807, + 0.008854061365127563, + 0.011586364358663559, + 0.012671188451349735, + 0.0183432474732399, + 0.007770521100610495, + 0.01157665066421032, + 0.010303252376616001, + 0.0118882330134511, + 0.003624730044975877, + 0.005177730694413185, + 0.005862304475158453, + 0.009758575819432735, + 0.003087154822424054, + 0.0055589922703802586, + 0.0058813923969864845, + 0.006088904105126858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.05496840178966522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632013574242592 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.008632013574242592, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632013574242592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19175684452056885, + 0.050568219274282455, + 0.24067482352256775, + 0.0961018055677414, + 0.0577363446354866, + 0.10728621482849121, + 0.04819988086819649, + 0.08607207238674164, + 0.07325576990842819, + 0.04834803566336632 + ], + "credit_map": [ + 0.019479895010590553, + 0.011296961456537247, + 0.01758280210196972, + 0.03117387555539608, + 0.0233962070196867, + 0.029084546491503716, + 0.03159155696630478, + 0.02815099246799946, + 0.004733681213110685, + 0.0045009502209723, + 0.006689806934446096, + 0.008268437348306179, + 0.004826118238270283, + 0.008565396070480347, + 0.007387399207800627, + 0.005596428643912077, + 0.046951718628406525, + 0.016998151317238808, + 0.023221280425786972, + 0.026740174740552902, + 0.017046187072992325, + 0.056738950312137604, + 0.028007542714476585, + 0.024970823898911476, + 0.008183000609278679, + 0.009969575330615044, + 0.00955415889620781, + 0.01589173637330532, + 0.006536155939102173, + 0.015565921552479267, + 0.017467379570007324, + 0.012933878228068352, + 0.004156364127993584, + 0.004387900698930025, + 0.00894613191485405, + 0.009175220504403114, + 0.006161184050142765, + 0.004974437411874533, + 0.011520086787641048, + 0.008415019139647484, + 0.013370425440371037, + 0.012051468715071678, + 0.012242194265127182, + 0.01627342775464058, + 0.011160850524902344, + 0.010637999512255192, + 0.018537228927016258, + 0.013012610375881195, + 0.0054521807469427586, + 0.005813259165734053, + 0.005352459382265806, + 0.005411254242062569, + 0.006519488524645567, + 0.007966776378452778, + 0.006016421131789684, + 0.00566803989931941, + 0.0069181607104837894, + 0.016842730343341827, + 0.011483591981232166, + 0.013607103377580643, + 0.008081771433353424, + 0.007293569855391979, + 0.010554791428148746, + 0.011290350928902626, + 0.004649383947253227, + 0.0084601491689682, + 0.00958336889743805, + 0.018381141126155853, + 0.005790875293314457, + 0.006187742576003075, + 0.008602375164628029, + 0.011600727215409279, + 0.004405595827847719, + 0.005373784806579351, + 0.005212339572608471, + 0.009141006506979465, + 0.0043087610974907875, + 0.005936774425208569, + 0.0065187569707632065, + 0.007451017387211323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.008632013574242592 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731334142386913 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731334142386913, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731334142386913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18097299337387085, + 0.05182080715894699, + 0.20940648019313812, + 0.13832898437976837, + 0.05274336040019989, + 0.09038040041923523, + 0.06438105553388596, + 0.08963758498430252, + 0.05350654944777489, + 0.06882171332836151 + ], + "credit_map": [ + 0.027183590456843376, + 0.010725434869527817, + 0.031649474054574966, + 0.022246405482292175, + 0.019583508372306824, + 0.025294248014688492, + 0.020837046205997467, + 0.023453300818800926, + 0.007039300166070461, + 0.00448534544557333, + 0.005494670942425728, + 0.006000563967972994, + 0.006056568119674921, + 0.011527180671691895, + 0.005626375786960125, + 0.00559080857783556, + 0.04100250080227852, + 0.016138212755322456, + 0.01885734312236309, + 0.019392114132642746, + 0.018244097009301186, + 0.054778508841991425, + 0.015614638105034828, + 0.025379087775945663, + 0.009350552223622799, + 0.00676721939817071, + 0.026453034952282906, + 0.023695247247815132, + 0.00905537884682417, + 0.01211358793079853, + 0.027305375784635544, + 0.02358861267566681, + 0.0033380379900336266, + 0.003482826054096222, + 0.011414957232773304, + 0.008317459374666214, + 0.0037890130188316107, + 0.004441522993147373, + 0.009134172461926937, + 0.008825378492474556, + 0.009864668361842632, + 0.009919512085616589, + 0.009215432219207287, + 0.011689716950058937, + 0.007630655076354742, + 0.017270606011152267, + 0.011461202055215836, + 0.01332861464470625, + 0.006044748704880476, + 0.00740321259945631, + 0.008040488697588444, + 0.00801808014512062, + 0.0068594468757510185, + 0.013948804698884487, + 0.006483839824795723, + 0.007582442834973335, + 0.008106022141873837, + 0.013804811984300613, + 0.015595782548189163, + 0.01359696127474308, + 0.008659356273710728, + 0.007938743568956852, + 0.009056995622813702, + 0.01287892460823059, + 0.004696746822446585, + 0.0066205961629748344, + 0.009047397412359715, + 0.006222956348210573, + 0.005682358518242836, + 0.006453998852521181, + 0.007898653857409954, + 0.006883847527205944, + 0.006022916175425053, + 0.011106692254543304, + 0.0073219905607402325, + 0.007600974757224321, + 0.007050485350191593, + 0.011485944502055645, + 0.01029402855783701, + 0.00793868862092495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.006731334142386913 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048633091151714325 + ], + "opposite_logprob_tokens": [ + -3.048633098602295 + ], + "logprob_reward": -0.048633091151714325, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048633091151714325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.201079860329628, + 0.048485469073057175, + 0.3431466221809387, + 0.09996218234300613, + 0.03481040894985199, + 0.08645618706941605, + 0.023262085393071175, + 0.03869395703077316, + 0.04342850297689438, + 0.08067473769187927 + ], + "credit_map": [ + 0.03173572197556496, + 0.011229225434362888, + 0.036811463534832, + 0.017234567552804947, + 0.024612946435809135, + 0.035876061767339706, + 0.018418295308947563, + 0.025161579251289368, + 0.006216150242835283, + 0.004790397360920906, + 0.007019505370408297, + 0.005269724875688553, + 0.005683739669620991, + 0.0072044264525175095, + 0.005822054110467434, + 0.006479470059275627, + 0.06960263848304749, + 0.02655182220041752, + 0.02956739254295826, + 0.03129260614514351, + 0.027894843369722366, + 0.09236326068639755, + 0.025283951312303543, + 0.04059011861681938, + 0.013210496865212917, + 0.006881197448819876, + 0.012476976029574871, + 0.012254404835402966, + 0.009419740177690983, + 0.02461749128997326, + 0.009438948705792427, + 0.011662928387522697, + 0.0024777702055871487, + 0.002256352687254548, + 0.007912179455161095, + 0.0064450413919985294, + 0.0027272712904959917, + 0.002352536655962467, + 0.005137924570590258, + 0.005501335021108389, + 0.009058111347258091, + 0.01067379117012024, + 0.00981998536735773, + 0.013362573459744453, + 0.007336871698498726, + 0.014611265622079372, + 0.010507293976843357, + 0.01108629535883665, + 0.0019713875371962786, + 0.0027912745717912912, + 0.0034471142571419477, + 0.0029048635624349117, + 0.0024078511632978916, + 0.00317732454277575, + 0.00376488477922976, + 0.0027973840478807688, + 0.0036392027977854013, + 0.0050467271357774734, + 0.00701588299125433, + 0.006333190482109785, + 0.003733839374035597, + 0.003210935043171048, + 0.00396840600296855, + 0.005745771806687117, + 0.0033606323413550854, + 0.004935206845402718, + 0.0061613465659320354, + 0.005045933183282614, + 0.005088545382022858, + 0.005837882868945599, + 0.007700106594711542, + 0.0052988482639193535, + 0.0053700548596680164, + 0.009734245017170906, + 0.012003200128674507, + 0.016118250787258148, + 0.006488096434623003, + 0.010070602409541607, + 0.009158059023320675, + 0.01173222903162241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.048633091151714325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, eye-level medium shot captures a mesmerizing underwater scene, rendered in a surreal anime aesthetic reminiscent of Makoto Shinkai's style. Perfectly centered in the frame, two barracudas engage in a captivating courtship dance. Their sleek, elongated bodies, marked by distinct black vertical bars, glide through the crystal-clear water in a slow, synchronized ballet. As they circle each other, their movements are fluid and deliberate, their translucent fins barely disturbing their surroundings. Sunlight filters through the surface above, creating shifting, ethereal volumetric rays that penetrate the deep blue depths and illuminate the scene with a divine glow. This dramatic lighting, combined with high-contrast shadows, causes the fish's iridescent silver scales to shimmer with a magical intensity. Countless tiny, jagged bubbles rise from their gentle movements, catching the light like scattered diamonds. In the background, the flooded riverbank, depicted in soft-focus, painted detail and featuring moss-covered stones and delicate aquatic plants, creates a stark, beautiful contrast with the crisp rendering of the fish, creating an atmosphere of profound tranquility and enchantment.", + "seed": 9821, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_2102_seed_9821_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": -0.08391026714816689, + "vlm_logprob": -0.08391026714816689, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.08391026714816689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256784439086914 + ], + "opposite_logprob_tokens": [ + -1.7256784439086914 + ], + "logprob_reward": -0.2256784439086914, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2256784439086914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5497927069664001, + 0.07162311673164368, + 0.028738003224134445, + 0.059862181544303894, + 0.030841710045933723, + 0.020480753853917122, + 0.05204974487423897, + 0.04757627099752426, + 0.043179746717214584, + 0.09585584700107574 + ], + "credit_map": [ + 0.1525149643421173, + 0.028366824612021446, + 0.0297661405056715, + 0.023872746154665947, + 0.026346290484070778, + 0.2144177109003067, + 0.040400389581918716, + 0.034107599407434464, + 0.009504050016403198, + 0.005739894695580006, + 0.010034236125648022, + 0.008032172918319702, + 0.00726971635594964, + 0.014590428210794926, + 0.0075315795838832855, + 0.00892104022204876, + 0.0034900184255093336, + 0.002999494317919016, + 0.003321866039186716, + 0.002793186344206333, + 0.0045812735334038734, + 0.004834237974137068, + 0.0030209128744900227, + 0.003697013482451439, + 0.006822042632848024, + 0.005871936678886414, + 0.008419403806328773, + 0.007234571967273951, + 0.007197521161288023, + 0.0070257000625133514, + 0.007895732298493385, + 0.009395275264978409, + 0.0038292561657726765, + 0.0025216408539563417, + 0.005308542400598526, + 0.0029010993894189596, + 0.004673724062740803, + 0.004265538416802883, + 0.0040114568546414375, + 0.0033304537646472454, + 0.002629463328048587, + 0.001967057818546891, + 0.0033558784052729607, + 0.0021608618553727865, + 0.002743666060268879, + 0.0019558651838451624, + 0.0024058662820607424, + 0.0032620958518236876, + 0.00632596667855978, + 0.006534547079354525, + 0.007481548935174942, + 0.004737295676022768, + 0.007746041752398014, + 0.005316170863807201, + 0.006216691341251135, + 0.007691481616348028, + 0.0051038675010204315, + 0.0044776201248168945, + 0.0037605364341288805, + 0.005614021793007851, + 0.009497077204287052, + 0.004964487627148628, + 0.007620092947036028, + 0.0065385643392801285, + 0.004652455449104309, + 0.003983843140304089, + 0.007319319527596235, + 0.004837573040276766, + 0.004695667885243893, + 0.005612206179648638, + 0.0051595838740468025, + 0.006919099949300289, + 0.008474191650748253, + 0.008269811049103737, + 0.0067182378843426704, + 0.005753114819526672, + 0.012144854292273521, + 0.010696901939809322, + 0.00989529862999916, + 0.03390343487262726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.2256784439086914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900883257389069 + ], + "opposite_logprob_tokens": [ + -2.5790088176727295 + ], + "logprob_reward": -0.07900883257389069, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900883257389069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2638407349586487, + 0.18342040479183197, + 0.07564433664083481, + 0.09259422868490219, + 0.037093598395586014, + 0.03578360378742218, + 0.0384460985660553, + 0.04062884300947189, + 0.17563559114933014, + 0.056912507861852646 + ], + "credit_map": [ + 0.07389145344495773, + 0.014108625240623951, + 0.014692811295390129, + 0.014195730909705162, + 0.014397887513041496, + 0.09766801446676254, + 0.019404521211981773, + 0.015481713227927685, + 0.018887484446167946, + 0.025196464732289314, + 0.016652565449476242, + 0.021740058436989784, + 0.042118269950151443, + 0.019973525777459145, + 0.02376721240580082, + 0.015084831975400448, + 0.007375065702944994, + 0.011565152555704117, + 0.011235409416258335, + 0.007152629550546408, + 0.013431514613330364, + 0.010159683413803577, + 0.007537969388067722, + 0.007186913397163153, + 0.009590324014425278, + 0.014790529385209084, + 0.011169139295816422, + 0.01266764011234045, + 0.011954830959439278, + 0.010258018039166927, + 0.011940576136112213, + 0.01022318471223116, + 0.005011351779103279, + 0.004195474088191986, + 0.005894662346690893, + 0.004267735406756401, + 0.0038204696029424667, + 0.003950971644371748, + 0.005248473957180977, + 0.004704461432993412, + 0.004585404880344868, + 0.003361831186339259, + 0.006992187816649675, + 0.0042062923312187195, + 0.0027775305788964033, + 0.0034494290594011545, + 0.004680649843066931, + 0.005730281118303537, + 0.004497320391237736, + 0.005180685315281153, + 0.004655048716813326, + 0.0035875062458217144, + 0.0095053780823946, + 0.003980899229645729, + 0.0033854309003800154, + 0.0036538359709084034, + 0.003681594505906105, + 0.004472062457352877, + 0.0038771224208176136, + 0.007069408427923918, + 0.0059656137600541115, + 0.0039962222799658775, + 0.004820764530450106, + 0.006746057886630297, + 0.02174399420619011, + 0.028941916301846504, + 0.01650441437959671, + 0.023400407284498215, + 0.027195220813155174, + 0.019210748374462128, + 0.02037867158651352, + 0.01826023869216442, + 0.0064653814770281315, + 0.008328575640916824, + 0.006800077855587006, + 0.006277251057326794, + 0.009670259431004524, + 0.006692543625831604, + 0.006613672710955143, + 0.006064754910767078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.07900883257389069 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594549672678113 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.00594549672678113, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594549672678113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2139545977115631, + 0.14020241796970367, + 0.0393136665225029, + 0.06438904255628586, + 0.08272568881511688, + 0.050988223403692245, + 0.05237716808915138, + 0.05811772495508194, + 0.23942597210407257, + 0.05850541964173317 + ], + "credit_map": [ + 0.051097579300403595, + 0.011552507989108562, + 0.02053138241171837, + 0.013570073060691357, + 0.012350491248071194, + 0.07011374831199646, + 0.017033886164426804, + 0.017704959958791733, + 0.015856150537729263, + 0.025064527988433838, + 0.010741323232650757, + 0.018473079428076744, + 0.024202799424529076, + 0.01570509932935238, + 0.017492180690169334, + 0.012667277827858925, + 0.006364046595990658, + 0.004869579803198576, + 0.004475585650652647, + 0.0045007336884737015, + 0.0048111192882061005, + 0.004675714299082756, + 0.003826489206403494, + 0.005790405906736851, + 0.0074732038192451, + 0.00583875784650445, + 0.012422051280736923, + 0.006494807079434395, + 0.008911538869142532, + 0.007712969556450844, + 0.006915646605193615, + 0.00862007774412632, + 0.007290063425898552, + 0.011735184118151665, + 0.011671663261950016, + 0.007474370300769806, + 0.013351697474718094, + 0.011391072534024715, + 0.009769685566425323, + 0.010041964240372181, + 0.006448112428188324, + 0.005264010280370712, + 0.009429709054529667, + 0.005608798004686832, + 0.005445848684757948, + 0.005312744993716478, + 0.00658411392942071, + 0.006894891615957022, + 0.006702245678752661, + 0.005371037870645523, + 0.009221239946782589, + 0.006566379684954882, + 0.006966634653508663, + 0.0046113464049994946, + 0.005229840520769358, + 0.007708447519689798, + 0.00608954718336463, + 0.006057747174054384, + 0.006153164431452751, + 0.00915859267115593, + 0.0082499198615551, + 0.005802155937999487, + 0.006396346725523472, + 0.010210261680185795, + 0.02795722335577011, + 0.035020653158426285, + 0.02570454776287079, + 0.029356159269809723, + 0.035877469927072525, + 0.026835547760128975, + 0.03193513676524162, + 0.026739267632365227, + 0.007408874109387398, + 0.008029423654079437, + 0.008297673426568508, + 0.0064859152771532536, + 0.00929249543696642, + 0.0056620691902935505, + 0.005779800936579704, + 0.007549178320914507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00594549672678113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633904159069061 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008633904159069061, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633904159069061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4086329936981201, + 0.10877522081136703, + 0.042434293776750565, + 0.05465128645300865, + 0.06404426693916321, + 0.06042821705341339, + 0.07710035890340805, + 0.029376760125160217, + 0.08303944766521454, + 0.07151719927787781 + ], + "credit_map": [ + 0.1070890724658966, + 0.02150721289217472, + 0.034497492015361786, + 0.01681762933731079, + 0.01937980204820633, + 0.15117235481739044, + 0.03196001425385475, + 0.026209399104118347, + 0.015048539265990257, + 0.011073008179664612, + 0.02108156681060791, + 0.013921712525188923, + 0.010250566527247429, + 0.012232095003128052, + 0.01168388407677412, + 0.013483839109539986, + 0.00728655094280839, + 0.0030796125065535307, + 0.006164140067994595, + 0.0044521368108689785, + 0.003874236950650811, + 0.006058997940272093, + 0.0037598884664475918, + 0.00775872590020299, + 0.007642053067684174, + 0.00409702816978097, + 0.010770957916975021, + 0.005410290323197842, + 0.003936353605240583, + 0.007716452237218618, + 0.005121053196489811, + 0.009957093745470047, + 0.007566007319837809, + 0.006501385476440191, + 0.012199444696307182, + 0.006240098271518946, + 0.008475151844322681, + 0.00914804171770811, + 0.00739159295335412, + 0.006522540934383869, + 0.005591181106865406, + 0.006380177568644285, + 0.012380294501781464, + 0.006996936164796352, + 0.005667702294886112, + 0.005561889614909887, + 0.007778041064739227, + 0.010071986354887486, + 0.01031864807009697, + 0.006535550579428673, + 0.013330938294529915, + 0.011304980143904686, + 0.008692609146237373, + 0.007997527718544006, + 0.007592145353555679, + 0.011327951215207577, + 0.003791100112721324, + 0.002491199877113104, + 0.0034176071640104055, + 0.005259239114820957, + 0.004134193062782288, + 0.0034234130289405584, + 0.0029955056961625814, + 0.00386450020596385, + 0.009933307766914368, + 0.010750934481620789, + 0.010519004426896572, + 0.008374763652682304, + 0.009416889399290085, + 0.010067417286336422, + 0.00873846746981144, + 0.015238653868436813, + 0.009585798718035221, + 0.008790690451860428, + 0.011621876619756222, + 0.007413382176309824, + 0.010394004173576832, + 0.006939635146409273, + 0.007479963358491659, + 0.009291844442486763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.008633904159069061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028465837240219 + ], + "opposite_logprob_tokens": [ + -2.3502845764160156 + ], + "logprob_reward": -0.10028465837240219, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028465837240219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4602296054363251, + 0.08201539516448975, + 0.03723907843232155, + 0.057204876095056534, + 0.07399153709411621, + 0.060625504702329636, + 0.04739736020565033, + 0.03327280655503273, + 0.09130318462848663, + 0.056720633059740067 + ], + "credit_map": [ + 0.13135255873203278, + 0.02249610237777233, + 0.02671317383646965, + 0.01846286654472351, + 0.023488547652959824, + 0.17774607241153717, + 0.03394247591495514, + 0.026027817279100418, + 0.01185462437570095, + 0.0071427482180297375, + 0.016765817999839783, + 0.009251076728105545, + 0.00687009934335947, + 0.010247515514492989, + 0.00907190702855587, + 0.010811610147356987, + 0.006453246343880892, + 0.003725921967998147, + 0.004791984800249338, + 0.003703515511006117, + 0.0037933653220534325, + 0.004596556071192026, + 0.0038576587103307247, + 0.006316829938441515, + 0.007915722206234932, + 0.004196698311716318, + 0.010485620237886906, + 0.005403005052357912, + 0.0041914391331374645, + 0.008649195544421673, + 0.005569710861891508, + 0.010793485678732395, + 0.009646858088672161, + 0.006589140277355909, + 0.014600341208279133, + 0.007884521037340164, + 0.009625856764614582, + 0.010518604889512062, + 0.00861759390681982, + 0.006508626509457827, + 0.008161941543221474, + 0.0047205057926476, + 0.01286772545427084, + 0.006078062113374472, + 0.005748742260038853, + 0.006102472078055143, + 0.007303021382540464, + 0.00964303407818079, + 0.0059707025066018105, + 0.00394822284579277, + 0.008437986485660076, + 0.006994676310569048, + 0.005186451133340597, + 0.004754848312586546, + 0.004858672618865967, + 0.007245799992233515, + 0.0037258861120790243, + 0.0027593793347477913, + 0.003859624732285738, + 0.005181269720196724, + 0.004205708857625723, + 0.003982703201472759, + 0.0034521284978836775, + 0.0061061023734509945, + 0.010804549790918827, + 0.01035116333514452, + 0.011426059529185295, + 0.00966630782932043, + 0.008766856044530869, + 0.011002302169799805, + 0.008829766884446144, + 0.020456179976463318, + 0.009306146763265133, + 0.00573864858597517, + 0.009299945086240768, + 0.004739710595458746, + 0.0074422042816877365, + 0.005886972416192293, + 0.006030834745615721, + 0.008276169188320637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.10028465837240219 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, eye-level medium shot captures a mesmerizing underwater scene, rendered in a surreal anime aesthetic reminiscent of Makoto Shinkai's style. Perfectly centered in the frame, two barracudas engage in a captivating courtship dance. Their sleek, elongated bodies, marked by distinct black vertical bars, glide through the crystal-clear water in a slow, synchronized ballet. As they circle each other, their movements are fluid and deliberate, their translucent fins barely disturbing their surroundings. Sunlight filters through the surface above, creating shifting, ethereal volumetric rays that penetrate the deep blue depths and illuminate the scene with a divine glow. This dramatic lighting, combined with high-contrast shadows, causes the fish's iridescent silver scales to shimmer with a magical intensity. Countless tiny, jagged bubbles rise from their gentle movements, catching the light like scattered diamonds. In the background, the flooded riverbank, depicted in soft-focus, painted detail and featuring moss-covered stones and delicate aquatic plants, creates a stark, beautiful contrast with the crisp rendering of the fish, creating an atmosphere of profound tranquility and enchantment.", + "seed": 9822, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_2102_seed_9822_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": -0.14158013444393874, + "vlm_logprob": -0.14158013444393874, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.14158013444393874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.49885034561157227 + ], + "opposite_logprob_tokens": [ + -0.9988503456115723 + ], + "logprob_reward": -0.49885034561157227, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.49885034561157227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1110268086194992, + 0.06569548696279526, + 0.1039341613650322, + 0.29410964250564575, + 0.11332438141107559, + 0.13715922832489014, + 0.08143410086631775, + 0.023526020348072052, + 0.05004134029150009, + 0.019748814404010773 + ], + "credit_map": [ + 0.00900178961455822, + 0.012068497017025948, + 0.015966633334755898, + 0.010641534812748432, + 0.018675554543733597, + 0.021661512553691864, + 0.011141405440866947, + 0.011869890615344048, + 0.0076694488525390625, + 0.004961130674928427, + 0.017080791294574738, + 0.009652003645896912, + 0.005599114112555981, + 0.009330928325653076, + 0.004583974834531546, + 0.006818097550421953, + 0.02190527878701687, + 0.023538900539278984, + 0.007114382926374674, + 0.006467984523624182, + 0.013037707656621933, + 0.014520568773150444, + 0.007266622502356768, + 0.01008271612226963, + 0.0933818370103836, + 0.06800246983766556, + 0.024251151829957962, + 0.02807433158159256, + 0.022951871156692505, + 0.029727470129728317, + 0.014241787604987621, + 0.013478707522153854, + 0.02862686477601528, + 0.03253825381398201, + 0.009936178103089333, + 0.010952589102089405, + 0.008121181279420853, + 0.010766943916678429, + 0.006038022693246603, + 0.006344350520521402, + 0.01658083125948906, + 0.019516998901963234, + 0.01413196511566639, + 0.025748759508132935, + 0.018852150067687035, + 0.01469543855637312, + 0.018564732745289803, + 0.00906835962086916, + 0.007167287636548281, + 0.0095747672021389, + 0.012324327602982521, + 0.007958762347698212, + 0.022864455357193947, + 0.008673837408423424, + 0.006798894610255957, + 0.006071772426366806, + 0.0027583669871091843, + 0.0022734126541763544, + 0.0026559289544820786, + 0.0022943930234760046, + 0.004284881521016359, + 0.002682188292965293, + 0.0034706145524978638, + 0.003106233198195696, + 0.005540644284337759, + 0.005572678055614233, + 0.008068829774856567, + 0.006388455163687468, + 0.007556891534477472, + 0.006047018337994814, + 0.0057194228284060955, + 0.005147404037415981, + 0.0017257212894037366, + 0.0028361149597913027, + 0.002802629256621003, + 0.0025186699349433184, + 0.0029043168760836124, + 0.0020789350382983685, + 0.002654130570590496, + 0.002228295663371682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.49885034561157227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905669301748276 + ], + "opposite_logprob_tokens": [ + -2.4640567302703857 + ], + "logprob_reward": -0.08905669301748276, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905669301748276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056484345346689224, + 0.01934572495520115, + 0.031154321506619453, + 0.5847243070602417, + 0.08935896307229996, + 0.05321173369884491, + 0.04846438765525818, + 0.0640857145190239, + 0.027391288429498672, + 0.02577929198741913 + ], + "credit_map": [ + 0.005139773245900869, + 0.007154922001063824, + 0.006889499258249998, + 0.007996916770935059, + 0.01090925745666027, + 0.009350203908979893, + 0.004002064000815153, + 0.005041706375777721, + 0.002554455539211631, + 0.001234276220202446, + 0.005559843964874744, + 0.0024277835618704557, + 0.001878012903034687, + 0.0023855511099100113, + 0.0014578498667106032, + 0.001847951440140605, + 0.0031959819607436657, + 0.003657514462247491, + 0.0033690077252686024, + 0.00264797848649323, + 0.003545640502125025, + 0.005491022951900959, + 0.002567243529483676, + 0.006679928861558437, + 0.1821180135011673, + 0.13229753077030182, + 0.048450883477926254, + 0.05781279131770134, + 0.04422280192375183, + 0.0650082603096962, + 0.029552293941378593, + 0.025261668488383293, + 0.006326702889055014, + 0.010353326797485352, + 0.02616775780916214, + 0.009937789291143417, + 0.008078020066022873, + 0.018008848652243614, + 0.005608934909105301, + 0.004877577535808086, + 0.0053507424890995026, + 0.0067703453823924065, + 0.005523772910237312, + 0.00830778107047081, + 0.007867688313126564, + 0.006520335096865892, + 0.007311358582228422, + 0.005559703800827265, + 0.003920459188520908, + 0.005396883003413677, + 0.005417525302618742, + 0.0070265852846205235, + 0.006900053471326828, + 0.0050043570809066296, + 0.009778227657079697, + 0.0050202906131744385, + 0.0048879156820476055, + 0.008497867733240128, + 0.006090749986469746, + 0.006963612977415323, + 0.00902994628995657, + 0.00943851936608553, + 0.010889166966080666, + 0.00828793365508318, + 0.002314518904313445, + 0.003243215149268508, + 0.0032769590616226196, + 0.0030036098323762417, + 0.0038156660739332438, + 0.004163977690041065, + 0.004390092566609383, + 0.0031832470558583736, + 0.002245493931695819, + 0.003341848263517022, + 0.0028251034673303366, + 0.0037523936480283737, + 0.0030861133709549904, + 0.00351189193315804, + 0.004148399457335472, + 0.002868045587092638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.08905669301748276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070497334003448 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070497334003448, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070497334003448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08628522604703903, + 0.017779437825083733, + 0.04156096652150154, + 0.4703531563282013, + 0.21428723633289337, + 0.05512511730194092, + 0.0248341616243124, + 0.03815344721078873, + 0.02041536197066307, + 0.03120584599673748 + ], + "credit_map": [ + 0.006417096592485905, + 0.009801485575735569, + 0.011122324503958225, + 0.013341660611331463, + 0.011625176295638084, + 0.012858123518526554, + 0.011811838485300541, + 0.009307516738772392, + 0.0025190021842718124, + 0.0016480576014146209, + 0.0026431505102664232, + 0.002124609425663948, + 0.002099035307765007, + 0.003188755828887224, + 0.001450387411750853, + 0.002106440719217062, + 0.005744417663663626, + 0.0034669211599975824, + 0.004218158777803183, + 0.0031490065157413483, + 0.00620076060295105, + 0.006638350430876017, + 0.0031214526388794184, + 0.009021902456879616, + 0.14791126549243927, + 0.10695125162601471, + 0.03739382326602936, + 0.0476369708776474, + 0.036441802978515625, + 0.04730655625462532, + 0.024723149836063385, + 0.02198837324976921, + 0.06500580161809921, + 0.08084207028150558, + 0.01624886691570282, + 0.009822888299822807, + 0.01637008786201477, + 0.014048351906239986, + 0.005868861451745033, + 0.006080310326069593, + 0.007449866738170385, + 0.007687740959227085, + 0.005765992682427168, + 0.006673287600278854, + 0.007458479143679142, + 0.007650683633983135, + 0.007731570862233639, + 0.004707492887973785, + 0.00182621693238616, + 0.0029976055957376957, + 0.0029518043156713247, + 0.003413600381463766, + 0.005635207053273916, + 0.003293716348707676, + 0.0028876790311187506, + 0.0018283342942595482, + 0.0035549176391214132, + 0.004690737463533878, + 0.0033235095907002687, + 0.0040266928263008595, + 0.006186716258525848, + 0.004952567163854837, + 0.006270939018577337, + 0.005147365387529135, + 0.0022531510330736637, + 0.002565252361819148, + 0.0025692572817206383, + 0.0022075839806348085, + 0.00440151197835803, + 0.0027318010106682777, + 0.0017531818011775613, + 0.0019336218247190118, + 0.002107898471876979, + 0.004021970555186272, + 0.004022645764052868, + 0.004675721749663353, + 0.004536351189017296, + 0.004196983762085438, + 0.004680550657212734, + 0.0029637254774570465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.011070497334003448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635676465928555 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635676465928555, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635676465928555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14672628045082092, + 0.06710650771856308, + 0.09750130027532578, + 0.16747452318668365, + 0.2574833333492279, + 0.050966840237379074, + 0.09479264914989471, + 0.03928244486451149, + 0.030797285959124565, + 0.047868795692920685 + ], + "credit_map": [ + 0.009476224891841412, + 0.014280061237514019, + 0.014116697944700718, + 0.01657738722860813, + 0.023908911272883415, + 0.031606003642082214, + 0.01869700290262699, + 0.018063990399241447, + 0.010174688883125782, + 0.009493965655565262, + 0.007856927812099457, + 0.0072293756529688835, + 0.007032534573227167, + 0.009063411504030228, + 0.008222230710089207, + 0.008033373393118382, + 0.013129420578479767, + 0.008873346261680126, + 0.008424020372331142, + 0.008798377588391304, + 0.01543176919221878, + 0.015676070004701614, + 0.008282492868602276, + 0.01888580247759819, + 0.05887434259057045, + 0.04226813092827797, + 0.013805825263261795, + 0.009445923380553722, + 0.013088555075228214, + 0.016181692481040955, + 0.0063058738596737385, + 0.007504183333367109, + 0.07624554634094238, + 0.09147890657186508, + 0.017271660268306732, + 0.01466679759323597, + 0.024070415645837784, + 0.015781350433826447, + 0.009855874814093113, + 0.008112790994346142, + 0.006677637342363596, + 0.006543489173054695, + 0.004711860325187445, + 0.008083060383796692, + 0.0062788426876068115, + 0.007684813812375069, + 0.005377181340008974, + 0.005609951913356781, + 0.007749008014798164, + 0.011511377058923244, + 0.010905027389526367, + 0.01310997735708952, + 0.020339811220765114, + 0.010437266901135445, + 0.012633602134883404, + 0.008106582798063755, + 0.003351096995174885, + 0.005694698076695204, + 0.003529828507453203, + 0.004242799244821072, + 0.005158834159374237, + 0.0047670151107013226, + 0.006505678407847881, + 0.006032492965459824, + 0.0033324137330055237, + 0.0039009323809295893, + 0.004203637596219778, + 0.003108975011855364, + 0.006961992010474205, + 0.003607348073273897, + 0.002618337282910943, + 0.0030636503361165524, + 0.0032196328975260258, + 0.00641561858355999, + 0.005844400729984045, + 0.0067670936696231365, + 0.007718794047832489, + 0.006137686315923929, + 0.007062741555273533, + 0.004702829755842686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.008635676465928555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028745979070663 + ], + "opposite_logprob_tokens": [ + -2.350287437438965 + ], + "logprob_reward": -0.10028745979070663, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028745979070663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13300250470638275, + 0.06070755049586296, + 0.08815062046051025, + 0.1173810362815857, + 0.3502994775772095, + 0.06301310658454895, + 0.06765739619731903, + 0.04160540550947189, + 0.02729758433997631, + 0.05088532716035843 + ], + "credit_map": [ + 0.00811794027686119, + 0.01439992431551218, + 0.009966352954506874, + 0.01071785669773817, + 0.026703927665948868, + 0.033478304743766785, + 0.013784239999949932, + 0.015833958983421326, + 0.013387488201260567, + 0.005982960108667612, + 0.005379533860832453, + 0.0061521511524915695, + 0.006662815809249878, + 0.01383144035935402, + 0.0035093941260129213, + 0.005801768973469734, + 0.012862581759691238, + 0.007638557814061642, + 0.0089199123904109, + 0.008040678687393665, + 0.011191981844604015, + 0.012077867984771729, + 0.00850443635135889, + 0.0189146026968956, + 0.033824920654296875, + 0.0240644421428442, + 0.009085280820727348, + 0.011049393564462662, + 0.01140767615288496, + 0.014722976833581924, + 0.004621877335011959, + 0.008604470640420914, + 0.10780202597379684, + 0.1316240429878235, + 0.025702018290758133, + 0.017028911039233208, + 0.026755748316645622, + 0.018024398013949394, + 0.01265672966837883, + 0.010705598630011082, + 0.006817770190536976, + 0.008270954713225365, + 0.007192283868789673, + 0.010107045993208885, + 0.007778393570333719, + 0.008474965579807758, + 0.007275465875864029, + 0.007096220273524523, + 0.005745699163526297, + 0.008336266502737999, + 0.008038506843149662, + 0.008290265686810017, + 0.016448672860860825, + 0.008107542060315609, + 0.006983096245676279, + 0.005707344971597195, + 0.0033808511216193438, + 0.005373081192374229, + 0.004021031316369772, + 0.004445315804332495, + 0.006071319337934256, + 0.0052605136297643185, + 0.007004311308264732, + 0.006048982497304678, + 0.003271752502769232, + 0.0033511833753436804, + 0.0031988765113055706, + 0.0027764884289354086, + 0.005270117428153753, + 0.0038419279735535383, + 0.0023394033778458834, + 0.0032478326465934515, + 0.0036391373723745346, + 0.006598575506359339, + 0.006308070849627256, + 0.008015529252588749, + 0.007424445357173681, + 0.006232663057744503, + 0.007556371856480837, + 0.00511053716763854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.10028745979070663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, eye-level medium shot captures a mesmerizing underwater scene, rendered in a surreal anime aesthetic reminiscent of Makoto Shinkai's style. Perfectly centered in the frame, two barracudas engage in a captivating courtship dance. Their sleek, elongated bodies, marked by distinct black vertical bars, glide through the crystal-clear water in a slow, synchronized ballet. As they circle each other, their movements are fluid and deliberate, their translucent fins barely disturbing their surroundings. Sunlight filters through the surface above, creating shifting, ethereal volumetric rays that penetrate the deep blue depths and illuminate the scene with a divine glow. This dramatic lighting, combined with high-contrast shadows, causes the fish's iridescent silver scales to shimmer with a magical intensity. Countless tiny, jagged bubbles rise from their gentle movements, catching the light like scattered diamonds. In the background, the flooded riverbank, depicted in soft-focus, painted detail and featuring moss-covered stones and delicate aquatic plants, creates a stark, beautiful contrast with the crisp rendering of the fish, creating an atmosphere of profound tranquility and enchantment.", + "seed": 9823, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_1/index_2102_seed_9823_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": -0.08891805768944323, + "vlm_logprob": -0.08891805768944323, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.08891805768944323, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2788017690181732 + ], + "opposite_logprob_tokens": [ + -1.5288017988204956 + ], + "logprob_reward": -0.2788017690181732, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2788017690181732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21023911237716675, + 0.09805922210216522, + 0.06739003956317902, + 0.1587456464767456, + 0.09055498242378235, + 0.11843829602003098, + 0.07158814370632172, + 0.08459862321615219, + 0.04511173814535141, + 0.055274177342653275 + ], + "credit_map": [ + 0.02665713243186474, + 0.02082929015159607, + 0.03565414994955063, + 0.02398751489818096, + 0.019527902826666832, + 0.03625357523560524, + 0.017910173162817955, + 0.02941938303411007, + 0.020493807271122932, + 0.019802480936050415, + 0.008058060891926289, + 0.008522004820406437, + 0.012211876921355724, + 0.011436804197728634, + 0.009104099124670029, + 0.008430089801549911, + 0.009865880012512207, + 0.012229207903146744, + 0.010086487047374249, + 0.008132373914122581, + 0.0067821284756064415, + 0.008030030876398087, + 0.0063046677969396114, + 0.005959264002740383, + 0.046755820512771606, + 0.04468926414847374, + 0.011949281208217144, + 0.009556529112160206, + 0.009422259405255318, + 0.013863328844308853, + 0.00887450110167265, + 0.013634675182402134, + 0.012173019349575043, + 0.012255748733878136, + 0.011595983058214188, + 0.008544466458261013, + 0.015928197652101517, + 0.013482414186000824, + 0.008924704976379871, + 0.007650451269000769, + 0.015806002542376518, + 0.011676766909658909, + 0.016479559242725372, + 0.014228450134396553, + 0.019803417846560478, + 0.016762612387537956, + 0.009445318952202797, + 0.014236166141927242, + 0.012430650182068348, + 0.005546051077544689, + 0.01106804609298706, + 0.010335128754377365, + 0.006612892262637615, + 0.014092864468693733, + 0.006249884143471718, + 0.005252629518508911, + 0.008520929142832756, + 0.006461836397647858, + 0.010002722963690758, + 0.015197432599961758, + 0.009052298963069916, + 0.00993554200977087, + 0.014505480416119099, + 0.010922381654381752, + 0.006139167118817568, + 0.0050635249353945255, + 0.004758904688060284, + 0.006773586850613356, + 0.005866673309355974, + 0.008846182376146317, + 0.004686306696385145, + 0.0029773919377475977, + 0.007312438916414976, + 0.0063743675127625465, + 0.005590302869677544, + 0.012894655577838421, + 0.005445762071758509, + 0.007869555614888668, + 0.004426421131938696, + 0.005360677372664213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.2788017690181732 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789993554353714 + ], + "opposite_logprob_tokens": [ + -2.5789992809295654 + ], + "logprob_reward": -0.0789993554353714, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789993554353714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11190449446439743, + 0.10942555963993073, + 0.13689033687114716, + 0.20098114013671875, + 0.11337725818157196, + 0.06771606206893921, + 0.09626360982656479, + 0.07743265479803085, + 0.01796910911798477, + 0.06803970783948898 + ], + "credit_map": [ + 0.01386086642742157, + 0.01141100749373436, + 0.015660634264349937, + 0.010579882189631462, + 0.01160405483096838, + 0.021575748920440674, + 0.009067932143807411, + 0.01814437471330166, + 0.01622992753982544, + 0.01557252462953329, + 0.021357214078307152, + 0.010939275845885277, + 0.011469997465610504, + 0.010535103268921375, + 0.012973731383681297, + 0.01034779753535986, + 0.030122855678200722, + 0.032779257744550705, + 0.01837032288312912, + 0.01103959046304226, + 0.013601385988295078, + 0.011486008763313293, + 0.01040696632117033, + 0.009083959273993969, + 0.039931245148181915, + 0.04370656982064247, + 0.01840159110724926, + 0.014270706102252007, + 0.01954818330705166, + 0.02052328735589981, + 0.02152676321566105, + 0.023072833195328712, + 0.013368172571063042, + 0.011436027474701405, + 0.015234447084367275, + 0.020581912249326706, + 0.0126514732837677, + 0.021943140774965286, + 0.008876699954271317, + 0.00928539503365755, + 0.00916395802050829, + 0.0051286728121340275, + 0.013270463794469833, + 0.007336210925132036, + 0.008198088966310024, + 0.010948044247925282, + 0.005258343182504177, + 0.008412289433181286, + 0.016505904495716095, + 0.006018065381795168, + 0.017619816586375237, + 0.01104854978621006, + 0.008988700807094574, + 0.02074670046567917, + 0.00809506606310606, + 0.0072408136911690235, + 0.007106628734618425, + 0.0055768610909581184, + 0.009550552815198898, + 0.010965430177748203, + 0.00971957016736269, + 0.012048298493027687, + 0.012904283590614796, + 0.009561034850776196, + 0.0023232861422002316, + 0.002249564277008176, + 0.0025177267380058765, + 0.0029844255186617374, + 0.0019894626457244158, + 0.0021563090849667788, + 0.002044598339125514, + 0.0017037370707839727, + 0.007718520704656839, + 0.007235347758978605, + 0.009479883126914501, + 0.012057567946612835, + 0.009034018032252789, + 0.008637120015919209, + 0.007243387866765261, + 0.00663386844098568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.0789993554353714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635659046471119 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635659046471119, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635659046471119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0940113440155983, + 0.11087390780448914, + 0.10546988248825073, + 0.14580321311950684, + 0.07929389923810959, + 0.041000835597515106, + 0.12144367396831512, + 0.19071617722511292, + 0.03042140044271946, + 0.08096565306186676 + ], + "credit_map": [ + 0.012806667014956474, + 0.010146441869437695, + 0.010013477876782417, + 0.009556392207741737, + 0.014057456515729427, + 0.014524280093610287, + 0.011415639892220497, + 0.011490988545119762, + 0.02816799283027649, + 0.027293799445033073, + 0.007768502924591303, + 0.008992450311779976, + 0.012014115229249, + 0.00973818264901638, + 0.006755745504051447, + 0.010143130086362362, + 0.011809092946350574, + 0.011705650016665459, + 0.011996849440038204, + 0.00801854021847248, + 0.01857423037290573, + 0.02211063727736473, + 0.010995621792972088, + 0.010259255766868591, + 0.01882869377732277, + 0.020954035222530365, + 0.015207944437861443, + 0.015009121038019657, + 0.0181143656373024, + 0.023304466158151627, + 0.016551168635487556, + 0.017833435907959938, + 0.00994674302637577, + 0.00901731476187706, + 0.008005760610103607, + 0.011966611258685589, + 0.014126532711088657, + 0.010535789653658867, + 0.007553341332823038, + 0.008141804486513138, + 0.004873061086982489, + 0.0039390441961586475, + 0.006813689600676298, + 0.006809769663959742, + 0.006708080880343914, + 0.004971490241587162, + 0.003554014256224036, + 0.003331685671582818, + 0.020623501390218735, + 0.00725276954472065, + 0.02102944254875183, + 0.012068112380802631, + 0.011377132497727871, + 0.02366398461163044, + 0.009892263449728489, + 0.015536469407379627, + 0.017837293446063995, + 0.017015032470226288, + 0.02028365433216095, + 0.03453920781612396, + 0.022257566452026367, + 0.02713586762547493, + 0.027462737634778023, + 0.02418481931090355, + 0.0029982756823301315, + 0.004896737635135651, + 0.0034049719106405973, + 0.003553597489371896, + 0.004769709426909685, + 0.0049541485495865345, + 0.0028998241759836674, + 0.002944135805591941, + 0.017561327666044235, + 0.007880831137299538, + 0.023679764941334724, + 0.009823067113757133, + 0.005858763586729765, + 0.009327284060418606, + 0.0041147819720208645, + 0.002719835378229618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.004635659046471119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191380063071847 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.003191380063071847, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191380063071847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18453140556812286, + 0.1302664428949356, + 0.11659703403711319, + 0.05688278377056122, + 0.15670202672481537, + 0.04363171383738518, + 0.0800829753279686, + 0.10478810966014862, + 0.030506547540426254, + 0.09601099044084549 + ], + "credit_map": [ + 0.022768963128328323, + 0.018725112080574036, + 0.023133987560868263, + 0.016685139387845993, + 0.021693244576454163, + 0.034097667783498764, + 0.01690050959587097, + 0.030526766553521156, + 0.035546693950891495, + 0.03271464258432388, + 0.017727330327033997, + 0.009805991314351559, + 0.010850108228623867, + 0.008216049522161484, + 0.007185098249465227, + 0.008220522664487362, + 0.022757306694984436, + 0.02429514192044735, + 0.017531923949718475, + 0.010507871396839619, + 0.012470396235585213, + 0.01166567113250494, + 0.008631990291178226, + 0.008736718446016312, + 0.007959057576954365, + 0.006954681593924761, + 0.005967631004750729, + 0.006566670723259449, + 0.007172896061092615, + 0.0051458231173455715, + 0.006941830273717642, + 0.010174188762903214, + 0.020933914929628372, + 0.028600791469216347, + 0.011100661009550095, + 0.012978760525584221, + 0.031139664351940155, + 0.017306102439761162, + 0.02102753333747387, + 0.013614591211080551, + 0.005204071290791035, + 0.004478273913264275, + 0.007191390730440617, + 0.006562473718076944, + 0.008058995939791203, + 0.004678010009229183, + 0.0037103856448084116, + 0.0037481116596609354, + 0.014822684228420258, + 0.0049206861294806, + 0.010097422637045383, + 0.00762387877330184, + 0.008045009337365627, + 0.019291624426841736, + 0.00665839621797204, + 0.008623268455266953, + 0.009039485827088356, + 0.009027237072587013, + 0.010567456483840942, + 0.020984644070267677, + 0.013200544752180576, + 0.01137844193726778, + 0.018990380689501762, + 0.011599916964769363, + 0.003262934507802129, + 0.0052148932591080666, + 0.0036594094708561897, + 0.003947809804230928, + 0.004774201661348343, + 0.004224314354360104, + 0.0028826871421188116, + 0.0025402973406016827, + 0.01674669049680233, + 0.008772174827754498, + 0.025943560525774956, + 0.011374633759260178, + 0.007294331677258015, + 0.016100913286209106, + 0.0053412895649671555, + 0.004437385126948357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.003191380063071847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two barracudas visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the barracudas have sleek, elongated bodies with distinct black vertical bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set underwater with sunlight filtering through the surface to create volumetric rays?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus flooded riverbank featuring moss-covered stones and aquatic plants?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896212488412857 + ], + "opposite_logprob_tokens": [ + -2.5789620876312256 + ], + "logprob_reward": -0.07896212488412857, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896212488412857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12337391823530197, + 0.1498156040906906, + 0.09820576012134552, + 0.11039318144321442, + 0.06592980027198792, + 0.07542426884174347, + 0.08705318719148636, + 0.12487754225730896, + 0.04721061885356903, + 0.11771618574857712 + ], + "credit_map": [ + 0.01697162538766861, + 0.01671469584107399, + 0.017809806391596794, + 0.01570126973092556, + 0.012096836231648922, + 0.018720054998993874, + 0.011067521758377552, + 0.01429210789501667, + 0.04042578116059303, + 0.03642880544066429, + 0.020543962717056274, + 0.01117207482457161, + 0.012472935020923615, + 0.011134383268654346, + 0.00746053596958518, + 0.010177122429013252, + 0.023643705993890762, + 0.02698308229446411, + 0.010038075037300587, + 0.007133182603865862, + 0.00948760099709034, + 0.007111650891602039, + 0.006758446339517832, + 0.007050026673823595, + 0.019015146419405937, + 0.018477926030755043, + 0.008488417603075504, + 0.013573992066085339, + 0.012469801120460033, + 0.015034683048725128, + 0.011813427321612835, + 0.011519786901772022, + 0.008314618840813637, + 0.01111975684762001, + 0.006966846529394388, + 0.007340976968407631, + 0.013049894943833351, + 0.007974415086209774, + 0.007276233751326799, + 0.0038870559073984623, + 0.0092058340087533, + 0.007295238319784403, + 0.0109022818505764, + 0.01028277724981308, + 0.011918327771127224, + 0.010301647707819939, + 0.006298375781625509, + 0.009219786152243614, + 0.012978368438780308, + 0.006355164106935263, + 0.012188373133540154, + 0.007602179888635874, + 0.01120301615446806, + 0.018577009439468384, + 0.008292068727314472, + 0.009857003577053547, + 0.011058193631470203, + 0.009327688254415989, + 0.013339805416762829, + 0.024720173329114914, + 0.014401105232536793, + 0.013229569420218468, + 0.023842543363571167, + 0.01495846826583147, + 0.00501787755638361, + 0.006696850061416626, + 0.004321071319282055, + 0.006396848242729902, + 0.005995732732117176, + 0.008131759241223335, + 0.006557808257639408, + 0.004092670511454344, + 0.017705369740724564, + 0.010398301295936108, + 0.029452966526150703, + 0.016478873789310455, + 0.010815013200044632, + 0.017436165362596512, + 0.008208641782402992, + 0.007220853120088577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.07896212488412857 + } + ], + "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/0000122/group_10/index_tensor([7601, 7601])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_tensor([7601, 7601])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4ca21873066eeca37dac14b807fa745d472c8aa7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_tensor([7601, 7601])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "perfectly static, extreme close-up, framing the scene with a distinctive retro-cyberpunk aesthetic, focuses on a cybernetic hand, its surface a perfect mirrored chrome sheen, broken by thin, dark seams. As persistent raindrops fall, they instantly hiss and evaporate into tiny vapors upon contact with the hand's superheated metal. The hand carefully grips a pair of intricate pruners, their delicate blades etched with microcircuits and emitting a steady, cool cyan glow. With mechanical precision, the pruners hover mere millimeters above a bioluminescent plant. The plant's petals are in a tragic state of decay; they slowly wilt, their edges curling inward, as the ethereal natural light fades, its color shifting to a dull, sickly copper hue. The entire scene is primarily illuminated by the moody, flickering purple-red glow of a large, out-of-focus holographic sign set against a distant, rain-slicked cityscape. Its light catches and reflects off the damp chrome surface, creating an atmosphere of gloom, concentration, and melancholy.", + "seed": 4027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_7601_seed_4027_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": -0.11973249251022935, + "vlm_logprob": -0.11973249251022935, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.11973249251022935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021398037672042847 + ], + "opposite_logprob_tokens": [ + -3.8963980674743652 + ], + "logprob_reward": -0.021398037672042847, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021398037672042847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38320988416671753, + 0.07336550951004028, + 0.04882744327187538, + 0.22664722800254822, + 0.05559201538562775, + 0.0805908590555191, + 0.06454090774059296, + 0.013157869689166546, + 0.019747190177440643, + 0.034321129322052 + ], + "credit_map": [ + 0.032670676708221436, + 0.04138708487153053, + 0.06252190470695496, + 0.053425971418619156, + 0.04773089662194252, + 0.06498090922832489, + 0.04936431348323822, + 0.03112811967730522, + 0.013749985955655575, + 0.008629905991256237, + 0.007176933344453573, + 0.016576597467064857, + 0.006145460996776819, + 0.006428190041333437, + 0.008526740595698357, + 0.006131697446107864, + 0.010992279276251793, + 0.008002509362995625, + 0.005859449505805969, + 0.006141036283224821, + 0.00542342197149992, + 0.005091147962957621, + 0.004066430032253265, + 0.0032511649187654257, + 0.026911746710538864, + 0.018491670489311218, + 0.08129503577947617, + 0.02738717570900917, + 0.012537208385765553, + 0.01597924344241619, + 0.017737185582518578, + 0.026307959109544754, + 0.007053472567349672, + 0.012214708141982555, + 0.012179727666079998, + 0.0030942193698138, + 0.004485017154365778, + 0.007334914989769459, + 0.0037815049290657043, + 0.005448451731353998, + 0.01184786856174469, + 0.010588063858449459, + 0.012402702122926712, + 0.005476528313010931, + 0.012431622482836246, + 0.016101684421300888, + 0.005999197717756033, + 0.0057431915774941444, + 0.009082188829779625, + 0.014199820347130299, + 0.007662233430892229, + 0.005918364506214857, + 0.008165747858583927, + 0.011541754007339478, + 0.0043748049065470695, + 0.0035959898959845304, + 0.0025103327352553606, + 0.002361362101510167, + 0.0014624845935031772, + 0.001383899711072445, + 0.001530165201984346, + 0.0017662789905443788, + 0.0010991223389282823, + 0.0010442244820296764, + 0.0038013088051229715, + 0.003189933719113469, + 0.0023300687316805124, + 0.00209063570946455, + 0.0018732282333076, + 0.0033405646681785583, + 0.0013329898938536644, + 0.0017884603003039956, + 0.006325362250208855, + 0.006930992007255554, + 0.004341959021985531, + 0.0033866181038320065, + 0.0037178804632276297, + 0.005161692388355732, + 0.0022117383778095245, + 0.002244886476546526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.021398037672042847 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163922011852264 + ], + "opposite_logprob_tokens": [ + -1.701639175415039 + ], + "logprob_reward": -0.20163922011852264, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163922011852264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27753356099128723, + 0.2793012857437134, + 0.13684485852718353, + 0.060622185468673706, + 0.06720567494630814, + 0.031223785132169724, + 0.05570971220731735, + 0.044155921787023544, + 0.02541576325893402, + 0.02198733575642109 + ], + "credit_map": [ + 0.04016037657856941, + 0.035749923437833786, + 0.06038595736026764, + 0.04136085882782936, + 0.02421470917761326, + 0.021563151851296425, + 0.03292124718427658, + 0.021177299320697784, + 0.02196984551846981, + 0.013626646250486374, + 0.017137039452791214, + 0.0691501572728157, + 0.013787461444735527, + 0.014871535822749138, + 0.0974629744887352, + 0.031295619904994965, + 0.015965616330504417, + 0.020671211183071136, + 0.01941605843603611, + 0.0196700319647789, + 0.01558594685047865, + 0.014820820651948452, + 0.018827632069587708, + 0.011887519620358944, + 0.00478476844727993, + 0.006776771508157253, + 0.013639098964631557, + 0.010086205787956715, + 0.0034833133686333895, + 0.006228919606655836, + 0.007518470287322998, + 0.008104628883302212, + 0.00686597591266036, + 0.010743260383605957, + 0.009267620742321014, + 0.006457280367612839, + 0.00874362327158451, + 0.013464090414345264, + 0.005963267292827368, + 0.005700547713786364, + 0.005309410393238068, + 0.006560664623975754, + 0.007824511267244816, + 0.0016875298460945487, + 0.0030518502462655306, + 0.003116011619567871, + 0.0016817109426483512, + 0.001992092002183199, + 0.008273041807115078, + 0.01042864192277193, + 0.006330334581434727, + 0.005082396790385246, + 0.007713516242802143, + 0.00909578800201416, + 0.004396295174956322, + 0.004389687906950712, + 0.007542031817138195, + 0.005796142388135195, + 0.004962489940226078, + 0.0047899531200528145, + 0.006346364505589008, + 0.006276116240769625, + 0.004067662637680769, + 0.004375155549496412, + 0.004966717679053545, + 0.0036636043805629015, + 0.003410044824704528, + 0.0031294701620936394, + 0.0024178274907171726, + 0.00373239372856915, + 0.0016517300391569734, + 0.0024439729750156403, + 0.003792630974203348, + 0.0026929599698632956, + 0.0026120732072740793, + 0.002043277258053422, + 0.0020915530622005463, + 0.004247160628437996, + 0.002066262299194932, + 0.002441415563225746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.20163922011852264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905069530010223 + ], + "opposite_logprob_tokens": [ + -2.464050769805908 + ], + "logprob_reward": -0.08905069530010223, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905069530010223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15530343353748322, + 0.15918122231960297, + 0.09655807167291641, + 0.13037851452827454, + 0.13167431950569153, + 0.06165675073862076, + 0.043069176375865936, + 0.06968330591917038, + 0.0796360895037651, + 0.07285913079977036 + ], + "credit_map": [ + 0.01462098490446806, + 0.02229989878833294, + 0.023423168808221817, + 0.02226852811872959, + 0.018033968284726143, + 0.019376646727323532, + 0.020995039492845535, + 0.014285195618867874, + 0.013056277297437191, + 0.01648947410285473, + 0.01370493695139885, + 0.04947348311543465, + 0.012500789016485214, + 0.01201668381690979, + 0.02861163578927517, + 0.013327937573194504, + 0.023590408265590668, + 0.015108860097825527, + 0.011130514554679394, + 0.010793033055961132, + 0.007827859371900558, + 0.009500810876488686, + 0.009840224869549274, + 0.00876635778695345, + 0.01720830425620079, + 0.01297011598944664, + 0.044693607836961746, + 0.015491420403122902, + 0.00676906481385231, + 0.011162418872117996, + 0.01090225763618946, + 0.01118132472038269, + 0.018955595791339874, + 0.026084445416927338, + 0.0196234080940485, + 0.014227161183953285, + 0.012034371495246887, + 0.01741543971002102, + 0.01154214609414339, + 0.011791755445301533, + 0.010328901000320911, + 0.008297000080347061, + 0.009810071438550949, + 0.004393625073134899, + 0.008899550884962082, + 0.010951383970677853, + 0.004433725494891405, + 0.0045424955897033215, + 0.005428253207355738, + 0.008203134872019291, + 0.005935818888247013, + 0.00437903730198741, + 0.005632906220853329, + 0.0074650258757174015, + 0.0035839960910379887, + 0.0024410064797848463, + 0.01084557082504034, + 0.009108361788094044, + 0.007447196636348963, + 0.008515908382833004, + 0.009269836358726025, + 0.011293631047010422, + 0.006813636049628258, + 0.00638916390016675, + 0.011135313659906387, + 0.013777181506156921, + 0.010297967121005058, + 0.00898426491767168, + 0.009572078473865986, + 0.013414551503956318, + 0.005806137342005968, + 0.006648591719567776, + 0.013164875097572803, + 0.012465712614357471, + 0.008289244025945663, + 0.00675212312489748, + 0.011070616543293, + 0.011832699179649353, + 0.004657357931137085, + 0.0046265022829174995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.08905069530010223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132522106170654 + ], + "opposite_logprob_tokens": [ + -1.4063252210617065 + ], + "logprob_reward": -0.28132522106170654, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132522106170654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26279449462890625, + 0.11926402151584625, + 0.07169615477323532, + 0.06892476975917816, + 0.06080086529254913, + 0.08600287139415741, + 0.10419568419456482, + 0.14058716595172882, + 0.035070691257715225, + 0.0506633035838604 + ], + "credit_map": [ + 0.0321180559694767, + 0.03738560900092125, + 0.03505351021885872, + 0.030977141112089157, + 0.03886236622929573, + 0.039759136736392975, + 0.028616566210985184, + 0.02002212032675743, + 0.008856509812176228, + 0.011608989909291267, + 0.009767611511051655, + 0.027070695534348488, + 0.014166896231472492, + 0.016513533890247345, + 0.01141565665602684, + 0.019864127039909363, + 0.007505911402404308, + 0.009518267586827278, + 0.008408633060753345, + 0.006725327577441931, + 0.008483747951686382, + 0.01192949153482914, + 0.010638762265443802, + 0.008486012928187847, + 0.0057942853309214115, + 0.00915033183991909, + 0.009725467301905155, + 0.01252239290624857, + 0.005405597388744354, + 0.011119838804006577, + 0.00935639813542366, + 0.005850452929735184, + 0.010524296201765537, + 0.011035802774131298, + 0.009123112075030804, + 0.0061462558805942535, + 0.006682023871690035, + 0.004668607376515865, + 0.005062940530478954, + 0.007557826116681099, + 0.010546782054007053, + 0.010441316291689873, + 0.012856114655733109, + 0.006545036565512419, + 0.009821097366511822, + 0.015447858721017838, + 0.006478662136942148, + 0.013866004534065723, + 0.022291753441095352, + 0.021733209490776062, + 0.012122103944420815, + 0.008486232720315456, + 0.009685331955552101, + 0.014777003787457943, + 0.00792269129306078, + 0.007177360355854034, + 0.033490557223558426, + 0.024721037596464157, + 0.015114666894078255, + 0.017712948843836784, + 0.013724327087402344, + 0.013877838850021362, + 0.010397837497293949, + 0.011547958478331566, + 0.005999480374157429, + 0.005288716405630112, + 0.004731415770947933, + 0.004236315842717886, + 0.0035894461907446384, + 0.004985397681593895, + 0.00292297825217247, + 0.0033169370144605637, + 0.009322777390480042, + 0.008650511503219604, + 0.005735981743782759, + 0.004175203386694193, + 0.007342453580349684, + 0.008456798270344734, + 0.0028791273944079876, + 0.00410044752061367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.28132522106170654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249288398772478 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249288398772478, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249288398772478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21496592462062836, + 0.15515068173408508, + 0.10598311573266983, + 0.12669022381305695, + 0.09590935707092285, + 0.08467225730419159, + 0.04597204923629761, + 0.06761015951633453, + 0.04210696369409561, + 0.0609392412006855 + ], + "credit_map": [ + 0.020775172859430313, + 0.029124485328793526, + 0.0348566472530365, + 0.024201007559895515, + 0.04272883012890816, + 0.026057444512844086, + 0.020954938605427742, + 0.016267402097582817, + 0.019902806729078293, + 0.014825846999883652, + 0.011558767408132553, + 0.010591931641101837, + 0.013282672502100468, + 0.0157648753374815, + 0.03550264984369278, + 0.033721111714839935, + 0.00956408679485321, + 0.0153200076892972, + 0.015714196488261223, + 0.014581333845853806, + 0.011547726579010487, + 0.01057379599660635, + 0.01820451021194458, + 0.010477452538907528, + 0.010589847341179848, + 0.017366209998726845, + 0.028185637667775154, + 0.017384370788931847, + 0.01178411953151226, + 0.014035643078386784, + 0.013615037314593792, + 0.013729347847402096, + 0.012080389074981213, + 0.01760455220937729, + 0.014347914606332779, + 0.008760737255215645, + 0.009386587888002396, + 0.016886448487639427, + 0.008624875918030739, + 0.008217837661504745, + 0.01170447189360857, + 0.010095741599798203, + 0.014588030986487865, + 0.005429308395832777, + 0.012327850796282291, + 0.01741826720535755, + 0.004889163188636303, + 0.008219423703849316, + 0.01062319427728653, + 0.007817179895937443, + 0.006330237723886967, + 0.0038678953424096107, + 0.0047945198602974415, + 0.005449487827718258, + 0.00373732834123075, + 0.0033522064331918955, + 0.013082703575491905, + 0.010404987260699272, + 0.007650271523743868, + 0.007705552037805319, + 0.0076193008571863174, + 0.008070004172623158, + 0.005926876328885555, + 0.007150453981012106, + 0.0053659165278077126, + 0.004834607243537903, + 0.005715054459869862, + 0.005058575421571732, + 0.00451182946562767, + 0.007841098122298717, + 0.0033625257201492786, + 0.005417353939265013, + 0.007097779307514429, + 0.009077323600649834, + 0.0093834875151515, + 0.006413191091269255, + 0.007008134387433529, + 0.009899350814521313, + 0.004489014856517315, + 0.007570951245725155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005249288398772478 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, extreme close-up, framing the scene with a distinctive retro-cyberpunk aesthetic, focuses on a cybernetic hand, its surface a perfect mirrored chrome sheen, broken by thin, dark seams. As persistent raindrops fall, they instantly hiss and evaporate into tiny vapors upon contact with the hand's superheated metal. The hand carefully grips a pair of intricate pruners, their delicate blades etched with microcircuits and emitting a steady, cool cyan glow. With mechanical precision, the pruners hover mere millimeters above a bioluminescent plant. The plant's petals are in a tragic state of decay; they slowly wilt, their edges curling inward, as the ethereal natural light fades, its color shifting to a dull, sickly copper hue. The entire scene is primarily illuminated by the moody, flickering purple-red glow of a large, out-of-focus holographic sign set against a distant, rain-slicked cityscape. Its light catches and reflects off the damp chrome surface, creating an atmosphere of gloom, concentration, and melancholy.", + "seed": 4028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_7601_seed_4028_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": -0.3092202732339501, + "vlm_logprob": -0.3092202732339501, + "vlm_yes_no_margin": 2.0, + "vlm_token_credit": -0.3092202732339501, + "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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0548629760742188 + ], + "opposite_logprob_tokens": [ + -0.42986297607421875 + ], + "logprob_reward": -1.0548629760742188, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0548629760742188, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2814837098121643, + 0.04007720574736595, + 0.08049020916223526, + 0.17192021012306213, + 0.054335009306669235, + 0.05492706969380379, + 0.10322708636522293, + 0.021418137475848198, + 0.06482557207345963, + 0.12729571759700775 + ], + "credit_map": [ + 0.048751939088106155, + 0.030802719295024872, + 0.048814091831445694, + 0.029328148812055588, + 0.021774886175990105, + 0.017431113868951797, + 0.030363457277417183, + 0.05421734228730202, + 0.005547528620809317, + 0.005329179111868143, + 0.005860032979398966, + 0.005110433325171471, + 0.0033463099971413612, + 0.0047739045694470406, + 0.004489867947995663, + 0.00561995105817914, + 0.012219921685755253, + 0.02169010415673256, + 0.014411728829145432, + 0.006352917291224003, + 0.004526249133050442, + 0.005973513238132, + 0.005565497558563948, + 0.00975027959793806, + 0.023751292377710342, + 0.03775995969772339, + 0.03212765231728554, + 0.016576232388615608, + 0.012211299501359463, + 0.012556575238704681, + 0.013667664490640163, + 0.023269539698958397, + 0.010565407574176788, + 0.00936983060091734, + 0.008478744886815548, + 0.0063303392380476, + 0.0031380148138850927, + 0.004317635670304298, + 0.004356937482953072, + 0.007778099272400141, + 0.00653206929564476, + 0.008188835345208645, + 0.008016359061002731, + 0.007773982360959053, + 0.004765747580677271, + 0.004558925982564688, + 0.007493056822568178, + 0.0075980983674526215, + 0.017795488238334656, + 0.01235205214470625, + 0.017317751422524452, + 0.013019734062254429, + 0.004085399676114321, + 0.004883671645075083, + 0.009364563040435314, + 0.024408427998423576, + 0.0031853914260864258, + 0.003381407354027033, + 0.0027376869693398476, + 0.0023665125481784344, + 0.0019914028234779835, + 0.001981632085517049, + 0.002878893632441759, + 0.0028952108696103096, + 0.00786567758768797, + 0.0049553546123206615, + 0.011131132021546364, + 0.0074307117611169815, + 0.006274332292377949, + 0.008306427858769894, + 0.009199859574437141, + 0.009662076830863953, + 0.013424062170088291, + 0.012734168209135532, + 0.02468547224998474, + 0.01479119248688221, + 0.015190516598522663, + 0.014226662926375866, + 0.01749301142990589, + 0.01475063432008028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -1.0548629760742188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31356924772262573 + ], + "opposite_logprob_tokens": [ + -1.3135693073272705 + ], + "logprob_reward": -0.31356924772262573, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31356924772262573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1456204205751419, + 0.08710241317749023, + 0.13056664168834686, + 0.07920338213443756, + 0.07437464594841003, + 0.06035098060965538, + 0.10223976522684097, + 0.08415232598781586, + 0.03278322145342827, + 0.2036062777042389 + ], + "credit_map": [ + 0.020831115543842316, + 0.02452419511973858, + 0.02028290554881096, + 0.021084370091557503, + 0.015080884099006653, + 0.01123716775327921, + 0.013559401035308838, + 0.019020376726984978, + 0.017558397725224495, + 0.013346810825169086, + 0.012460664846003056, + 0.007269918918609619, + 0.008487113751471043, + 0.0062528494745492935, + 0.007717379834502935, + 0.014009271748363972, + 0.01967903971672058, + 0.02814355678856373, + 0.019223835319280624, + 0.012802754528820515, + 0.012247663922607899, + 0.008930456824600697, + 0.009625367820262909, + 0.01991395466029644, + 0.012603135779500008, + 0.014478940516710281, + 0.013044155202805996, + 0.00877812784165144, + 0.006661641877144575, + 0.00553059671074152, + 0.006567937787622213, + 0.011538838967680931, + 0.015194817446172237, + 0.011033358052372932, + 0.00995673332363367, + 0.009696689434349537, + 0.004688755609095097, + 0.0050317938439548016, + 0.008844862692058086, + 0.00992762390524149, + 0.006765502039343119, + 0.012921231798827648, + 0.008099911734461784, + 0.007576447445899248, + 0.004832336213439703, + 0.00451222388073802, + 0.00715286610648036, + 0.008490455336868763, + 0.023246921598911285, + 0.01254958100616932, + 0.015418396331369877, + 0.010559097863733768, + 0.007722232956439257, + 0.005701281595975161, + 0.009508227929472923, + 0.017534013837575912, + 0.01584467478096485, + 0.012138685211539268, + 0.01156365405768156, + 0.012052328325808048, + 0.006162042263895273, + 0.00639929948374629, + 0.008342941291630268, + 0.011648692190647125, + 0.004216943401843309, + 0.003212915500625968, + 0.005140882451087236, + 0.0036740154027938843, + 0.0023387775290757418, + 0.00272486824542284, + 0.005454635247588158, + 0.006020179018378258, + 0.023710252717137337, + 0.0161293838173151, + 0.03361671045422554, + 0.020841514691710472, + 0.022969378158450127, + 0.0194364245980978, + 0.031562622636556625, + 0.035339973866939545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.31356924772262573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496795102953911 + ], + "opposite_logprob_tokens": [ + -2.9299678802490234 + ], + "logprob_reward": -0.05496795102953911, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496795102953911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38506895303726196, + 0.0248758215457201, + 0.12017425894737244, + 0.1003892794251442, + 0.05416932702064514, + 0.025052331387996674, + 0.13122349977493286, + 0.01986483857035637, + 0.02297121100127697, + 0.1162104457616806 + ], + "credit_map": [ + 0.058192458003759384, + 0.03199823945760727, + 0.06509031355381012, + 0.053072553128004074, + 0.029145346954464912, + 0.03142508864402771, + 0.03642001003026962, + 0.07972494512796402, + 0.003913909196853638, + 0.003587208455428481, + 0.003599968971684575, + 0.0024323391262441874, + 0.0019460230832919478, + 0.0023040100932121277, + 0.0024705748073756695, + 0.0046217916533350945, + 0.015434152446687222, + 0.02200046367943287, + 0.01867186278104782, + 0.013694515451788902, + 0.013851059600710869, + 0.009282667189836502, + 0.010775092989206314, + 0.01646445319056511, + 0.014215145260095596, + 0.021328484639525414, + 0.018445316702127457, + 0.010002447292208672, + 0.0072661032900214195, + 0.007281527854502201, + 0.008295952342450619, + 0.013554305769503117, + 0.00785556435585022, + 0.009164303541183472, + 0.007635787129402161, + 0.008199782110750675, + 0.0038437945768237114, + 0.004765010438859463, + 0.004819165915250778, + 0.00788592454046011, + 0.00352139538154006, + 0.004319903440773487, + 0.003934331703931093, + 0.0028511786367744207, + 0.0020446046255528927, + 0.001727040158584714, + 0.0031511674169451, + 0.003502710722386837, + 0.010052921250462532, + 0.01777385175228119, + 0.022219764068722725, + 0.016123289242386818, + 0.008318914100527763, + 0.01099589467048645, + 0.012331507168710232, + 0.03340737149119377, + 0.0028157152701169252, + 0.002698490396142006, + 0.0027779729571193457, + 0.0022650870960205793, + 0.0021107725333422422, + 0.0024030134081840515, + 0.002215850865468383, + 0.0025779360439628363, + 0.0027281378861516714, + 0.0017487616278231144, + 0.003951346036046743, + 0.002694781869649887, + 0.001996031729504466, + 0.002846507355570793, + 0.003665566211566329, + 0.0033400796819478273, + 0.009684698656201363, + 0.012035194784402847, + 0.019108612090349197, + 0.013678807765245438, + 0.018510257825255394, + 0.011166511103510857, + 0.015521437861025333, + 0.016504937782883644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.05496795102953911 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292415112257004 + ], + "opposite_logprob_tokens": [ + -2.237924098968506 + ], + "logprob_reward": -0.11292415112257004, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292415112257004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13591347634792328, + 0.1613283008337021, + 0.07799401879310608, + 0.08456666022539139, + 0.1986156404018402, + 0.06335075944662094, + 0.10862890630960464, + 0.040091242641210556, + 0.04553339257836342, + 0.08397765457630157 + ], + "credit_map": [ + 0.02173854596912861, + 0.02559272013604641, + 0.022541647776961327, + 0.015332365408539772, + 0.011760415509343147, + 0.009067497216165066, + 0.013771471567451954, + 0.01610880345106125, + 0.03912161663174629, + 0.026126975193619728, + 0.02523975260555744, + 0.01419053040444851, + 0.01009301096200943, + 0.010797891765832901, + 0.015288942493498325, + 0.02046956866979599, + 0.018989747390151024, + 0.013007622212171555, + 0.014685974456369877, + 0.0055413939990103245, + 0.005178338382393122, + 0.004528750199824572, + 0.00525579834356904, + 0.010806381702423096, + 0.011347501538693905, + 0.021371301263570786, + 0.015406832098960876, + 0.007647040765732527, + 0.0057847159914672375, + 0.005412427242845297, + 0.006358223035931587, + 0.011238601058721542, + 0.049810826778411865, + 0.037885382771492004, + 0.029676970094442368, + 0.016537081450223923, + 0.010318182408809662, + 0.009314160794019699, + 0.015555812045931816, + 0.029517194256186485, + 0.012085577473044395, + 0.01679673232138157, + 0.010013031773269176, + 0.004697864875197411, + 0.0037027059588581324, + 0.0036107187625020742, + 0.0053276605904102325, + 0.0071164583787322044, + 0.01455764751881361, + 0.021344250068068504, + 0.017352094873785973, + 0.010621333494782448, + 0.007545533590018749, + 0.008174293674528599, + 0.00954584963619709, + 0.0194878987967968, + 0.008804701268672943, + 0.004603974986821413, + 0.005386754404753447, + 0.005217639729380608, + 0.003158747451379895, + 0.0035098937805742025, + 0.004109393339604139, + 0.005300130695104599, + 0.005665878299623728, + 0.005223297048360109, + 0.007080822251737118, + 0.0056927502155303955, + 0.0037037250585854053, + 0.005401800852268934, + 0.006451538763940334, + 0.0063135758973658085, + 0.01676803268492222, + 0.013770178891718388, + 0.012330735102295876, + 0.011743145994842052, + 0.007514670956879854, + 0.004621165804564953, + 0.008567843586206436, + 0.008661882020533085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.11292415112257004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777040220797062 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.009777040220797062, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777040220797062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1779763102531433, + 0.09581469744443893, + 0.06052326038479805, + 0.15089315176010132, + 0.05871012434363365, + 0.09443771839141846, + 0.18914785981178284, + 0.037404272705316544, + 0.0422447994351387, + 0.09284783154726028 + ], + "credit_map": [ + 0.02903909981250763, + 0.03531461954116821, + 0.032069362699985504, + 0.014666381292045116, + 0.014621674083173275, + 0.010030909441411495, + 0.016650471836328506, + 0.02558378502726555, + 0.01882012002170086, + 0.00978852715343237, + 0.016242682933807373, + 0.00824650563299656, + 0.00858734454959631, + 0.00648429524153471, + 0.010003351606428623, + 0.01764187030494213, + 0.01281814370304346, + 0.0071715195663273335, + 0.008316714316606522, + 0.007058423478156328, + 0.005905597470700741, + 0.004046775866299868, + 0.005181417800486088, + 0.01002467144280672, + 0.022103145718574524, + 0.028612257912755013, + 0.026907583698630333, + 0.015433351509273052, + 0.010095065459609032, + 0.010251960717141628, + 0.014312301762402058, + 0.023177480325102806, + 0.00871607568114996, + 0.009225555695593357, + 0.009733336977660656, + 0.006883027032017708, + 0.0033757290802896023, + 0.004934167955070734, + 0.005738414824008942, + 0.010103820823132992, + 0.014146124012768269, + 0.010100772604346275, + 0.016266649588942528, + 0.01340221893042326, + 0.007817236706614494, + 0.006103605963289738, + 0.007905924692749977, + 0.018695184960961342, + 0.05425006523728371, + 0.03691811487078667, + 0.02874147519469261, + 0.014846635051071644, + 0.008918401785194874, + 0.008576760068535805, + 0.010808253660798073, + 0.026088150218129158, + 0.006396139971911907, + 0.004161118995398283, + 0.004602930974215269, + 0.004227948375046253, + 0.0039075566455721855, + 0.0033649387769401073, + 0.005045795813202858, + 0.0056978436186909676, + 0.00487213209271431, + 0.0044937445782125, + 0.006981092970818281, + 0.005175093654543161, + 0.0040200697258114815, + 0.00419125659391284, + 0.006402024533599615, + 0.006109386216849089, + 0.01552471425384283, + 0.015059281140565872, + 0.015822388231754303, + 0.013232595287263393, + 0.006684086751192808, + 0.004932792391628027, + 0.01067543774843216, + 0.010916538536548615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.009777040220797062 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, extreme close-up, framing the scene with a distinctive retro-cyberpunk aesthetic, focuses on a cybernetic hand, its surface a perfect mirrored chrome sheen, broken by thin, dark seams. As persistent raindrops fall, they instantly hiss and evaporate into tiny vapors upon contact with the hand's superheated metal. The hand carefully grips a pair of intricate pruners, their delicate blades etched with microcircuits and emitting a steady, cool cyan glow. With mechanical precision, the pruners hover mere millimeters above a bioluminescent plant. The plant's petals are in a tragic state of decay; they slowly wilt, their edges curling inward, as the ethereal natural light fades, its color shifting to a dull, sickly copper hue. The entire scene is primarily illuminated by the moody, flickering purple-red glow of a large, out-of-focus holographic sign set against a distant, rain-slicked cityscape. Its light catches and reflects off the damp chrome surface, creating an atmosphere of gloom, concentration, and melancholy.", + "seed": 4029, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_7601_seed_4029_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": -0.09620121819898486, + "vlm_logprob": -0.09620121819898486, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.09620121819898486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03880197927355766 + ], + "opposite_logprob_tokens": [ + -3.288801908493042 + ], + "logprob_reward": -0.03880197927355766, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03880197927355766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25257670879364014, + 0.12684893608093262, + 0.15755192935466766, + 0.08655884861946106, + 0.0536111518740654, + 0.042759474366903305, + 0.05858935788273811, + 0.04691440239548683, + 0.10588159412145615, + 0.06870761513710022 + ], + "credit_map": [ + 0.03552056849002838, + 0.024076415225863457, + 0.033433523029088974, + 0.02517276629805565, + 0.039466653019189835, + 0.028234580531716347, + 0.038209784775972366, + 0.028462426736950874, + 0.008859098888933659, + 0.007257051300257444, + 0.02972245402634144, + 0.010938574559986591, + 0.007732234429568052, + 0.03328332304954529, + 0.01566474139690399, + 0.0133914640173316, + 0.028877275064587593, + 0.018091673031449318, + 0.02170196920633316, + 0.023668264970183372, + 0.015303457155823708, + 0.023697469383478165, + 0.015111335553228855, + 0.011100481264293194, + 0.009986505843698978, + 0.008655154146254063, + 0.010391246527433395, + 0.017742568626999855, + 0.009858474135398865, + 0.011655498296022415, + 0.011515681631863117, + 0.0067537156865000725, + 0.007077164016664028, + 0.005802595056593418, + 0.005846912506967783, + 0.00845738872885704, + 0.006426122505217791, + 0.0076597039587795734, + 0.008565452881157398, + 0.0037758126854896545, + 0.006529301404953003, + 0.004233479034155607, + 0.0058801136910915375, + 0.005033234134316444, + 0.004319099243730307, + 0.008472606539726257, + 0.0055271717719733715, + 0.0027644678484648466, + 0.009593411348760128, + 0.005532467272132635, + 0.006630587857216597, + 0.008437331765890121, + 0.006757786497473717, + 0.010771344415843487, + 0.007478620857000351, + 0.003387812292203307, + 0.006435276474803686, + 0.004764167591929436, + 0.004507431760430336, + 0.010982926934957504, + 0.0030938670970499516, + 0.006813296116888523, + 0.00805744156241417, + 0.00225999322719872, + 0.012886623851954937, + 0.009836042299866676, + 0.012105575762689114, + 0.02847355045378208, + 0.008695108816027641, + 0.016065889969468117, + 0.011542503722012043, + 0.006276295054703951, + 0.00646498566493392, + 0.004981827922165394, + 0.0163138285279274, + 0.007891574874520302, + 0.005856749136000872, + 0.01858377642929554, + 0.0048085786402225494, + 0.003806284861639142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03880197927355766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08911328762769699 + ], + "opposite_logprob_tokens": [ + -2.464113235473633 + ], + "logprob_reward": -0.08911328762769699, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08911328762769699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2123597413301468, + 0.1661272943019867, + 0.1337173581123352, + 0.1051134541630745, + 0.04537652060389519, + 0.03304912522435188, + 0.08763575553894043, + 0.05312345549464226, + 0.13258060812950134, + 0.030916687101125717 + ], + "credit_map": [ + 0.026957189664244652, + 0.01906212605535984, + 0.03840513527393341, + 0.04147268086671829, + 0.015463600866496563, + 0.027725202962756157, + 0.02711578831076622, + 0.016158025711774826, + 0.016876263543963432, + 0.011318624950945377, + 0.047350745648145676, + 0.00916763860732317, + 0.008091716095805168, + 0.05771205946803093, + 0.007484028115868568, + 0.008126228116452694, + 0.017720282077789307, + 0.013254511170089245, + 0.019437657669186592, + 0.034266650676727295, + 0.00872545875608921, + 0.020092761144042015, + 0.012860454618930817, + 0.007359587587416172, + 0.020950494334101677, + 0.010782308876514435, + 0.011100939474999905, + 0.024576492607593536, + 0.007533616852015257, + 0.015362362377345562, + 0.00992534589022398, + 0.004881903529167175, + 0.004747353494167328, + 0.004179544746875763, + 0.0053927842527627945, + 0.011228823103010654, + 0.0036823658738285303, + 0.006994415540248156, + 0.0068179951049387455, + 0.0023332415148615837, + 0.004858052358031273, + 0.0036694230511784554, + 0.004293056670576334, + 0.004811299033463001, + 0.003426689188927412, + 0.00510288355872035, + 0.0047949436120688915, + 0.0020927786827087402, + 0.01271645538508892, + 0.005927893798798323, + 0.009643078781664371, + 0.01955912634730339, + 0.005602374207228422, + 0.01438258495181799, + 0.015344920568168163, + 0.004459324758499861, + 0.0067259627394378185, + 0.0037148145493119955, + 0.006478644907474518, + 0.013241578824818134, + 0.0030759915243834257, + 0.007825215347111225, + 0.009652896784245968, + 0.002408349420875311, + 0.016222333535552025, + 0.011195861734449863, + 0.014827577397227287, + 0.037541914731264114, + 0.010797785595059395, + 0.020991267636418343, + 0.014136580750346184, + 0.0068673002533614635, + 0.0027432083152234554, + 0.0027803725097328424, + 0.006341337691992521, + 0.002374662784859538, + 0.0032444726675748825, + 0.006982570514082909, + 0.004084376152604818, + 0.002365688094869256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.08911328762769699 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370270878076553 + ], + "opposite_logprob_tokens": [ + -3.4087026119232178 + ], + "logprob_reward": -0.03370270878076553, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370270878076553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16452878713607788, + 0.2141832858324051, + 0.17381958663463593, + 0.06589118391275406, + 0.07832702994346619, + 0.07921641319990158, + 0.03349282965064049, + 0.04968784376978874, + 0.06811147928237915, + 0.07274157553911209 + ], + "credit_map": [ + 0.02136881649494171, + 0.01620926521718502, + 0.02524818666279316, + 0.021611781790852547, + 0.01674373634159565, + 0.022948037832975388, + 0.02126581408083439, + 0.019133159890770912, + 0.015422995202243328, + 0.008905578404664993, + 0.05967621132731438, + 0.013712026178836823, + 0.013645283877849579, + 0.07071050256490707, + 0.01876905933022499, + 0.013341620564460754, + 0.021613555029034615, + 0.014782414771616459, + 0.024031026288866997, + 0.04603805020451546, + 0.015714263543486595, + 0.026585889980196953, + 0.015054335817694664, + 0.01000005379319191, + 0.006745730992406607, + 0.006554041523486376, + 0.007538330741226673, + 0.016523929312825203, + 0.00612366059795022, + 0.009580123238265514, + 0.008551785722374916, + 0.004273583181202412, + 0.010288122110068798, + 0.006710133049637079, + 0.007148796692490578, + 0.015086141414940357, + 0.0072969673201441765, + 0.013645109720528126, + 0.013298118487000465, + 0.004853640217334032, + 0.012127800844609737, + 0.006833734456449747, + 0.008384103886783123, + 0.018372103571891785, + 0.006423181854188442, + 0.012701672501862049, + 0.010529343970119953, + 0.003844469552859664, + 0.0063487146981060505, + 0.0038275967817753553, + 0.0046190484426915646, + 0.003984163049608469, + 0.002809482626616955, + 0.00573605764657259, + 0.004107722081243992, + 0.002060046186670661, + 0.003979956731200218, + 0.004228426143527031, + 0.005995460320264101, + 0.012781773693859577, + 0.0033819940872490406, + 0.007301644887775183, + 0.009416542947292328, + 0.0026020400691777468, + 0.0077103436924517155, + 0.009266800247132778, + 0.00592048792168498, + 0.014634383842349052, + 0.007790593430399895, + 0.010025943629443645, + 0.008744349703192711, + 0.004018578678369522, + 0.005002978257834911, + 0.004429192747920752, + 0.016288945451378822, + 0.008569959551095963, + 0.005705185700207949, + 0.020336555317044258, + 0.005374590400606394, + 0.007034169510006905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.03370270878076553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31344273686408997 + ], + "opposite_logprob_tokens": [ + -1.3134427070617676 + ], + "logprob_reward": -0.31344273686408997, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344273686408997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1356876790523529, + 0.12009056657552719, + 0.08747431635856628, + 0.21437129378318787, + 0.1018521785736084, + 0.04830258712172508, + 0.08257556706666946, + 0.15411242842674255, + 0.017291665077209473, + 0.03824169933795929 + ], + "credit_map": [ + 0.019033515825867653, + 0.01441147830337286, + 0.023851493373513222, + 0.022552408277988434, + 0.013853740878403187, + 0.019023779779672623, + 0.012693909928202629, + 0.010267357341945171, + 0.01205579936504364, + 0.011165850795805454, + 0.016400178894400597, + 0.02107526920735836, + 0.012965093366801739, + 0.01959175243973732, + 0.01273635309189558, + 0.014100270345807076, + 0.011346503160893917, + 0.009278994053602219, + 0.011162937618792057, + 0.017764344811439514, + 0.007234921678900719, + 0.01412276178598404, + 0.009003286249935627, + 0.007560567930340767, + 0.03172270581126213, + 0.018806319683790207, + 0.02523823082447052, + 0.060654737055301666, + 0.016394509002566338, + 0.030695533379912376, + 0.02250935137271881, + 0.008349914103746414, + 0.015294866636395454, + 0.011794619262218475, + 0.010055853985249996, + 0.0238487608730793, + 0.0068986183032393456, + 0.0161448922008276, + 0.013653199188411236, + 0.004161365330219269, + 0.008991802111268044, + 0.004662038292735815, + 0.0065465522930026054, + 0.006279091816395521, + 0.0053238896653056145, + 0.008017209358513355, + 0.005658390000462532, + 0.00282361195422709, + 0.012741115875542164, + 0.006327800452709198, + 0.007705638650804758, + 0.01833895780146122, + 0.007390644401311874, + 0.013627488166093826, + 0.012499033473432064, + 0.0039448924362659454, + 0.020359216257929802, + 0.012451757676899433, + 0.020064665004611015, + 0.03558046370744705, + 0.00869076419621706, + 0.024436164647340775, + 0.026954706758260727, + 0.005574691575020552, + 0.0019397326977923512, + 0.0021218021865934134, + 0.002137948526069522, + 0.0030272183939814568, + 0.002065773820504546, + 0.0022868874948471785, + 0.0024474943056702614, + 0.0012648062547668815, + 0.003018632996827364, + 0.0029808238614350557, + 0.007621074095368385, + 0.0035908371210098267, + 0.004482287913560867, + 0.009493534453213215, + 0.003865752136334777, + 0.0031887537334114313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.31344273686408997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945378448814154 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.005945378448814154, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945378448814154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14478828012943268, + 0.055614061653614044, + 0.08344519138336182, + 0.08748674392700195, + 0.29159682989120483, + 0.042695775628089905, + 0.10103505849838257, + 0.05392415076494217, + 0.05433756113052368, + 0.08507633209228516 + ], + "credit_map": [ + 0.017804067581892014, + 0.015249323099851608, + 0.025480831041932106, + 0.022593043744564056, + 0.01641651801764965, + 0.019982926547527313, + 0.016832131892442703, + 0.01042943075299263, + 0.005264005158096552, + 0.004205980338156223, + 0.010119219310581684, + 0.007027347572147846, + 0.00426936661824584, + 0.013342075981199741, + 0.005509816110134125, + 0.005876250099390745, + 0.014830938540399075, + 0.010308190248906612, + 0.011073296889662743, + 0.01379968412220478, + 0.006952779833227396, + 0.01285475306212902, + 0.00870680995285511, + 0.004918738268315792, + 0.017581168562173843, + 0.010377487167716026, + 0.009217402897775173, + 0.017502956092357635, + 0.007695530075579882, + 0.012875898741185665, + 0.007291974034160376, + 0.004944325890392065, + 0.04105638340115547, + 0.02461596392095089, + 0.02621348761022091, + 0.07547634094953537, + 0.019821878522634506, + 0.05029367282986641, + 0.04134152829647064, + 0.012777568772435188, + 0.0050780996680259705, + 0.00444232951849699, + 0.0059684389270842075, + 0.008597567677497864, + 0.0027198405005037785, + 0.007547378074377775, + 0.005889721680432558, + 0.00245239888317883, + 0.015808073803782463, + 0.007158690132200718, + 0.009888281114399433, + 0.0230293869972229, + 0.0076766847632825375, + 0.01593048684298992, + 0.01655416563153267, + 0.004989293869584799, + 0.009167639538645744, + 0.0037812215741723776, + 0.005873875226825476, + 0.013001730665564537, + 0.003216729499399662, + 0.008626426570117474, + 0.007880504243075848, + 0.002376020886003971, + 0.006291213911026716, + 0.004523547366261482, + 0.00597020098939538, + 0.015038315206766129, + 0.004328891634941101, + 0.008547523058950901, + 0.006817542016506195, + 0.002820325316861272, + 0.00598868727684021, + 0.0047876122407615185, + 0.020841017365455627, + 0.007943934760987759, + 0.0069301011972129345, + 0.025881921872496605, + 0.007859807461500168, + 0.004843253176659346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.005945378448814154 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly static, extreme close-up, framing the scene with a distinctive retro-cyberpunk aesthetic, focuses on a cybernetic hand, its surface a perfect mirrored chrome sheen, broken by thin, dark seams. As persistent raindrops fall, they instantly hiss and evaporate into tiny vapors upon contact with the hand's superheated metal. The hand carefully grips a pair of intricate pruners, their delicate blades etched with microcircuits and emitting a steady, cool cyan glow. With mechanical precision, the pruners hover mere millimeters above a bioluminescent plant. The plant's petals are in a tragic state of decay; they slowly wilt, their edges curling inward, as the ethereal natural light fades, its color shifting to a dull, sickly copper hue. The entire scene is primarily illuminated by the moody, flickering purple-red glow of a large, out-of-focus holographic sign set against a distant, rain-slicked cityscape. Its light catches and reflects off the damp chrome surface, creating an atmosphere of gloom, concentration, and melancholy.", + "seed": 4030, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_10/index_7601_seed_4030_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": -0.15089367642067372, + "vlm_logprob": -0.15089367642067372, + "vlm_yes_no_margin": 2.5999999761581423, + "vlm_token_credit": -0.15089367642067372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12831149995326996 + ], + "opposite_logprob_tokens": [ + -2.1283113956451416 + ], + "logprob_reward": -0.12831149995326996, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12831149995326996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45880943536758423, + 0.05330893024802208, + 0.025562124326825142, + 0.03218350186944008, + 0.043760400265455246, + 0.02809634618461132, + 0.03517920896410942, + 0.01679561287164688, + 0.05260327458381653, + 0.25370118021965027 + ], + "credit_map": [ + 0.05105258896946907, + 0.018187616020441055, + 0.16291485726833344, + 0.03213809058070183, + 0.0387495718896389, + 0.026198020204901695, + 0.027112441137433052, + 0.10245627164840698, + 0.008281899616122246, + 0.002136257477104664, + 0.004515775479376316, + 0.004867741838097572, + 0.004578960593789816, + 0.010359360836446285, + 0.006858409382402897, + 0.011710528284311295, + 0.0038132648915052414, + 0.0030233038123697042, + 0.002656439086422324, + 0.002818661741912365, + 0.0028157010674476624, + 0.002414856106042862, + 0.003479222534224391, + 0.004540675785392523, + 0.005587518680840731, + 0.004383854568004608, + 0.005505005829036236, + 0.0034023146145045757, + 0.0029624709859490395, + 0.0027937821578234434, + 0.0040584043599665165, + 0.0034901518374681473, + 0.006947270128875971, + 0.003693697042763233, + 0.007267811801284552, + 0.004381007049232721, + 0.007889932952821255, + 0.003728155279532075, + 0.0038030026480555534, + 0.0060495235957205296, + 0.005097015295177698, + 0.0023749752435833216, + 0.005242897663265467, + 0.0028419552836567163, + 0.002187122358009219, + 0.0034285206347703934, + 0.002924202475696802, + 0.003999659791588783, + 0.006887160241603851, + 0.0019703847356140614, + 0.010904351249337196, + 0.003741072490811348, + 0.002776411361992359, + 0.003278032410889864, + 0.0021464640740305185, + 0.003475333796814084, + 0.0026798166800290346, + 0.001546312472783029, + 0.0019413413247093558, + 0.0024594482965767384, + 0.001971522578969598, + 0.0022183461114764214, + 0.0014838273636996746, + 0.0024949980434030294, + 0.008152109570801258, + 0.003941302187740803, + 0.009938864968717098, + 0.005151489283889532, + 0.00504326680675149, + 0.0036726908292621374, + 0.010820663534104824, + 0.005882887169718742, + 0.033990323543548584, + 0.014014174230396748, + 0.044417258352041245, + 0.03337052837014198, + 0.01824960485100746, + 0.03100772202014923, + 0.02736835740506649, + 0.051283203065395355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.12831149995326996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487464189529419 + ], + "opposite_logprob_tokens": [ + -1.223746418952942 + ], + "logprob_reward": -0.3487464189529419, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3487464189529419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42797768115997314, + 0.047478266060352325, + 0.0538986437022686, + 0.02323816902935505, + 0.05370301753282547, + 0.038097038865089417, + 0.05184606835246086, + 0.015566017478704453, + 0.07524412125349045, + 0.2129509150981903 + ], + "credit_map": [ + 0.046597983688116074, + 0.015491386875510216, + 0.15520717203617096, + 0.03249151259660721, + 0.03726474940776825, + 0.023618848994374275, + 0.022626781836152077, + 0.09467925131320953, + 0.00943994615226984, + 0.004155795555561781, + 0.005620818585157394, + 0.006976251024752855, + 0.0039327433332800865, + 0.005257468204945326, + 0.006240545306354761, + 0.005854697432368994, + 0.009411913342773914, + 0.008168067783117294, + 0.006166275590658188, + 0.005463572219014168, + 0.006510571576654911, + 0.0038421691861003637, + 0.005163973197340965, + 0.009172103367745876, + 0.0030145577620714903, + 0.0028532398864626884, + 0.0036577321588993073, + 0.0025529423728585243, + 0.0026397884357720613, + 0.0020801862701773643, + 0.0034501354675740004, + 0.0029895866755396128, + 0.010823929682374, + 0.007921704091131687, + 0.007867383770644665, + 0.005470278672873974, + 0.005305304192006588, + 0.003932871390134096, + 0.003435818711295724, + 0.008945731446146965, + 0.005015464499592781, + 0.0025650723837316036, + 0.009020340628921986, + 0.0042070187628269196, + 0.0026928421575576067, + 0.0030346184503287077, + 0.0032217931002378464, + 0.008339890278875828, + 0.008393608964979649, + 0.0025644239503890276, + 0.018724137917160988, + 0.0047470200806856155, + 0.002699471777305007, + 0.0031270848121494055, + 0.002336559584364295, + 0.00925376359373331, + 0.0027499005664139986, + 0.0015444174641743302, + 0.0024471592623740435, + 0.001941362046636641, + 0.001701654982753098, + 0.0013549852883443236, + 0.0014171904185786843, + 0.002409347565844655, + 0.011822779662907124, + 0.008990485221147537, + 0.019221406430006027, + 0.005857693497091532, + 0.007224747911095619, + 0.004987560678273439, + 0.009700308553874493, + 0.0074391416274011135, + 0.029653124511241913, + 0.014151503331959248, + 0.03513826057314873, + 0.025663232430815697, + 0.01576288975775242, + 0.02490254119038582, + 0.02665262296795845, + 0.04102674499154091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.3487464189529419 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865897819399834 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865897819399834, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865897819399834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.263043075799942, + 0.10053517669439316, + 0.10605562478303909, + 0.0650683268904686, + 0.0704742893576622, + 0.030844412744045258, + 0.10310651361942291, + 0.020629795268177986, + 0.08544329553842545, + 0.15479949116706848 + ], + "credit_map": [ + 0.027837222442030907, + 0.009843636304140091, + 0.09972123056650162, + 0.018657632172107697, + 0.02159496583044529, + 0.01376670878380537, + 0.011102145537734032, + 0.060519538819789886, + 0.013996904715895653, + 0.005288667976856232, + 0.0099541787058115, + 0.009262114763259888, + 0.008876076899468899, + 0.015717172995209694, + 0.021823616698384285, + 0.01561644859611988, + 0.014836913906037807, + 0.015105096623301506, + 0.009721397422254086, + 0.012542657554149628, + 0.01031310111284256, + 0.011676791124045849, + 0.017799708992242813, + 0.014059951528906822, + 0.01564301922917366, + 0.003859155112877488, + 0.0080432603135705, + 0.004663153551518917, + 0.00529502285644412, + 0.004979276098310947, + 0.011052721180021763, + 0.011532719247043133, + 0.0060517010278999805, + 0.0049177296459674835, + 0.01812759041786194, + 0.004936144687235355, + 0.004920912440866232, + 0.005569893401116133, + 0.006820365320891142, + 0.019129948690533638, + 0.0049880677834153175, + 0.0035822573117911816, + 0.006336280610412359, + 0.003125052200630307, + 0.0031799606513231993, + 0.0030730627477169037, + 0.0032460971269756556, + 0.00331363407894969, + 0.020962338894605637, + 0.005571560934185982, + 0.037767767906188965, + 0.009498698636889458, + 0.006038756109774113, + 0.005510483402758837, + 0.005535805132240057, + 0.012221098877489567, + 0.0031050865072757006, + 0.0016537007177248597, + 0.0031048106029629707, + 0.0030552432872354984, + 0.0024895011447370052, + 0.002710318425670266, + 0.00233500637114048, + 0.002176129026338458, + 0.01688416674733162, + 0.004125719889998436, + 0.009586772881448269, + 0.0070699346251785755, + 0.007271775510162115, + 0.005128985736519098, + 0.025018351152539253, + 0.01035759225487709, + 0.01825331337749958, + 0.00990776065737009, + 0.026955073699355125, + 0.02387261763215065, + 0.015452084131538868, + 0.019941309466958046, + 0.018774239346385002, + 0.021643085405230522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.04865897819399834 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556200623512268 + ], + "opposite_logprob_tokens": [ + -1.6005619764328003 + ], + "logprob_reward": -0.22556200623512268, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556200623512268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15568269789218903, + 0.04882313683629036, + 0.08774901181459427, + 0.10482729971408844, + 0.14592988789081573, + 0.12135213613510132, + 0.1566171795129776, + 0.046606242656707764, + 0.044697973877191544, + 0.08771445602178574 + ], + "credit_map": [ + 0.019274650141596794, + 0.012522230856120586, + 0.04607320576906204, + 0.01692253164947033, + 0.015793612226843834, + 0.0075574154034256935, + 0.012565874494612217, + 0.02497316151857376, + 0.008425923995673656, + 0.004881072323769331, + 0.0050721983425319195, + 0.006264023017138243, + 0.004691637121140957, + 0.007597764953970909, + 0.006265443749725819, + 0.005625074729323387, + 0.017151057720184326, + 0.01344167161732912, + 0.013555645942687988, + 0.01135227084159851, + 0.007056869566440582, + 0.006656329147517681, + 0.008730201981961727, + 0.009804955683648586, + 0.03277665376663208, + 0.007268455810844898, + 0.010603933595120907, + 0.004842259921133518, + 0.00761027168482542, + 0.004707415122538805, + 0.005124990828335285, + 0.03189331293106079, + 0.022449225187301636, + 0.020740076899528503, + 0.031022246927022934, + 0.01253865659236908, + 0.010808737017214298, + 0.008666099980473518, + 0.006884010508656502, + 0.03282083198428154, + 0.025533093139529228, + 0.01498450804501772, + 0.01994358003139496, + 0.015452984720468521, + 0.008960328996181488, + 0.008482677862048149, + 0.009895660914480686, + 0.018099302425980568, + 0.027812210842967033, + 0.007954019121825695, + 0.0545339360833168, + 0.014238977804780006, + 0.010119880549609661, + 0.010626235976815224, + 0.00845332257449627, + 0.022878581658005714, + 0.005621269345283508, + 0.003641456598415971, + 0.0055093299597501755, + 0.005759035237133503, + 0.005865118931978941, + 0.007893037050962448, + 0.006074642762541771, + 0.006242354400455952, + 0.008134358562529087, + 0.002104690531268716, + 0.004750904627144337, + 0.0035215988755226135, + 0.00339693995192647, + 0.00268668495118618, + 0.01448510866612196, + 0.005617690738290548, + 0.012111504562199116, + 0.0062470026314258575, + 0.015131879597902298, + 0.013796196319162846, + 0.006885765586048365, + 0.010189232416450977, + 0.0077962311916053295, + 0.015556642785668373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.22556200623512268 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 cybernetic hand with a mirrored chrome surface and dark seams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do raindrops fall and instantly hiss/evaporate into vapor upon contact with the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the hand gripping intricate pruners with microcircuit-etched blades emitting a steady cyan glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a bioluminescent plant visible with wilting petals shifting from natural light to a dull copper hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a large, out-of-focus holographic sign with a flickering purple-red glow against a rain-slicked cityscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031894787680357695 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.0031894787680357695, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031894787680357695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5775046944618225, + 0.03243406116962433, + 0.06158977001905441, + 0.05253244563937187, + 0.05445510521531105, + 0.039384208619594574, + 0.04988078773021698, + 0.016721319407224655, + 0.03648674860596657, + 0.07901082932949066 + ], + "credit_map": [ + 0.06623279303312302, + 0.01993986964225769, + 0.22750729322433472, + 0.04125867784023285, + 0.04205644503235817, + 0.027239087969064713, + 0.021700501441955566, + 0.13157008588314056, + 0.005355730187147856, + 0.0040761386044323444, + 0.004169109743088484, + 0.0048126280307769775, + 0.0022166180424392223, + 0.0030490458011627197, + 0.0030562791507691145, + 0.005698511842638254, + 0.009179389104247093, + 0.006482488941401243, + 0.007586919702589512, + 0.011062498204410076, + 0.007339721545577049, + 0.005475305952131748, + 0.006187064573168755, + 0.008276388049125671, + 0.01817578822374344, + 0.004023220390081406, + 0.011175808496773243, + 0.0028805481269955635, + 0.00280999974347651, + 0.002159436931833625, + 0.002558202715590596, + 0.008749446831643581, + 0.006327657029032707, + 0.005472645163536072, + 0.013231960125267506, + 0.00457811402156949, + 0.004656499717384577, + 0.004262782167643309, + 0.0035765995271503925, + 0.01234885212033987, + 0.007702535018324852, + 0.00428807595744729, + 0.0057295202277600765, + 0.004872716497629881, + 0.002785998862236738, + 0.004493669141083956, + 0.003634283784776926, + 0.005877410061657429, + 0.008722205646336079, + 0.0031828780192881823, + 0.016912363469600677, + 0.0049859448336064816, + 0.003101028734818101, + 0.003488959977403283, + 0.0025755739770829678, + 0.006911839824169874, + 0.0022879657335579395, + 0.002387701068073511, + 0.0025426261126995087, + 0.0024879418779164553, + 0.001416853629052639, + 0.002069623675197363, + 0.0013981329975649714, + 0.0021304755937308073, + 0.006991219706833363, + 0.002768266247585416, + 0.0071207634173333645, + 0.0032029864378273487, + 0.0030183102935552597, + 0.0029407301917672157, + 0.006921213120222092, + 0.0035232596565037966, + 0.010468476451933384, + 0.007253523916006088, + 0.011865253560245037, + 0.013898337259888649, + 0.006083329673856497, + 0.00862442422658205, + 0.012732071802020073, + 0.008085415698587894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0031894787680357695 + } + ], + "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/0000122/group_11/index_tensor([2708, 2708])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_11/index_tensor([2708, 2708])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bb0a926bc7bea6332dd1d662dc32077b4034e551 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_11/index_tensor([2708, 2708])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled, extreme close-up captures a scene of controlled destruction in dramatic slow motion. A thin, light-colored pine dowel, with a smooth surface and distinct grain, rests horizontally on two supporting blocks on a worn, scarred wooden workbench in a garage. Viewed from directly above, a solid, rectangular steel block drops swiftly and decisively toward the dowel's midpoint, its dark gray surface revealing a dull metallic sheen and slight wear. The entire scene is harshly illuminated by bright, direct light from an overhead fixture, casting stark, deep shadows. As the block impacts with such force, the dowel visibly bends under the immense pressure, its wood fibers straining with a sound before a catastrophic crack erupts from its center. The dowel then explodes into a cloud of sharp, jagged fragments and several larger splinters. The raw fiber texture of the freshly broken wood catches the clinical light as it is propelled outward before gracefully tumbling downward, dispersing and coming to rest on the dusty surface of the workbench, creating a focused and intense atmosphere of the physical failure of material.", + "seed": 20055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_11/index_2708_seed_20055_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": -0.14684246182441713, + "vlm_logprob": -0.14684246182441713, + "vlm_yes_no_margin": 2.4999999523162844, + "vlm_token_credit": -0.14684246182441713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22401778399944305 + ], + "opposite_logprob_tokens": [ + -1.8490177392959595 + ], + "logprob_reward": -0.22401778399944305, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.22401778399944305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25828874111175537, + 0.12530329823493958, + 0.095861054956913, + 0.04961521178483963, + 0.09510371088981628, + 0.08935889601707458, + 0.13184687495231628, + 0.05551294982433319, + 0.07578670978546143, + 0.023322517052292824 + ], + "credit_map": [ + 0.03356175497174263, + 0.041645925492048264, + 0.04557543247938156, + 0.02380925416946411, + 0.03134511783719063, + 0.03089928813278675, + 0.029604671522974968, + 0.021847333759069443, + 0.020442204549908638, + 0.014216519892215729, + 0.014984951354563236, + 0.01815587654709816, + 0.017578275874257088, + 0.012154308147728443, + 0.01463743019849062, + 0.013133733533322811, + 0.014376932755112648, + 0.011412352323532104, + 0.014638038352131844, + 0.010774370282888412, + 0.013391649350523949, + 0.012055433355271816, + 0.011901332065463066, + 0.007310961373150349, + 0.006222639698535204, + 0.005282280966639519, + 0.0076060607098042965, + 0.006031989585608244, + 0.005449915304780006, + 0.005686596967279911, + 0.004729426000267267, + 0.008606307208538055, + 0.014484187588095665, + 0.012591525912284851, + 0.015415429137647152, + 0.012948105111718178, + 0.014674583449959755, + 0.008139879442751408, + 0.008557705208659172, + 0.008292300626635551, + 0.020234184339642525, + 0.011765670962631702, + 0.011591027490794659, + 0.010063731111586094, + 0.00956074707210064, + 0.00684722326695919, + 0.011702444404363632, + 0.007593875285238028, + 0.02939510904252529, + 0.019155386835336685, + 0.017671214416623116, + 0.012311598286032677, + 0.018455497920513153, + 0.006914329249411821, + 0.02066265419125557, + 0.007281097117811441, + 0.01667610928416252, + 0.007501273415982723, + 0.012045213021337986, + 0.004850325640290976, + 0.003908642567694187, + 0.0024287626147270203, + 0.005016888491809368, + 0.003085739677771926, + 0.01359483227133751, + 0.014124181121587753, + 0.011441532522439957, + 0.008638783358037472, + 0.00888795405626297, + 0.005136088002473116, + 0.008712252601981163, + 0.005251097958534956, + 0.006156054325401783, + 0.0036008781753480434, + 0.0047265999019145966, + 0.0025223607663065195, + 0.001641272334381938, + 0.0013605866115540266, + 0.0021225048694759607, + 0.0011922618141397834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.22401778399944305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014246894046664238 + ], + "opposite_logprob_tokens": [ + -4.264246940612793 + ], + "logprob_reward": -0.014246894046664238, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014246894046664238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4042215347290039, + 0.08579936623573303, + 0.12932860851287842, + 0.08530088514089584, + 0.07627897709608078, + 0.061948515474796295, + 0.05359198525547981, + 0.04115535691380501, + 0.027758847922086716, + 0.03461594134569168 + ], + "credit_map": [ + 0.04819786548614502, + 0.05024075135588646, + 0.06496330350637436, + 0.03639059513807297, + 0.0290068332105875, + 0.07878890633583069, + 0.042080026119947433, + 0.05455329269170761, + 0.007132548838853836, + 0.009107360616326332, + 0.00832371972501278, + 0.012211288325488567, + 0.02044157311320305, + 0.007335295435041189, + 0.014778297394514084, + 0.006469277665019035, + 0.01672733947634697, + 0.015110990963876247, + 0.016576189547777176, + 0.011716069653630257, + 0.021864280104637146, + 0.017637722194194794, + 0.014795275405049324, + 0.014900743961334229, + 0.009562771767377853, + 0.011234228499233723, + 0.013161207549273968, + 0.0100044971331954, + 0.014284125529229641, + 0.009892897680401802, + 0.008960678242146969, + 0.008200477808713913, + 0.009532381780445576, + 0.009159551002085209, + 0.011993994005024433, + 0.009403567761182785, + 0.009982731193304062, + 0.008201614953577518, + 0.006999111268669367, + 0.011006023734807968, + 0.0105452174320817, + 0.0076670520938932896, + 0.005000385455787182, + 0.007577163632959127, + 0.009503050707280636, + 0.005721720866858959, + 0.009443286806344986, + 0.006490635685622692, + 0.010277044959366322, + 0.00930388830602169, + 0.005975492764264345, + 0.006289578042924404, + 0.00710983294993639, + 0.003916375804692507, + 0.00666285352781415, + 0.004056913778185844, + 0.009879847057163715, + 0.005312527995556593, + 0.0064979903399944305, + 0.003858025185763836, + 0.004983525723218918, + 0.0035392222926020622, + 0.004444815218448639, + 0.002639405196532607, + 0.004428445361554623, + 0.004622319713234901, + 0.004570590797811747, + 0.0034931013360619545, + 0.00327004655264318, + 0.0027433123905211687, + 0.002382574835792184, + 0.002248456235975027, + 0.010157712735235691, + 0.0036600311286747456, + 0.008472193963825703, + 0.003811580128967762, + 0.001990938326343894, + 0.0018672747537493706, + 0.002911002142354846, + 0.0017452063038945198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014246894046664238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202053382992744 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.018202053382992744, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018202053382992744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17085319757461548, + 0.15371772646903992, + 0.23624984920024872, + 0.15757478773593903, + 0.11352211982011795, + 0.050912678241729736, + 0.03862234577536583, + 0.02294008992612362, + 0.0343717597424984, + 0.02123543992638588 + ], + "credit_map": [ + 0.017750253900885582, + 0.024855490773916245, + 0.029625236988067627, + 0.018953047692775726, + 0.017476018518209457, + 0.021187402307987213, + 0.021218392997980118, + 0.01978735439479351, + 0.012468227185308933, + 0.016929425299167633, + 0.012314308434724808, + 0.020232191309332848, + 0.042949724942445755, + 0.01553226076066494, + 0.019599171355366707, + 0.013692417182028294, + 0.03403059020638466, + 0.03858032450079918, + 0.026147767901420593, + 0.03027411922812462, + 0.03612370789051056, + 0.027125248685479164, + 0.02875576913356781, + 0.01521232072263956, + 0.01678595319390297, + 0.01797567680478096, + 0.025180919095873833, + 0.01416873000562191, + 0.024052951484918594, + 0.01972951740026474, + 0.019440781325101852, + 0.020240262150764465, + 0.013741630129516125, + 0.013310391455888748, + 0.019319122657179832, + 0.015580985695123672, + 0.012255633249878883, + 0.01186563540250063, + 0.011950507760047913, + 0.01549821998924017, + 0.008089260198175907, + 0.006013576406985521, + 0.004486755467951298, + 0.005651901941746473, + 0.008671330288052559, + 0.0040519400499761105, + 0.008758580312132835, + 0.00518933217972517, + 0.0064753759652376175, + 0.006838895380496979, + 0.0044676451943814754, + 0.004013525787740946, + 0.0059337797574698925, + 0.003616778878495097, + 0.004500706680119038, + 0.0027756360359489918, + 0.005134905688464642, + 0.0030164928175508976, + 0.003550443332642317, + 0.0026408862322568893, + 0.00249724043533206, + 0.0015725125558674335, + 0.002626189263537526, + 0.0019014194840565324, + 0.009226233698427677, + 0.007355152163654566, + 0.004536885768175125, + 0.003550149267539382, + 0.00250131799839437, + 0.0022274679504334927, + 0.0026222853921353817, + 0.002352269133552909, + 0.004439578391611576, + 0.0026539387181401253, + 0.003987613134086132, + 0.002584111178293824, + 0.0018806522712111473, + 0.0015056662959977984, + 0.0028174922335892916, + 0.00136638677213341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.018202053382992744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561956942081451 + ], + "opposite_logprob_tokens": [ + -1.6006195545196533 + ], + "logprob_reward": -0.22561956942081451, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561956942081451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14523397386074066, + 0.11151085793972015, + 0.2093658149242401, + 0.06676387041807175, + 0.2273721992969513, + 0.057533685117959976, + 0.06278499960899353, + 0.04637548699975014, + 0.05055712163448334, + 0.02250191755592823 + ], + "credit_map": [ + 0.015676718205213547, + 0.01998625323176384, + 0.021512283012270927, + 0.016263458877801895, + 0.016038233414292336, + 0.016580326482653618, + 0.020662721246480942, + 0.018513988703489304, + 0.00907042995095253, + 0.014979105442762375, + 0.01046717818826437, + 0.013350991532206535, + 0.03223251923918724, + 0.011167590506374836, + 0.012862849049270153, + 0.007380195893347263, + 0.03952670842409134, + 0.028111247345805168, + 0.028817180544137955, + 0.023802807554602623, + 0.026224756613373756, + 0.021948806941509247, + 0.019706951454281807, + 0.02122737281024456, + 0.008641427382826805, + 0.005910954438149929, + 0.013736589811742306, + 0.007835057564079762, + 0.006779864430427551, + 0.009606161154806614, + 0.0058248285204172134, + 0.008428988978266716, + 0.028593759983778, + 0.027888290584087372, + 0.044474005699157715, + 0.03041563183069229, + 0.023789089173078537, + 0.02423071675002575, + 0.022174140438437462, + 0.025806577876210213, + 0.007843630388379097, + 0.006888923700898886, + 0.007266167085617781, + 0.009113510139286518, + 0.007258503697812557, + 0.004063015338033438, + 0.009476378560066223, + 0.005623556207865477, + 0.00840757042169571, + 0.007748475763946772, + 0.006585278548300266, + 0.007826845161616802, + 0.010176403447985649, + 0.005247298162430525, + 0.01184041053056717, + 0.0049527189694345, + 0.009941892698407173, + 0.007202737033367157, + 0.008057383820414543, + 0.005487657152116299, + 0.0039513856172561646, + 0.002867279574275017, + 0.005241596605628729, + 0.0036255549639463425, + 0.007768200244754553, + 0.008889122866094112, + 0.006754203699529171, + 0.005962627939879894, + 0.006282484624534845, + 0.003842288628220558, + 0.0068996199406683445, + 0.004158573225140572, + 0.003517815377563238, + 0.0035585416480898857, + 0.004163919482380152, + 0.0028943412471562624, + 0.0019073705188930035, + 0.0017284664791077375, + 0.003105051349848509, + 0.0016264114528894424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.22561956942081451 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252126008272171 + ], + "opposite_logprob_tokens": [ + -1.5021259784698486 + ], + "logprob_reward": -0.252126008272171, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252126008272171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11595385521650314, + 0.13911481201648712, + 0.1727612167596817, + 0.22420382499694824, + 0.07513787597417831, + 0.08245635777711868, + 0.06118987500667572, + 0.046763788908720016, + 0.03932928293943405, + 0.04308909550309181 + ], + "credit_map": [ + 0.013800877146422863, + 0.01230455283075571, + 0.02508985809981823, + 0.010623091831803322, + 0.013445071876049042, + 0.013018256053328514, + 0.013561278581619263, + 0.0141108687967062, + 0.029638350009918213, + 0.017406601458787918, + 0.01727929711341858, + 0.015954086557030678, + 0.016784146428108215, + 0.01778610236942768, + 0.011487530544400215, + 0.012778710573911667, + 0.027569031342864037, + 0.024220183491706848, + 0.029590871185064316, + 0.02513773739337921, + 0.016635490581393242, + 0.01832985319197178, + 0.016055110841989517, + 0.015222935006022453, + 0.0309385284781456, + 0.01748744770884514, + 0.04967333748936653, + 0.026789283379912376, + 0.021407485008239746, + 0.020150132477283478, + 0.028551584109663963, + 0.02920602075755596, + 0.010946836322546005, + 0.008751184679567814, + 0.01269559096544981, + 0.00810596439987421, + 0.008056995458900928, + 0.010282134637236595, + 0.006685316096991301, + 0.009613853879272938, + 0.01011930126696825, + 0.009371519088745117, + 0.009851407259702682, + 0.010652318596839905, + 0.012206627056002617, + 0.0089262081310153, + 0.01186872087419033, + 0.00946025364100933, + 0.00930674560368061, + 0.006926046218723059, + 0.007776505313813686, + 0.006527514196932316, + 0.009780081920325756, + 0.0045253741554915905, + 0.011725001968443394, + 0.004622613079845905, + 0.014358416199684143, + 0.011087403632700443, + 0.005907767917960882, + 0.0038708667270839214, + 0.0038242610171437263, + 0.0025003058835864067, + 0.0029174124356359243, + 0.002297351835295558, + 0.005289622116833925, + 0.00536568695679307, + 0.0044068628922104836, + 0.004271877929568291, + 0.005794229451566935, + 0.003662440227344632, + 0.006693651899695396, + 0.0038449133280664682, + 0.01083120983093977, + 0.004059642553329468, + 0.011022940278053284, + 0.005413152743130922, + 0.0020390073768794537, + 0.0031964650843292475, + 0.0037611292209476233, + 0.0027655463200062513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.252126008272171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, extreme close-up captures a scene of controlled destruction in dramatic slow motion. A thin, light-colored pine dowel, with a smooth surface and distinct grain, rests horizontally on two supporting blocks on a worn, scarred wooden workbench in a garage. Viewed from directly above, a solid, rectangular steel block drops swiftly and decisively toward the dowel's midpoint, its dark gray surface revealing a dull metallic sheen and slight wear. The entire scene is harshly illuminated by bright, direct light from an overhead fixture, casting stark, deep shadows. As the block impacts with such force, the dowel visibly bends under the immense pressure, its wood fibers straining with a sound before a catastrophic crack erupts from its center. The dowel then explodes into a cloud of sharp, jagged fragments and several larger splinters. The raw fiber texture of the freshly broken wood catches the clinical light as it is propelled outward before gracefully tumbling downward, dispersing and coming to rest on the dusty surface of the workbench, creating a focused and intense atmosphere of the physical failure of material.", + "seed": 20056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_11/index_2708_seed_20056_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": -0.08339704964309931, + "vlm_logprob": -0.08339704964309931, + "vlm_yes_no_margin": 3.0999999761581423, + "vlm_token_credit": -0.08339704964309931, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1418507695198059 + ], + "opposite_logprob_tokens": [ + -2.391850709915161 + ], + "logprob_reward": -0.1418507695198059, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1418507695198059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11092638224363327, + 0.05264398083090782, + 0.08203050494194031, + 0.08000654727220535, + 0.09524306654930115, + 0.08050759881734848, + 0.07442133873701096, + 0.39692404866218567, + 0.018362611532211304, + 0.008933981880545616 + ], + "credit_map": [ + 0.020386433228850365, + 0.022210340946912766, + 0.01153010968118906, + 0.009525350295007229, + 0.014184907078742981, + 0.017290573567152023, + 0.006171956658363342, + 0.009626707062125206, + 0.011382448486983776, + 0.019427252933382988, + 0.003503247629851103, + 0.002912584226578474, + 0.005606251768767834, + 0.003991059958934784, + 0.002319207414984703, + 0.003501929109916091, + 0.012307310476899147, + 0.008376440033316612, + 0.010003003291785717, + 0.008706173859536648, + 0.011752885766327381, + 0.008933698758482933, + 0.01272632461041212, + 0.0092246662825346, + 0.009687948040664196, + 0.011483252979815006, + 0.010847944766283035, + 0.01547760609537363, + 0.0063890558667480946, + 0.009496275335550308, + 0.00999751128256321, + 0.0066269501112401485, + 0.023624354973435402, + 0.01421472616493702, + 0.013964139856398106, + 0.007350066676735878, + 0.011155890300869942, + 0.008280938491225243, + 0.01167057454586029, + 0.004982380662113428, + 0.016518699005246162, + 0.01216453779488802, + 0.014643803238868713, + 0.007784565910696983, + 0.010547011159360409, + 0.00639820471405983, + 0.006841436959803104, + 0.005609339103102684, + 0.015467534773051739, + 0.013733979314565659, + 0.012241100892424583, + 0.00815626047551632, + 0.008086595684289932, + 0.0064031523652374744, + 0.005534073803573847, + 0.004798645619302988, + 0.12169725447893143, + 0.028819628059864044, + 0.15202432870864868, + 0.047337573021650314, + 0.007832122966647148, + 0.012962461449205875, + 0.016700735315680504, + 0.009549932554364204, + 0.0032252860255539417, + 0.005133429542183876, + 0.0021628758404403925, + 0.002707090927287936, + 0.0017207132186740637, + 0.0013735366519540548, + 0.001272844267077744, + 0.0007668354664929211, + 0.0014671673998236656, + 0.0019217332592234015, + 0.0010360423475503922, + 0.0009813070064410567, + 0.0012590017868205905, + 0.0008634731057099998, + 0.000849352334626019, + 0.000555904523935169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.1418507695198059 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00679693091660738 + ], + "opposite_logprob_tokens": [ + -5.006796836853027 + ], + "logprob_reward": -0.00679693091660738, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00679693091660738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18096992373466492, + 0.1515277475118637, + 0.14286062121391296, + 0.08976510912179947, + 0.16491757333278656, + 0.06113346293568611, + 0.04460154473781586, + 0.0818498283624649, + 0.04986020177602768, + 0.03251395747065544 + ], + "credit_map": [ + 0.028889594599604607, + 0.031207624822854996, + 0.023246683180332184, + 0.017594361677765846, + 0.023650944232940674, + 0.025890616700053215, + 0.01298463810235262, + 0.017505459487438202, + 0.022939881309866905, + 0.020123740658164024, + 0.048170965164899826, + 0.016257038339972496, + 0.013472731225192547, + 0.009638245217502117, + 0.009493870660662651, + 0.011431284248828888, + 0.027886563912034035, + 0.012241327203810215, + 0.012671142816543579, + 0.0180924441665411, + 0.026908455416560173, + 0.011421340517699718, + 0.019006283953785896, + 0.014633052051067352, + 0.013540167361497879, + 0.012361436150968075, + 0.011233663186430931, + 0.009560741484165192, + 0.013853457756340504, + 0.007538413628935814, + 0.01494043879210949, + 0.006736791227012873, + 0.037436019629240036, + 0.023895321413874626, + 0.02291860803961754, + 0.013175318017601967, + 0.018657205626368523, + 0.017186066135764122, + 0.019572552293539047, + 0.012076481245458126, + 0.013538254424929619, + 0.008482095785439014, + 0.011382766999304295, + 0.006954485550522804, + 0.007776445709168911, + 0.0042399144731462, + 0.005394658073782921, + 0.003364840056747198, + 0.008118161000311375, + 0.008698100224137306, + 0.0077085960656404495, + 0.004444890655577183, + 0.004370969254523516, + 0.003918354399502277, + 0.0039120991714298725, + 0.003430373501032591, + 0.015982486307621002, + 0.013854202814400196, + 0.01916489750146866, + 0.007171065546572208, + 0.013151083141565323, + 0.004713484086096287, + 0.00485168257728219, + 0.0029609256889671087, + 0.00776738440617919, + 0.009542753919959068, + 0.007018422242254019, + 0.007706124801188707, + 0.006296022329479456, + 0.004202036652714014, + 0.004972807131707668, + 0.002354649594053626, + 0.004383148159831762, + 0.0053736064583063126, + 0.003930245526134968, + 0.005758484359830618, + 0.0035909225698560476, + 0.0027791019529104233, + 0.004076290875673294, + 0.0026221589650958776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.00679693091660738 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211401343345642 + ], + "opposite_logprob_tokens": [ + -4.264211177825928 + ], + "logprob_reward": -0.014211401343345642, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211401343345642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17917080223560333, + 0.11168968677520752, + 0.2271907925605774, + 0.10039693862199783, + 0.09451822191476822, + 0.0276192519813776, + 0.02003081701695919, + 0.11338575929403305, + 0.10094192624092102, + 0.025055795907974243 + ], + "credit_map": [ + 0.028850411996245384, + 0.02813480980694294, + 0.020958665758371353, + 0.014875267632305622, + 0.024482183158397675, + 0.03133979067206383, + 0.012047769501805305, + 0.018481900915503502, + 0.02301814779639244, + 0.043557699769735336, + 0.01556460838764906, + 0.005496564786881208, + 0.010031400248408318, + 0.006105415523052216, + 0.00421913294121623, + 0.0036967128980904818, + 0.034987468272447586, + 0.017686637118458748, + 0.03140246495604515, + 0.016128452494740486, + 0.052438415586948395, + 0.02576437033712864, + 0.02780001610517502, + 0.02098296396434307, + 0.016304023563861847, + 0.012047713622450829, + 0.009211559779942036, + 0.009221282787621021, + 0.017859870567917824, + 0.007732962723821402, + 0.02107236161828041, + 0.006947167217731476, + 0.012205494567751884, + 0.018356241285800934, + 0.007377413101494312, + 0.011137043125927448, + 0.012733816169202328, + 0.011649668216705322, + 0.011441205628216267, + 0.009617330506443977, + 0.005677493289113045, + 0.005819413345307112, + 0.0036316963378340006, + 0.0032661096192896366, + 0.002599545056000352, + 0.0023359451442956924, + 0.0026616896502673626, + 0.0016273605870082974, + 0.003468736307695508, + 0.0048543368466198444, + 0.0016992018790915608, + 0.003068693447858095, + 0.0017916371580213308, + 0.001972004072740674, + 0.001631651888601482, + 0.0015445544850081205, + 0.026499584317207336, + 0.02088647522032261, + 0.03198304399847984, + 0.010989144444465637, + 0.009599251672625542, + 0.005763189867138863, + 0.004763331729918718, + 0.002901747589930892, + 0.01923796907067299, + 0.029881352558732033, + 0.009197751991450787, + 0.01416966412216425, + 0.010397197678685188, + 0.007391709368675947, + 0.007512089796364307, + 0.0031541844364255667, + 0.0035258587449789047, + 0.004257762338966131, + 0.003361154580488801, + 0.004591465927660465, + 0.0024507909547537565, + 0.0018178906757384539, + 0.0033830401953309774, + 0.0016678324900567532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.014211401343345642 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706170976161957 + ], + "opposite_logprob_tokens": [ + -2.127061605453491 + ], + "logprob_reward": -0.12706170976161957, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706170976161957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12985694408416748, + 0.11884753406047821, + 0.05576878786087036, + 0.05099264904856682, + 0.07615186274051666, + 0.15077458322048187, + 0.04618888348340988, + 0.17154274880886078, + 0.16670164465904236, + 0.03317436948418617 + ], + "credit_map": [ + 0.019594019278883934, + 0.0209791362285614, + 0.013984008692204952, + 0.009142148308455944, + 0.02240217663347721, + 0.02198260836303234, + 0.008686981163918972, + 0.01308586448431015, + 0.016011681407690048, + 0.034930404275655746, + 0.03611321374773979, + 0.009465971030294895, + 0.008576278574764729, + 0.004643861670047045, + 0.004764558747410774, + 0.004341568797826767, + 0.01021635439246893, + 0.007092751562595367, + 0.005070993211120367, + 0.004413745831698179, + 0.008608427830040455, + 0.006794890854507685, + 0.00792247336357832, + 0.005649147555232048, + 0.009401075541973114, + 0.008509548380970955, + 0.0029799961484968662, + 0.004565030802041292, + 0.008026200346648693, + 0.00450020981952548, + 0.009557659737765789, + 0.0034529275726526976, + 0.011525757610797882, + 0.01927982270717621, + 0.006262399721890688, + 0.0068787261843681335, + 0.008294557221233845, + 0.009827342815697193, + 0.009103046730160713, + 0.004980206955224276, + 0.03427766263484955, + 0.023903382942080498, + 0.02539907582104206, + 0.012869548052549362, + 0.02045200765132904, + 0.012757360935211182, + 0.012776301242411137, + 0.008339239284396172, + 0.009103136137127876, + 0.008523151278495789, + 0.008265262469649315, + 0.005858976393938065, + 0.0033953008241951466, + 0.004981843288987875, + 0.003198283025994897, + 0.002862931927666068, + 0.04314576834440231, + 0.025268908590078354, + 0.05353469029068947, + 0.01825077086687088, + 0.012765329331159592, + 0.0075514172203838825, + 0.007079517003148794, + 0.003946344833821058, + 0.030590802431106567, + 0.04709467664361, + 0.01610252447426319, + 0.024520423263311386, + 0.0156075619161129, + 0.013098111376166344, + 0.01372314989566803, + 0.0059644076973199844, + 0.004896209109574556, + 0.005717587657272816, + 0.0032446575351059437, + 0.004929258022457361, + 0.00422624172642827, + 0.0030795796774327755, + 0.0045972298830747604, + 0.00248360657133162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.12706170976161957 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706443667411804 + ], + "opposite_logprob_tokens": [ + -2.1270644664764404 + ], + "logprob_reward": -0.12706443667411804, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706443667411804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11741721630096436, + 0.2142362743616104, + 0.09122513234615326, + 0.0273964274674654, + 0.04747370630502701, + 0.18422554433345795, + 0.04675574600696564, + 0.20718684792518616, + 0.04980466142296791, + 0.014278403483331203 + ], + "credit_map": [ + 0.016101429238915443, + 0.02190084382891655, + 0.014177603647112846, + 0.011011444963514805, + 0.02036932110786438, + 0.01692371815443039, + 0.008090371266007423, + 0.008842486888170242, + 0.04539733752608299, + 0.051261644810438156, + 0.030668282881379128, + 0.016987167298793793, + 0.023278094828128815, + 0.016609877347946167, + 0.015244792215526104, + 0.014789070934057236, + 0.015933094546198845, + 0.005599621683359146, + 0.016259709373116493, + 0.009533542208373547, + 0.009929562918841839, + 0.0145125025883317, + 0.009851565584540367, + 0.009605534374713898, + 0.004478720482438803, + 0.003964022733271122, + 0.0024451748467981815, + 0.0023860735818743706, + 0.003978385124355555, + 0.0027810800820589066, + 0.0053584822453558445, + 0.0020044888369739056, + 0.010439609177410603, + 0.007815874181687832, + 0.006997423712164164, + 0.005157185718417168, + 0.0037129230331629515, + 0.004484608769416809, + 0.00569097651168704, + 0.0031751079950481653, + 0.039062440395355225, + 0.03403425216674805, + 0.031010475009679794, + 0.019653664901852608, + 0.016067102551460266, + 0.01487177424132824, + 0.0187007375061512, + 0.010825102217495441, + 0.007827182300388813, + 0.007526641711592674, + 0.009436405263841152, + 0.005808060057461262, + 0.003549321088939905, + 0.004358888603746891, + 0.004277207423001528, + 0.003972035367041826, + 0.056934092193841934, + 0.02601003274321556, + 0.07014941424131393, + 0.022493377327919006, + 0.010967813432216644, + 0.007479452062398195, + 0.008434937335550785, + 0.004717727657407522, + 0.006847850047051907, + 0.01189118530601263, + 0.008050545118749142, + 0.007149392273277044, + 0.004172860644757748, + 0.003678110195323825, + 0.00465629156678915, + 0.0033584260381758213, + 0.0020569495391100645, + 0.0028430295642465353, + 0.002397082047536969, + 0.002328204456716776, + 0.001330617000348866, + 0.0011609215289354324, + 0.0014429615112021565, + 0.0007186383008956909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.12706443667411804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, extreme close-up captures a scene of controlled destruction in dramatic slow motion. A thin, light-colored pine dowel, with a smooth surface and distinct grain, rests horizontally on two supporting blocks on a worn, scarred wooden workbench in a garage. Viewed from directly above, a solid, rectangular steel block drops swiftly and decisively toward the dowel's midpoint, its dark gray surface revealing a dull metallic sheen and slight wear. The entire scene is harshly illuminated by bright, direct light from an overhead fixture, casting stark, deep shadows. As the block impacts with such force, the dowel visibly bends under the immense pressure, its wood fibers straining with a sound before a catastrophic crack erupts from its center. The dowel then explodes into a cloud of sharp, jagged fragments and several larger splinters. The raw fiber texture of the freshly broken wood catches the clinical light as it is propelled outward before gracefully tumbling downward, dispersing and coming to rest on the dusty surface of the workbench, creating a focused and intense atmosphere of the physical failure of material.", + "seed": 20057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_11/index_2708_seed_20057_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": -0.09579848796129227, + "vlm_logprob": -0.09579848796129227, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.09579848796129227, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13584034144878387 + ], + "opposite_logprob_tokens": [ + -2.51084041595459 + ], + "logprob_reward": -0.13584034144878387, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.13584034144878387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14385241270065308, + 0.09628299623727798, + 0.07773066312074661, + 0.09015908092260361, + 0.17929847538471222, + 0.11146388947963715, + 0.07090222835540771, + 0.10599573701620102, + 0.05816110596060753, + 0.06615333259105682 + ], + "credit_map": [ + 0.014191552065312862, + 0.010794277302920818, + 0.008718884550035, + 0.010233056731522083, + 0.013485626317560673, + 0.02789423242211342, + 0.022868463769555092, + 0.0356663279235363, + 0.009090554900467396, + 0.01027853973209858, + 0.008508607745170593, + 0.009128748439252377, + 0.013278870843350887, + 0.01337067224085331, + 0.01751692034304142, + 0.015110088512301445, + 0.009163981303572655, + 0.0079496456310153, + 0.011005192995071411, + 0.005385586526244879, + 0.014448113739490509, + 0.010636582970619202, + 0.00962131004780531, + 0.009520254097878933, + 0.011302325874567032, + 0.009247788228094578, + 0.011363119818270206, + 0.008791087195277214, + 0.01333608478307724, + 0.010431064292788506, + 0.013921546749770641, + 0.011766073293983936, + 0.018685782328248024, + 0.029575234279036522, + 0.01493164524435997, + 0.015047836117446423, + 0.04146846383810043, + 0.025870978832244873, + 0.017953919246792793, + 0.0157646331936121, + 0.010049505159258842, + 0.018425095826387405, + 0.0076699452474713326, + 0.008558519184589386, + 0.008286170661449432, + 0.015851924195885658, + 0.009575238451361656, + 0.03304750472307205, + 0.010136867873370647, + 0.010671843774616718, + 0.00707919429987669, + 0.0069139935076236725, + 0.01294251624494791, + 0.005588901229202747, + 0.008033555932343006, + 0.009535356424748898, + 0.014519890770316124, + 0.022636063396930695, + 0.009200387634336948, + 0.012468233704566956, + 0.01689130812883377, + 0.0108347712084651, + 0.008168726228177547, + 0.011276368983089924, + 0.00845319777727127, + 0.0090857595205307, + 0.005732755176723003, + 0.0069618234410882, + 0.006897938437759876, + 0.009206525981426239, + 0.005286581814289093, + 0.0065365321934223175, + 0.00843764003366232, + 0.008861619979143143, + 0.006855030544102192, + 0.00956709310412407, + 0.00900904182344675, + 0.007607685402035713, + 0.006733895745128393, + 0.009081331081688404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.13584034144878387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020638711750507355 + ], + "opposite_logprob_tokens": [ + -3.8956387042999268 + ], + "logprob_reward": -0.020638711750507355, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020638711750507355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2415209710597992, + 0.10911448299884796, + 0.07569774240255356, + 0.07881226390600204, + 0.09814494848251343, + 0.11405614018440247, + 0.0936422049999237, + 0.08077187091112137, + 0.05929799750447273, + 0.048941366374492645 + ], + "credit_map": [ + 0.02421717718243599, + 0.03168859705328941, + 0.016188455745577812, + 0.014804409816861153, + 0.03363119810819626, + 0.03835960477590561, + 0.03423924371600151, + 0.04839230328798294, + 0.011345937848091125, + 0.012273674830794334, + 0.008621476590633392, + 0.008370415307581425, + 0.012577829882502556, + 0.016424283385276794, + 0.020996009930968285, + 0.018504861742258072, + 0.008068130351603031, + 0.006433993112295866, + 0.00702471612021327, + 0.004538516514003277, + 0.020289720967411995, + 0.010371208190917969, + 0.009944481775164604, + 0.009026974439620972, + 0.0071162753738462925, + 0.007317552343010902, + 0.011057348921895027, + 0.011708824895322323, + 0.012150856666266918, + 0.009328448213636875, + 0.008513247594237328, + 0.01161971129477024, + 0.010268401354551315, + 0.012522151693701744, + 0.011809156276285648, + 0.008864505216479301, + 0.01576353795826435, + 0.022349823266267776, + 0.006755253300070763, + 0.009812130592763424, + 0.008215677924454212, + 0.023441854864358902, + 0.028933467343449593, + 0.01616808958351612, + 0.006596767343580723, + 0.011361121200025082, + 0.007357199676334858, + 0.011981968767940998, + 0.015818459913134575, + 0.014066342264413834, + 0.010128769092261791, + 0.010240485891699791, + 0.017057595774531364, + 0.007133579812943935, + 0.009592832997441292, + 0.009604147635400295, + 0.013263328932225704, + 0.017845885828137398, + 0.005566548090428114, + 0.006041181739419699, + 0.015251856297254562, + 0.006150031927973032, + 0.007314404007047415, + 0.00933863501995802, + 0.00883298646658659, + 0.01017790287733078, + 0.004499830771237612, + 0.005736920516937971, + 0.007527467794716358, + 0.007580956909805536, + 0.006554702762514353, + 0.008387232199311256, + 0.006574651226401329, + 0.007248123176395893, + 0.004042781889438629, + 0.006004505790770054, + 0.008959791623055935, + 0.005125843919813633, + 0.005090543534606695, + 0.005895128473639488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.020638711750507355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600058138370514 + ], + "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": [ + 0.12774237990379333, + 0.1737087517976761, + 0.1617836207151413, + 0.07103206217288971, + 0.03446151688694954, + 0.23135943710803986, + 0.04757906496524811, + 0.07565153390169144, + 0.030775759369134903, + 0.04590585455298424 + ], + "credit_map": [ + 0.014173535630106926, + 0.01107086706906557, + 0.010293864645063877, + 0.008783090859651566, + 0.012256461195647717, + 0.027702614665031433, + 0.01746775582432747, + 0.025994181632995605, + 0.016076713800430298, + 0.028830043971538544, + 0.011874240823090076, + 0.01118761207908392, + 0.013581370934844017, + 0.030998164787888527, + 0.023233959451317787, + 0.03792663291096687, + 0.015290046110749245, + 0.016283627599477768, + 0.015852436423301697, + 0.011187417432665825, + 0.030977042391896248, + 0.02226255275309086, + 0.024170992895960808, + 0.0257595032453537, + 0.006808256730437279, + 0.008254016749560833, + 0.00845409743487835, + 0.009033631533384323, + 0.011355836875736713, + 0.008700259029865265, + 0.007527295965701342, + 0.010898671112954617, + 0.004222933202981949, + 0.006598127074539661, + 0.0025627687573432922, + 0.0039253244176507, + 0.003862655023112893, + 0.005294091068208218, + 0.0039097764529287815, + 0.004085839726030827, + 0.014145106077194214, + 0.04962233826518059, + 0.05261838436126709, + 0.029147790744900703, + 0.013921014033257961, + 0.024496449157595634, + 0.011789021082222462, + 0.035619329661130905, + 0.007819619961082935, + 0.006585490889847279, + 0.005598608870059252, + 0.005710211116820574, + 0.007672911509871483, + 0.0032745457720011473, + 0.005182257853448391, + 0.005735417362302542, + 0.012318853288888931, + 0.016406264156103134, + 0.005854308605194092, + 0.007866663858294487, + 0.013579889200627804, + 0.006059837061911821, + 0.005894669331610203, + 0.007671049330383539, + 0.004788055084645748, + 0.005558710545301437, + 0.0021288027055561543, + 0.0023916810750961304, + 0.005106361582875252, + 0.004525912459939718, + 0.0022784348111599684, + 0.003997799474745989, + 0.0059399981983006, + 0.007459629327058792, + 0.003697402775287628, + 0.006036409176886082, + 0.008525780402123928, + 0.005026019178330898, + 0.0038222523871809244, + 0.005398358218371868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.020600058138370514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015935182571411 + ], + "opposite_logprob_tokens": [ + -1.7015935182571411 + ], + "logprob_reward": -0.2015935182571411, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015935182571411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16049280762672424, + 0.1446746587753296, + 0.07939773797988892, + 0.11314649134874344, + 0.13799265027046204, + 0.07599018514156342, + 0.06352462619543076, + 0.1083766520023346, + 0.046207454055547714, + 0.07019678503274918 + ], + "credit_map": [ + 0.02009301260113716, + 0.017462313175201416, + 0.013634568080306053, + 0.01034526340663433, + 0.01763276569545269, + 0.03413461893796921, + 0.02740931510925293, + 0.019780948758125305, + 0.011581062339246273, + 0.023386096581816673, + 0.011742474511265755, + 0.008199460804462433, + 0.01133692916482687, + 0.023028569296002388, + 0.019083287566900253, + 0.03631676733493805, + 0.00833996944129467, + 0.008202871307730675, + 0.010858126915991306, + 0.0070320856757462025, + 0.011635462753474712, + 0.010351755656301975, + 0.010633900761604309, + 0.012343560345470905, + 0.008800949901342392, + 0.008214189670979977, + 0.0151709895581007, + 0.017539121210575104, + 0.012171430513262749, + 0.015340223908424377, + 0.016686271876096725, + 0.01922331005334854, + 0.01720043085515499, + 0.02289513312280178, + 0.014664572663605213, + 0.012470635585486889, + 0.01643633469939232, + 0.026789018884301186, + 0.010766983963549137, + 0.016769520938396454, + 0.01038163248449564, + 0.008646681904792786, + 0.011089421808719635, + 0.008551791310310364, + 0.00445948401466012, + 0.007380297407507896, + 0.006751627661287785, + 0.01872924342751503, + 0.010250955820083618, + 0.008871393278241158, + 0.005947079975157976, + 0.005634389817714691, + 0.010167106054723263, + 0.005204897373914719, + 0.007540885824710131, + 0.009907912462949753, + 0.017437772825360298, + 0.019526217132806778, + 0.006850016303360462, + 0.007448058109730482, + 0.016322949901223183, + 0.010871169157326221, + 0.011172673664987087, + 0.018747787922620773, + 0.006043813191354275, + 0.006701747886836529, + 0.004170247353613377, + 0.005222219508141279, + 0.006571939680725336, + 0.005280285608023405, + 0.006559309083968401, + 0.005657888483256102, + 0.00816474761813879, + 0.00963597185909748, + 0.007058256771415472, + 0.010693911463022232, + 0.009967443533241749, + 0.00820415373891592, + 0.007323197089135647, + 0.009149097837507725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.2015935182571411 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031981021165848 + ], + "opposite_logprob_tokens": [ + -2.3503198623657227 + ], + "logprob_reward": -0.10031981021165848, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031981021165848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13808025419712067, + 0.10888853669166565, + 0.0604463629424572, + 0.02723116986453533, + 0.278215229511261, + 0.027550291270017624, + 0.09149886667728424, + 0.161924347281456, + 0.05253182351589203, + 0.053633157163858414 + ], + "credit_map": [ + 0.014997829683125019, + 0.015414970926940441, + 0.017414242029190063, + 0.008654423989355564, + 0.01670841872692108, + 0.02666553296148777, + 0.01719869114458561, + 0.021026141941547394, + 0.010082677006721497, + 0.01281072199344635, + 0.0072284783236682415, + 0.00621307548135519, + 0.010660110041499138, + 0.01997542195022106, + 0.01077678520232439, + 0.031141266226768494, + 0.008156068623065948, + 0.00804811343550682, + 0.005910445470362902, + 0.005412121303379536, + 0.010613013990223408, + 0.006787956692278385, + 0.008244656026363373, + 0.007273987866938114, + 0.0022407809738069773, + 0.0028643545228987932, + 0.004398800898343325, + 0.0036556157283484936, + 0.003968601580709219, + 0.0032414149027317762, + 0.002448738319799304, + 0.004412864334881306, + 0.025121578946709633, + 0.04220178723335266, + 0.018501469865441322, + 0.02572479285299778, + 0.06924309581518173, + 0.03757763281464577, + 0.032997794449329376, + 0.02684706076979637, + 0.002718273550271988, + 0.003665863536298275, + 0.0031824251636862755, + 0.003036646405234933, + 0.002415855647996068, + 0.0035722502507269382, + 0.0026704082265496254, + 0.006288567557930946, + 0.014360646717250347, + 0.018357234075665474, + 0.0074435011483728886, + 0.01125210803002119, + 0.01738748326897621, + 0.007435414008796215, + 0.006882410496473312, + 0.008380064740777016, + 0.024670183658599854, + 0.02919565699994564, + 0.013691097497940063, + 0.02260025404393673, + 0.032022468745708466, + 0.011956963688135147, + 0.013814861886203289, + 0.013972860760986805, + 0.008084272034466267, + 0.008766713552176952, + 0.003419461427256465, + 0.004145162645727396, + 0.007631671614944935, + 0.008498075418174267, + 0.004506226163357496, + 0.007480239495635033, + 0.006742035038769245, + 0.008257598616182804, + 0.005311363842338324, + 0.0061567495577037334, + 0.008970212191343307, + 0.006187992636114359, + 0.00569281168282032, + 0.006314393132925034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.10031981021165848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled, extreme close-up captures a scene of controlled destruction in dramatic slow motion. A thin, light-colored pine dowel, with a smooth surface and distinct grain, rests horizontally on two supporting blocks on a worn, scarred wooden workbench in a garage. Viewed from directly above, a solid, rectangular steel block drops swiftly and decisively toward the dowel's midpoint, its dark gray surface revealing a dull metallic sheen and slight wear. The entire scene is harshly illuminated by bright, direct light from an overhead fixture, casting stark, deep shadows. As the block impacts with such force, the dowel visibly bends under the immense pressure, its wood fibers straining with a sound before a catastrophic crack erupts from its center. The dowel then explodes into a cloud of sharp, jagged fragments and several larger splinters. The raw fiber texture of the freshly broken wood catches the clinical light as it is propelled outward before gracefully tumbling downward, dispersing and coming to rest on the dusty surface of the workbench, creating a focused and intense atmosphere of the physical failure of material.", + "seed": 20058, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_11/index_2708_seed_20058_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": -0.07278237733989953, + "vlm_logprob": -0.07278237733989953, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.07278237733989953, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19147183001041412 + ], + "opposite_logprob_tokens": [ + -1.941471815109253 + ], + "logprob_reward": -0.19147183001041412, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.19147183001041412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042015016078948975, + 0.057888463139534, + 0.05750378593802452, + 0.06575743108987808, + 0.07783673703670502, + 0.11335545033216476, + 0.060955777764320374, + 0.04415246471762657, + 0.42456552386283875, + 0.055969372391700745 + ], + "credit_map": [ + 0.00521884998306632, + 0.007320151198655367, + 0.006422156002372503, + 0.005431153811514378, + 0.003534299088642001, + 0.004037969745695591, + 0.0037912968546152115, + 0.006259136833250523, + 0.009705968201160431, + 0.009092845022678375, + 0.006057830061763525, + 0.006506535690277815, + 0.0060297348536551, + 0.007019458804279566, + 0.009318552911281586, + 0.004157532472163439, + 0.005816607736051083, + 0.01151292398571968, + 0.009230634197592735, + 0.009314813651144505, + 0.006133008748292923, + 0.004559424705803394, + 0.007188090123236179, + 0.00374828209169209, + 0.01091762911528349, + 0.012531066313385963, + 0.008780598640441895, + 0.006665130145847797, + 0.00576361920684576, + 0.004662978928536177, + 0.01019486878067255, + 0.0062415460124611855, + 0.005884687416255474, + 0.011820204555988312, + 0.010416489094495773, + 0.005942551884800196, + 0.015113533474504948, + 0.013089600950479507, + 0.00880894809961319, + 0.006760725285857916, + 0.015792768448591232, + 0.028291109949350357, + 0.012434596195816994, + 0.008553879335522652, + 0.017724880948662758, + 0.013846038840711117, + 0.00885880459100008, + 0.007853370159864426, + 0.009020374156534672, + 0.0097700459882617, + 0.011873845010995865, + 0.003017015988007188, + 0.010223383083939552, + 0.008821899071335793, + 0.003390057710930705, + 0.0048391553573310375, + 0.004240730777382851, + 0.01493987999856472, + 0.0046635945327579975, + 0.002301617991179228, + 0.006452220492064953, + 0.006432092748582363, + 0.002720693126320839, + 0.002401633420959115, + 0.038157813251018524, + 0.0174667127430439, + 0.12037094682455063, + 0.015645598992705345, + 0.0739070251584053, + 0.1384323090314865, + 0.006571761332452297, + 0.014013377949595451, + 0.00692371092736721, + 0.007888066582381725, + 0.010085083544254303, + 0.002354917349293828, + 0.016849596053361893, + 0.006827878765761852, + 0.0023297418374568224, + 0.002710378495976329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.19147183001041412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00983582716435194 + ], + "opposite_logprob_tokens": [ + -4.634835720062256 + ], + "logprob_reward": -0.00983582716435194, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00983582716435194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07818548381328583, + 0.11500167846679688, + 0.06334999948740005, + 0.1628924012184143, + 0.26889780163764954, + 0.04818108677864075, + 0.07952432334423065, + 0.025603162124753, + 0.13193899393081665, + 0.026425033807754517 + ], + "credit_map": [ + 0.011391024105250835, + 0.013045812025666237, + 0.013104238547384739, + 0.007547485642135143, + 0.005369121674448252, + 0.012929674237966537, + 0.005903114099055529, + 0.008895018137991428, + 0.018978891894221306, + 0.019650394096970558, + 0.013762550428509712, + 0.009025431238114834, + 0.013141407631337643, + 0.014166968874633312, + 0.0181885939091444, + 0.008087443187832832, + 0.007700156886130571, + 0.008801409974694252, + 0.011287663131952286, + 0.01016855426132679, + 0.008196708746254444, + 0.004557806998491287, + 0.008069559931755066, + 0.0045681484043598175, + 0.028668900951743126, + 0.03601443022489548, + 0.02238854393362999, + 0.0159097108989954, + 0.013097146525979042, + 0.010277563706040382, + 0.022839674726128578, + 0.013696438632905483, + 0.026996951550245285, + 0.034338150173425674, + 0.0452752523124218, + 0.02513061836361885, + 0.037137214094400406, + 0.037628162652254105, + 0.031992845237255096, + 0.03039863519370556, + 0.006741818971931934, + 0.01081113051623106, + 0.005967959761619568, + 0.004220903385430574, + 0.007103403564542532, + 0.005406699609011412, + 0.004109624307602644, + 0.003819553181529045, + 0.011793936602771282, + 0.01061875931918621, + 0.009089398197829723, + 0.003102523274719715, + 0.01386556588113308, + 0.02314821071922779, + 0.003301111049950123, + 0.004604826681315899, + 0.002495360327884555, + 0.006664769258350134, + 0.0027949754148721695, + 0.0018529791850596666, + 0.002552897436544299, + 0.004587720613926649, + 0.0025853579863905907, + 0.0020691051613539457, + 0.013732898980379105, + 0.006059081293642521, + 0.04004298150539398, + 0.005301946308463812, + 0.018551699817180634, + 0.040762946009635925, + 0.002867182483896613, + 0.004620272200554609, + 0.0036293852608650923, + 0.0018929245416074991, + 0.007347262930124998, + 0.0023950652685016394, + 0.004566059447824955, + 0.0017276335274800658, + 0.002870015101507306, + 0.0019966887775808573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.00983582716435194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304961621761322 + ], + "opposite_logprob_tokens": [ + -3.7733049392700195 + ], + "logprob_reward": -0.023304961621761322, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304961621761322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06754673272371292, + 0.12935051321983337, + 0.08765465766191483, + 0.10920607298612595, + 0.17158542573451996, + 0.07458049058914185, + 0.17160487174987793, + 0.042534779757261276, + 0.12628448009490967, + 0.019651910290122032 + ], + "credit_map": [ + 0.009332729503512383, + 0.011795771308243275, + 0.00987349171191454, + 0.012463150545954704, + 0.005907005164772272, + 0.0061057270504534245, + 0.006180176045745611, + 0.005888689309358597, + 0.01384146697819233, + 0.01984991505742073, + 0.01692802645266056, + 0.009896180592477322, + 0.01939878985285759, + 0.016996337100863457, + 0.018766537308692932, + 0.013673294335603714, + 0.010852484963834286, + 0.014417985454201698, + 0.012143122963607311, + 0.012557233683764935, + 0.016171423718333244, + 0.006091861519962549, + 0.008723640814423561, + 0.006696911063045263, + 0.0189774502068758, + 0.02233225293457508, + 0.0148288169875741, + 0.010250736027956009, + 0.009047485888004303, + 0.00725709181278944, + 0.0166587233543396, + 0.009853520430624485, + 0.015906695276498795, + 0.021255459636449814, + 0.028638163581490517, + 0.015689261257648468, + 0.024603620171546936, + 0.02562613971531391, + 0.020832723006606102, + 0.019033387303352356, + 0.010586705058813095, + 0.012305972166359425, + 0.0097229378297925, + 0.0063005778938531876, + 0.012487214989960194, + 0.008603067137300968, + 0.008275074884295464, + 0.006298951338976622, + 0.025010844692587852, + 0.020517094060778618, + 0.017132094129920006, + 0.006004866678267717, + 0.03143928572535515, + 0.05489692464470863, + 0.006983340717852116, + 0.009620449505746365, + 0.0038762439507991076, + 0.008444852195680141, + 0.004192848224192858, + 0.0016949468990787864, + 0.005756494589149952, + 0.012891996651887894, + 0.002955998992547393, + 0.0027214018628001213, + 0.013900870457291603, + 0.008226066827774048, + 0.03539947420358658, + 0.0051162512972950935, + 0.018146753311157227, + 0.03696925565600395, + 0.003936694003641605, + 0.004589123651385307, + 0.003192167729139328, + 0.0036275270394980907, + 0.0028235733043402433, + 0.0021149953827261925, + 0.001916120876558125, + 0.0018199854530394077, + 0.0020973910577595234, + 0.002060151658952236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.023304961621761322 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129535362124443 + ], + "opposite_logprob_tokens": [ + -2.2379534244537354 + ], + "logprob_reward": -0.1129535362124443, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129535362124443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0523197278380394, + 0.09879664331674576, + 0.04443666338920593, + 0.05589870363473892, + 0.0811232402920723, + 0.0863819345831871, + 0.11504638940095901, + 0.032291874289512634, + 0.3039868175983429, + 0.12971797585487366 + ], + "credit_map": [ + 0.007996932603418827, + 0.007666615769267082, + 0.007683138828724623, + 0.00627901079133153, + 0.005055389367043972, + 0.007812433876097202, + 0.004647102206945419, + 0.005179107189178467, + 0.018562879413366318, + 0.016607616096735, + 0.01135092880576849, + 0.011362262070178986, + 0.00871992763131857, + 0.009925098158419132, + 0.012624255381524563, + 0.009643680416047573, + 0.004512496292591095, + 0.0079273022711277, + 0.005945247132331133, + 0.0071568056009709835, + 0.007043877150863409, + 0.003416330786421895, + 0.004494382534176111, + 0.003940221853554249, + 0.009203211404383183, + 0.010736525058746338, + 0.0078359330072999, + 0.004795132204890251, + 0.00460099196061492, + 0.004141437355428934, + 0.008905325084924698, + 0.005680146161466837, + 0.0061587439849972725, + 0.008229231461882591, + 0.01319529302418232, + 0.013811755925416946, + 0.008523859083652496, + 0.012604467570781708, + 0.010134893469512463, + 0.0084649957716465, + 0.010197559371590614, + 0.015130630694329739, + 0.009992134757339954, + 0.009120394475758076, + 0.014871873892843723, + 0.010790664702653885, + 0.00963078998029232, + 0.00664789043366909, + 0.014795447699725628, + 0.018555520102381706, + 0.019508354365825653, + 0.009983859024941921, + 0.013406706973910332, + 0.012504799291491508, + 0.009998197667300701, + 0.01629350148141384, + 0.004305114503949881, + 0.006978020537644625, + 0.004330482333898544, + 0.001941656693816185, + 0.00475281709805131, + 0.003577290801331401, + 0.0038090581074357033, + 0.0025974351447075605, + 0.030727986246347427, + 0.016149258241057396, + 0.08586344867944717, + 0.011894318275153637, + 0.048074446618556976, + 0.09505793452262878, + 0.005794172175228596, + 0.010425255633890629, + 0.013981967233121395, + 0.011546316556632519, + 0.032789621502161026, + 0.008055866695940495, + 0.0383293479681015, + 0.011956893838942051, + 0.007423147559165955, + 0.005634813103824854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.1129535362124443 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot as a static, slightly high-angled extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thin, light-colored pine dowel with visible grain resting horizontally on two blocks on a worn wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid, rectangular dark gray steel block drop swiftly and decisively onto the midpoint of the dowel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dowel visibly bend, crack, and explode into sharp fragments and splinters upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh, direct overhead light casting stark, deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634573169052601 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.02634573169052601, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634573169052601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057128991931676865, + 0.09514224529266357, + 0.0321255624294281, + 0.054346706718206406, + 0.07512534409761429, + 0.20729847252368927, + 0.0268455408513546, + 0.07722541689872742, + 0.24022088944911957, + 0.13454091548919678 + ], + "credit_map": [ + 0.006477701477706432, + 0.009717660956084728, + 0.007803162094205618, + 0.012315637432038784, + 0.005204558838158846, + 0.005640293471515179, + 0.004324799403548241, + 0.005645178724080324, + 0.01301688700914383, + 0.016134683042764664, + 0.015090788714587688, + 0.006733964197337627, + 0.012030445970594883, + 0.011243373155593872, + 0.010993428528308868, + 0.00989866629242897, + 0.0036304695531725883, + 0.004796671215444803, + 0.005235259886831045, + 0.006175564136356115, + 0.0039448062889277935, + 0.0021774698980152607, + 0.003353490261361003, + 0.002811828162521124, + 0.010659497231245041, + 0.011627239175140858, + 0.007435692474246025, + 0.004967390559613705, + 0.0043706139549613, + 0.0038085891865193844, + 0.007181114517152309, + 0.004296565428376198, + 0.0070353043265640736, + 0.01177883893251419, + 0.010383393615484238, + 0.013159713707864285, + 0.01250385120511055, + 0.007340297102928162, + 0.007044014986604452, + 0.0058799246326088905, + 0.029528947547078133, + 0.04509742185473442, + 0.026659948751330376, + 0.01736357994377613, + 0.028528451919555664, + 0.02702660858631134, + 0.016560394316911697, + 0.01653309538960457, + 0.004221312701702118, + 0.0034140627831220627, + 0.005397853907197714, + 0.0022574581671506166, + 0.003552340902388096, + 0.0030593599658459425, + 0.0024292240850627422, + 0.002513926476240158, + 0.0055514187552034855, + 0.03280554711818695, + 0.005644988734275103, + 0.0024664353113621473, + 0.011358455754816532, + 0.014748250134289265, + 0.0025779821444302797, + 0.002072337083518505, + 0.026447203010320663, + 0.010606938041746616, + 0.07770208269357681, + 0.009884564206004143, + 0.03329939395189285, + 0.07017698884010315, + 0.0039443764835596085, + 0.008159303106367588, + 0.015987155959010124, + 0.020621588453650475, + 0.02065928466618061, + 0.007699701003730297, + 0.0362870991230011, + 0.017413239926099777, + 0.007988039404153824, + 0.007884789258241653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.02634573169052601 + } + ], + "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/0000122/group_12/index_tensor([2956, 2956])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_12/index_tensor([2956, 2956])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..94b1c900871fc6e002bcdd84cc016e6f3c8feb99 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_12/index_tensor([2956, 2956])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In the gentle, Hayao Miyazaki animation style, the scene begins with a medium shot, immediately launching into a slow, wondrous advance toward a giant, vibrant green leaf shimmering within a lush, sun-dappled forest. Curled quietly at the leaf's center lies a tiny, fragile Partula gibba snail, its body contracted in slumber. As the camera glides closer with a quiet awe, the exquisite details of its miniature world are revealed; the snail's shell, a beautiful, translucent spiral composed of pale cream and soft amber hues, glows warmly as it catches the golden light filtering through the lofty canopy. The broad leaf it rests on is a landscape in itself, its surface dotted with perfectly spherical, crystalline dewdrops that magnify the leaf's intricate textures and refract sunlight into miniature rainbows. Against a soft-focus backdrop, the ancient forest, breathing life, rises majestically with mossy tree trunks, and the forest floor is a tapestry of soft mosses and tiny wildflowers. The camera's steady, deliberate movement toward the snail enhances the dreamlike atmosphere of tranquility, culminating in an intimate extreme close-up that fills the frame with the subtle pearlescent texture of the translucent shell, capturing a moment of profound and magical stillness.", + "seed": 20295, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_12/index_2956_seed_20295_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": -0.017672580922953783, + "vlm_logprob": -0.017672580922953783, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.017672580922953783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027946066111326218 + ], + "opposite_logprob_tokens": [ + -3.7779459953308105 + ], + "logprob_reward": -0.027946066111326218, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027946066111326218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18920333683490753, + 0.05833011120557785, + 0.024864576756954193, + 0.15296903252601624, + 0.09391859918832779, + 0.14952674508094788, + 0.16393668949604034, + 0.0613337866961956, + 0.04445195198059082, + 0.061465200036764145 + ], + "credit_map": [ + 0.023736948147416115, + 0.01616225205361843, + 0.01735449954867363, + 0.033571790903806686, + 0.016593681648373604, + 0.027148403227329254, + 0.02975974790751934, + 0.02487599477171898, + 0.006116525735706091, + 0.0055978670716285706, + 0.010602950118482113, + 0.008096028119325638, + 0.004499330185353756, + 0.008407379500567913, + 0.007283783983439207, + 0.007726239040493965, + 0.002696405164897442, + 0.003511692164465785, + 0.0029582453425973654, + 0.0031527176033705473, + 0.0029088961891829967, + 0.003189230104908347, + 0.003324171295389533, + 0.0031232181936502457, + 0.0062905242666602135, + 0.00374213675968349, + 0.00807237345725298, + 0.004659454803913832, + 0.004430434200912714, + 0.060883406549692154, + 0.0051786997355520725, + 0.059711992740631104, + 0.008456523530185223, + 0.009251104667782784, + 0.01589413359761238, + 0.01429075375199318, + 0.010168349370360374, + 0.010170431807637215, + 0.012757958844304085, + 0.012929331511259079, + 0.01709190011024475, + 0.01797381415963173, + 0.019427316263318062, + 0.027674803510308266, + 0.01244707778096199, + 0.017282187938690186, + 0.019682688638567924, + 0.017946947365999222, + 0.016973678022623062, + 0.028741272166371346, + 0.02199900522828102, + 0.021548423916101456, + 0.020894067361950874, + 0.011912154033780098, + 0.020351536571979523, + 0.021516533568501472, + 0.007517809979617596, + 0.009850549511611462, + 0.006110295653343201, + 0.0062552024610340595, + 0.009568308480083942, + 0.006946904119104147, + 0.008383565582334995, + 0.006701142992824316, + 0.005107278935611248, + 0.009082973934710026, + 0.0056183417327702045, + 0.00495954928919673, + 0.006036942824721336, + 0.004234100691974163, + 0.005580286495387554, + 0.003832475980743766, + 0.0056757996790111065, + 0.01220753975212574, + 0.005714952480047941, + 0.005429969634860754, + 0.005603987257927656, + 0.004592168144881725, + 0.015637725591659546, + 0.006603045854717493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.027946066111326218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001209838199429214 + ], + "opposite_logprob_tokens": [ + -6.751209735870361 + ], + "logprob_reward": -0.001209838199429214, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001209838199429214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3392159938812256, + 0.24484016001224518, + 0.029277939349412918, + 0.12592363357543945, + 0.02298249490559101, + 0.055092111229896545, + 0.04221666231751442, + 0.027319855988025665, + 0.016169680282473564, + 0.09696143865585327 + ], + "credit_map": [ + 0.026540739461779594, + 0.02335147187113762, + 0.029663925990462303, + 0.07040242850780487, + 0.02820477820932865, + 0.05330062657594681, + 0.07231926918029785, + 0.03543274477124214, + 0.017443181946873665, + 0.015595936216413975, + 0.03588169068098068, + 0.03626903519034386, + 0.01694251410663128, + 0.03654326871037483, + 0.04434456676244736, + 0.0418199747800827, + 0.003511974820867181, + 0.0033105730544775724, + 0.00511532835662365, + 0.00237338338047266, + 0.0027498300187289715, + 0.005007097963243723, + 0.0030106937047094107, + 0.0041990564204752445, + 0.005474897567182779, + 0.003242180682718754, + 0.006620804779231548, + 0.003603511955589056, + 0.0032411750871688128, + 0.05030660703778267, + 0.004428743850439787, + 0.04900571331381798, + 0.0034415742848068476, + 0.002111174399033189, + 0.004021446220576763, + 0.0017146756872534752, + 0.001784353400580585, + 0.003301357850432396, + 0.0030442834831774235, + 0.0035636296961456537, + 0.006481538061052561, + 0.006090459879487753, + 0.0073357257060706615, + 0.009469673037528992, + 0.004796769469976425, + 0.007197296246886253, + 0.007064249832183123, + 0.006656395271420479, + 0.005217675119638443, + 0.00780237652361393, + 0.0055502173490822315, + 0.005374271422624588, + 0.004719125106930733, + 0.0031176249030977488, + 0.004945582244545221, + 0.005489792209118605, + 0.00236846087500453, + 0.0035172891803085804, + 0.0036709504202008247, + 0.00459741847589612, + 0.002827505813911557, + 0.0026889953296631575, + 0.004694188013672829, + 0.002955049742013216, + 0.0016067484393715858, + 0.003663596697151661, + 0.0016978539060801268, + 0.0019452982814982533, + 0.002202176721766591, + 0.0013509535929188132, + 0.0020283565390855074, + 0.0016746947076171637, + 0.010571594350039959, + 0.00811462290585041, + 0.020095737650990486, + 0.007013492751866579, + 0.006968303117901087, + 0.006540968082845211, + 0.017407679930329323, + 0.020249029621481895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.001209838199429214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546654790639877 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546654790639877, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546654790639877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3732423484325409, + 0.27231788635253906, + 0.017575204372406006, + 0.09910902380943298, + 0.05413688346743584, + 0.025132164359092712, + 0.009409751743078232, + 0.02132064662873745, + 0.017104700207710266, + 0.1106514111161232 + ], + "credit_map": [ + 0.02547203190624714, + 0.021542683243751526, + 0.03500128164887428, + 0.06921480596065521, + 0.02934439480304718, + 0.05898602679371834, + 0.07251372933387756, + 0.0611673966050148, + 0.01941020041704178, + 0.01662386953830719, + 0.0396004244685173, + 0.03925246000289917, + 0.020320117473602295, + 0.0416857972741127, + 0.047905333340168, + 0.04751968011260033, + 0.0015818693209439516, + 0.0013436067383736372, + 0.0038292768876999617, + 0.0013427865924313664, + 0.0016482442151755095, + 0.004019840154796839, + 0.0011774675222113729, + 0.0026321138720959425, + 0.00506745046004653, + 0.0029072556644678116, + 0.005519375670701265, + 0.004211812745779753, + 0.005964937154203653, + 0.035499121993780136, + 0.003810563590377569, + 0.03612850233912468, + 0.00482738995924592, + 0.0054271710105240345, + 0.008450780995190144, + 0.007802390493452549, + 0.005958940833806992, + 0.0062094698660075665, + 0.008012372069060802, + 0.007448362186551094, + 0.0020775096490979195, + 0.00297381985001266, + 0.0029273100662976503, + 0.005581495352089405, + 0.002680704230442643, + 0.001950757810845971, + 0.0037155651953071356, + 0.003225001273676753, + 0.0013614424970000982, + 0.0007276146207004786, + 0.001619157032109797, + 0.000897019577678293, + 0.0011345393722876906, + 0.0010078687919303775, + 0.0011578052071854472, + 0.0015043048188090324, + 0.002546759322285652, + 0.00395696796476841, + 0.0025944411754608154, + 0.00302134919911623, + 0.0023894014302641153, + 0.002227226970717311, + 0.0025695571675896645, + 0.0020149415358901024, + 0.00169037701562047, + 0.005289542023092508, + 0.0015966480132192373, + 0.00171028939075768, + 0.002425621962174773, + 0.0010169022716581821, + 0.0017472078325226903, + 0.0016281121643260121, + 0.013406113721430302, + 0.0138516491279006, + 0.010771970264613628, + 0.010086342692375183, + 0.009470432065427303, + 0.010991129092872143, + 0.030251745134592056, + 0.011822018772363663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.012546654790639877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08819300681352615, + 0.19077594578266144, + 0.026686619967222214, + 0.5515367388725281, + 0.02439657784998417, + 0.012052739039063454, + 0.024207325652241707, + 0.043668534606695175, + 0.014265037141740322, + 0.02421741746366024 + ], + "credit_map": [ + 0.00665257079526782, + 0.005652075167745352, + 0.01040658075362444, + 0.014890545979142189, + 0.007834170013666153, + 0.01398529578000307, + 0.015431972220540047, + 0.013339811936020851, + 0.014322960749268532, + 0.011703080497682095, + 0.028211092576384544, + 0.027322715148329735, + 0.014447311870753765, + 0.028880994766950607, + 0.03309917077422142, + 0.03278864547610283, + 0.0037218763027340174, + 0.003058639820665121, + 0.0037826362531632185, + 0.002595331519842148, + 0.003375587286427617, + 0.004106165841221809, + 0.0025590478908270597, + 0.0034873380791395903, + 0.023167753592133522, + 0.015063012950122356, + 0.030978413298726082, + 0.014924384653568268, + 0.012604944407939911, + 0.22458752989768982, + 0.016897384077310562, + 0.21331335604190826, + 0.002715480048209429, + 0.0030448543839156628, + 0.004324282985180616, + 0.0026040608063340187, + 0.001986897084861994, + 0.0033510185312479734, + 0.003280892036855221, + 0.0030890940688550472, + 0.0014073450583964586, + 0.0013619797537103295, + 0.0015136144356802106, + 0.0018058867426589131, + 0.0012376559898257256, + 0.0013551271986216307, + 0.0016801847377792, + 0.0016909465193748474, + 0.0026082596741616726, + 0.0048288218677043915, + 0.0032015151809901, + 0.0031525101512670517, + 0.0028995838947594166, + 0.0020011882297694683, + 0.0024188722018152475, + 0.0030965774785727262, + 0.005300001706928015, + 0.00603881198912859, + 0.0048699877224862576, + 0.005571397021412849, + 0.005677226930856705, + 0.005420338362455368, + 0.005680556409060955, + 0.0051102181896567345, + 0.0015310889575630426, + 0.0035633426159620285, + 0.0013937478652223945, + 0.0015678995987400413, + 0.002083193277940154, + 0.0010682889260351658, + 0.001808151020668447, + 0.001249326509423554, + 0.002710111904889345, + 0.0029427669942378998, + 0.0036899782717227936, + 0.0017126239836215973, + 0.002313492586836219, + 0.002568091033026576, + 0.005528561305254698, + 0.0027517941780388355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.04304661601781845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036137294955551624 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036137294955551624, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036137294955551624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2747355103492737, + 0.10219567269086838, + 0.07071629166603088, + 0.1181141585111618, + 0.08023373782634735, + 0.12252737581729889, + 0.04817960038781166, + 0.08040899783372879, + 0.01980348490178585, + 0.08308516442775726 + ], + "credit_map": [ + 0.02171149104833603, + 0.014456638135015965, + 0.024264471605420113, + 0.047818202525377274, + 0.018156740814447403, + 0.06108653545379639, + 0.05772440508008003, + 0.02951704151928425, + 0.008774117566645145, + 0.007597692776471376, + 0.015193616040050983, + 0.013362190686166286, + 0.009287294931709766, + 0.015544638969004154, + 0.016221806406974792, + 0.016214314848184586, + 0.008993426337838173, + 0.012714644894003868, + 0.00976136140525341, + 0.004378180019557476, + 0.0064548710361123085, + 0.012370854616165161, + 0.006156813353300095, + 0.009886140003800392, + 0.00794295035302639, + 0.012386275455355644, + 0.011054814793169498, + 0.014133504591882229, + 0.00797367189079523, + 0.027385078370571136, + 0.013968135230243206, + 0.02326972223818302, + 0.006079092621803284, + 0.009198746643960476, + 0.011826496571302414, + 0.014021310955286026, + 0.009326387196779251, + 0.009853878058493137, + 0.00992151815444231, + 0.010006306692957878, + 0.016456227749586105, + 0.01276225782930851, + 0.015273666940629482, + 0.02384784072637558, + 0.010100076906383038, + 0.013773137703537941, + 0.015976764261722565, + 0.014337407425045967, + 0.008929810486733913, + 0.006979790516197681, + 0.004678138066083193, + 0.00492242444306612, + 0.005955549422651529, + 0.0047117373906075954, + 0.006226590368896723, + 0.005775561090558767, + 0.010340950451791286, + 0.015654683113098145, + 0.0059135290794074535, + 0.007454035338014364, + 0.015297400765120983, + 0.0066889626905322075, + 0.010783595032989979, + 0.008275838568806648, + 0.003161993343383074, + 0.0040313974022865295, + 0.0022718538530170918, + 0.0015186318196356297, + 0.0024293537717312574, + 0.0016682035056874156, + 0.002746598795056343, + 0.001975452061742544, + 0.009449470788240433, + 0.006645474582910538, + 0.018327517434954643, + 0.005549357272684574, + 0.005826596636325121, + 0.0058589838445186615, + 0.014465040527284145, + 0.016962723806500435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0036137294955551624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the gentle, Hayao Miyazaki animation style, the scene begins with a medium shot, immediately launching into a slow, wondrous advance toward a giant, vibrant green leaf shimmering within a lush, sun-dappled forest. Curled quietly at the leaf's center lies a tiny, fragile Partula gibba snail, its body contracted in slumber. As the camera glides closer with a quiet awe, the exquisite details of its miniature world are revealed; the snail's shell, a beautiful, translucent spiral composed of pale cream and soft amber hues, glows warmly as it catches the golden light filtering through the lofty canopy. The broad leaf it rests on is a landscape in itself, its surface dotted with perfectly spherical, crystalline dewdrops that magnify the leaf's intricate textures and refract sunlight into miniature rainbows. Against a soft-focus backdrop, the ancient forest, breathing life, rises majestically with mossy tree trunks, and the forest floor is a tapestry of soft mosses and tiny wildflowers. The camera's steady, deliberate movement toward the snail enhances the dreamlike atmosphere of tranquility, culminating in an intimate extreme close-up that fills the frame with the subtle pearlescent texture of the translucent shell, capturing a moment of profound and magical stillness.", + "seed": 20296, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_12/index_2956_seed_20296_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": -0.06523423762992024, + "vlm_logprob": -0.06523423762992024, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.06523423762992024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1030714139342308 + ], + "opposite_logprob_tokens": [ + -2.353071451187134 + ], + "logprob_reward": -0.1030714139342308, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1030714139342308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21207593381404877, + 0.04727334529161453, + 0.02341836504638195, + 0.04484301805496216, + 0.10357192158699036, + 0.3442244529724121, + 0.025131380185484886, + 0.025412634015083313, + 0.07360899448394775, + 0.10043992102146149 + ], + "credit_map": [ + 0.020436227321624756, + 0.014068135991692543, + 0.05998953431844711, + 0.011515654623508453, + 0.022052399814128876, + 0.019389772787690163, + 0.011598841287195683, + 0.05302538350224495, + 0.0041774264536798, + 0.004776735790073872, + 0.005971542093902826, + 0.005022468511015177, + 0.00477683637291193, + 0.007296945434063673, + 0.0061228228732943535, + 0.009128566831350327, + 0.0034429472871124744, + 0.0022644714917987585, + 0.0026638659182935953, + 0.0020120434928685427, + 0.0035059021320194006, + 0.004151566419750452, + 0.0018376065418124199, + 0.0035399619955569506, + 0.0063340612687170506, + 0.004197075497359037, + 0.005860579665750265, + 0.004915534518659115, + 0.006828793790191412, + 0.006791679188609123, + 0.004949320573359728, + 0.0049659782089293, + 0.013426164165139198, + 0.006839674897491932, + 0.015164039097726345, + 0.009105195291340351, + 0.015162582509219646, + 0.013654917478561401, + 0.016523662954568863, + 0.013695684261620045, + 0.02717273123562336, + 0.008189823478460312, + 0.016369977965950966, + 0.008705893531441689, + 0.03008507750928402, + 0.12125410884618759, + 0.011534675024449825, + 0.12091219425201416, + 0.0022582754027098417, + 0.0020338944159448147, + 0.002586968708783388, + 0.0034049474634230137, + 0.003029854269698262, + 0.0029364333022385836, + 0.003818394150584936, + 0.00506261270493269, + 0.005289346911013126, + 0.002933073090389371, + 0.004373714327812195, + 0.0027270223945379257, + 0.0019642668776214123, + 0.002341401530429721, + 0.0023788511753082275, + 0.003404958639293909, + 0.007642294745892286, + 0.016828766092658043, + 0.009566210210323334, + 0.0072791907005012035, + 0.008343633264303207, + 0.006880070548504591, + 0.008062629029154778, + 0.00900620874017477, + 0.008390449918806553, + 0.004865412134677172, + 0.02893487736582756, + 0.005325201898813248, + 0.005772980861365795, + 0.0347684770822525, + 0.0063672964461147785, + 0.006015230435878038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.1030714139342308 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008669831790030003 + ], + "opposite_logprob_tokens": [ + -4.758669853210449 + ], + "logprob_reward": -0.008669831790030003, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008669831790030003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3235318958759308, + 0.04933619871735573, + 0.06630269438028336, + 0.026157384738326073, + 0.20424294471740723, + 0.012934243306517601, + 0.025894319638609886, + 0.05421218276023865, + 0.08624504506587982, + 0.15114308893680573 + ], + "credit_map": [ + 0.04063979536294937, + 0.029680240899324417, + 0.025763003155589104, + 0.024957749992609024, + 0.05748807638883591, + 0.06366681307554245, + 0.027384014800190926, + 0.05395220220088959, + 0.005393681116402149, + 0.005476037040352821, + 0.0035945314448326826, + 0.004410640802234411, + 0.003949376288801432, + 0.009842022322118282, + 0.004723479971289635, + 0.011946424841880798, + 0.009942763485014439, + 0.004673105198889971, + 0.006855775136500597, + 0.004287974443286657, + 0.011631946079432964, + 0.014291628263890743, + 0.0051590208895504475, + 0.00946048367768526, + 0.002511642873287201, + 0.001615105546079576, + 0.002060616621747613, + 0.0020707498770207167, + 0.0022061944473534822, + 0.0072624823078513145, + 0.0018979663727805018, + 0.00653262622654438, + 0.02561338059604168, + 0.016068654134869576, + 0.01919623278081417, + 0.014426130801439285, + 0.04858873039484024, + 0.033191267400979996, + 0.02407361939549446, + 0.02308494783937931, + 0.001377338427118957, + 0.0009951930260285735, + 0.0017948977183550596, + 0.0010620371904224157, + 0.0014191841473802924, + 0.0023977120872586966, + 0.001386002404615283, + 0.0025018786545842886, + 0.003077323315665126, + 0.002762878779321909, + 0.003134192433208227, + 0.0025866059586405754, + 0.0026531864423304796, + 0.0043244147673249245, + 0.0033541591838002205, + 0.004001557361334562, + 0.007300728000700474, + 0.007087163161486387, + 0.008590999990701675, + 0.006720896810293198, + 0.006963977124541998, + 0.004627564921975136, + 0.0063652219250798225, + 0.006555627565830946, + 0.010532788932323456, + 0.009911938570439816, + 0.013340131379663944, + 0.009087360464036465, + 0.011137145571410656, + 0.008959319442510605, + 0.010906524024903774, + 0.012369837611913681, + 0.016185063868761063, + 0.008433620445430279, + 0.04261944442987442, + 0.008225301280617714, + 0.008387935347855091, + 0.04741017147898674, + 0.009650756604969501, + 0.010230795480310917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.008669831790030003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007637348957359791 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637348957359791, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637348957359791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10986421257257462, + 0.03224167600274086, + 0.016366124153137207, + 0.04027582332491875, + 0.08334745466709137, + 0.12775513529777527, + 0.016901271417737007, + 0.025555990636348724, + 0.0891350582242012, + 0.45855727791786194 + ], + "credit_map": [ + 0.01070456299930811, + 0.007640536408871412, + 0.014257228001952171, + 0.007099470589309931, + 0.016521645709872246, + 0.01963350735604763, + 0.0078536756336689, + 0.026153579354286194, + 0.002407743828371167, + 0.0014032720355316997, + 0.00235911482013762, + 0.0014946531737223268, + 0.002276842948049307, + 0.00919716339558363, + 0.001844621729105711, + 0.0112582603469491, + 0.0020605020690709352, + 0.0016417467268183827, + 0.0026446888223290443, + 0.0022097935434430838, + 0.0015193251892924309, + 0.001760109094902873, + 0.002211212180554867, + 0.002318745478987694, + 0.006023705005645752, + 0.003911934327334166, + 0.0051560611464083195, + 0.0038479375652968884, + 0.00682336650788784, + 0.004863472655415535, + 0.00460397033020854, + 0.00504537345841527, + 0.010792694985866547, + 0.006836636923253536, + 0.0070067415945231915, + 0.005457357969135046, + 0.020823555067181587, + 0.013627647422254086, + 0.009060878306627274, + 0.009741936810314655, + 0.009789972566068172, + 0.006019944790750742, + 0.01142494659870863, + 0.010453732684254646, + 0.009731011465191841, + 0.03433581069111824, + 0.011625745333731174, + 0.034373968839645386, + 0.001540784491226077, + 0.00172986404504627, + 0.001938957255333662, + 0.002191838575527072, + 0.0024306257255375385, + 0.002556690713390708, + 0.002297271741554141, + 0.0022152389865368605, + 0.003484954359009862, + 0.0028945838566869497, + 0.004890451207756996, + 0.0032666416373103857, + 0.0021631724666804075, + 0.002784163225442171, + 0.0026941841933876276, + 0.0033778382930904627, + 0.010920914821326733, + 0.01573827676475048, + 0.011854884214699268, + 0.007615076843649149, + 0.011168877594172955, + 0.00898660346865654, + 0.009468497708439827, + 0.013381913304328918, + 0.03383565694093704, + 0.026455404236912727, + 0.1331273764371872, + 0.022144755348563194, + 0.02736460603773594, + 0.16002316772937775, + 0.02821745164692402, + 0.02738882228732109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.007637348957359791 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154401659965515 + ], + "opposite_logprob_tokens": [ + -1.7015440464019775 + ], + "logprob_reward": -0.20154401659965515, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154401659965515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11431731283664703, + 0.08499772846698761, + 0.031101064756512642, + 0.11574205756187439, + 0.1321103274822235, + 0.09908250719308853, + 0.1603228747844696, + 0.05114210024476051, + 0.09283733367919922, + 0.11834676563739777 + ], + "credit_map": [ + 0.013682670891284943, + 0.011888602748513222, + 0.01813305728137493, + 0.009304608218371868, + 0.019694535061717033, + 0.019998887553811073, + 0.008971359580755234, + 0.01264357753098011, + 0.006258686073124409, + 0.004597107879817486, + 0.005690078251063824, + 0.004750641994178295, + 0.005094151012599468, + 0.024050842970609665, + 0.005745384842157364, + 0.028810834512114525, + 0.003185757203027606, + 0.0022525740787386894, + 0.00446890527382493, + 0.0037475142162293196, + 0.003629822749644518, + 0.006218269933015108, + 0.003101769834756851, + 0.004496448207646608, + 0.016186797991394997, + 0.011360814794898033, + 0.012851591221988201, + 0.009443933144211769, + 0.01811094768345356, + 0.021696336567401886, + 0.010134548880159855, + 0.015957074239850044, + 0.01873776689171791, + 0.009577436372637749, + 0.021621664986014366, + 0.016027692705392838, + 0.013283330015838146, + 0.013696196489036083, + 0.022304395213723183, + 0.01686183549463749, + 0.008054123260080814, + 0.004683397710323334, + 0.00642369082197547, + 0.0061489129438996315, + 0.008456717245280743, + 0.03126714006066322, + 0.005092335864901543, + 0.028956176713109016, + 0.012364190071821213, + 0.017055736854672432, + 0.014570637606084347, + 0.030347511172294617, + 0.025603055953979492, + 0.0178353413939476, + 0.026272345334291458, + 0.016274044290184975, + 0.008978504687547684, + 0.006277823355048895, + 0.009775529615581036, + 0.006412610877305269, + 0.005411252845078707, + 0.004787762183696032, + 0.004946415312588215, + 0.004552199970930815, + 0.011588096618652344, + 0.011070488020777702, + 0.013282853178679943, + 0.009502174332737923, + 0.010968565940856934, + 0.010577589273452759, + 0.01132942084223032, + 0.014518133364617825, + 0.011352173052728176, + 0.019404077902436256, + 0.014270974323153496, + 0.010800850577652454, + 0.014931516721844673, + 0.02110012248158455, + 0.011625038459897041, + 0.014862006530165672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.20154401659965515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07569181174039841, + 0.06096259504556656, + 0.022883238270878792, + 0.07069183140993118, + 0.04940416291356087, + 0.568383514881134, + 0.019562454894185066, + 0.01434137299656868, + 0.05137403681874275, + 0.06670497357845306 + ], + "credit_map": [ + 0.006778711918741465, + 0.007227669004350901, + 0.010048200376331806, + 0.006345926783978939, + 0.008157563395798206, + 0.012374277226626873, + 0.0046489546075463295, + 0.02011050470173359, + 0.005299524869769812, + 0.00604697410017252, + 0.0045470199547708035, + 0.0043013617396354675, + 0.005571043584495783, + 0.015836339443922043, + 0.004950044676661491, + 0.014410285279154778, + 0.0026769402902573347, + 0.0014767369721084833, + 0.00391501234844327, + 0.0027613455895334482, + 0.0019113316666334867, + 0.002835965948179364, + 0.0037664775736629963, + 0.0035394281148910522, + 0.008746305480599403, + 0.005933656822890043, + 0.006403051316738129, + 0.005604710895568132, + 0.01028879452496767, + 0.01574484072625637, + 0.006125925108790398, + 0.011844546534121037, + 0.008406098000705242, + 0.0043724230490624905, + 0.005462204106152058, + 0.004053213633596897, + 0.009977535344660282, + 0.005463603883981705, + 0.004412192851305008, + 0.007256888784468174, + 0.04401605203747749, + 0.014149758033454418, + 0.02557244896888733, + 0.013141629286110401, + 0.05065372958779335, + 0.2032550424337387, + 0.017220599576830864, + 0.20037426054477692, + 0.002027438022196293, + 0.00222996948286891, + 0.002848836127668619, + 0.0025669694878160954, + 0.002500313799828291, + 0.0019432433182373643, + 0.002805377123877406, + 0.002640306716784835, + 0.0020460146479308605, + 0.0015074415132403374, + 0.0020940180402249098, + 0.0018202904611825943, + 0.0012702789390459657, + 0.0013310805661603808, + 0.0018090056255459785, + 0.0024632434360682964, + 0.005642375908792019, + 0.01049257442355156, + 0.005958808120340109, + 0.005087576806545258, + 0.0069763511419296265, + 0.005406040232628584, + 0.005846200045198202, + 0.005964108742773533, + 0.006770895794034004, + 0.007494838908314705, + 0.015531699173152447, + 0.005712026730179787, + 0.005907626822590828, + 0.01357095967978239, + 0.005113037768751383, + 0.006603890564292669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.005248576868325472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the gentle, Hayao Miyazaki animation style, the scene begins with a medium shot, immediately launching into a slow, wondrous advance toward a giant, vibrant green leaf shimmering within a lush, sun-dappled forest. Curled quietly at the leaf's center lies a tiny, fragile Partula gibba snail, its body contracted in slumber. As the camera glides closer with a quiet awe, the exquisite details of its miniature world are revealed; the snail's shell, a beautiful, translucent spiral composed of pale cream and soft amber hues, glows warmly as it catches the golden light filtering through the lofty canopy. The broad leaf it rests on is a landscape in itself, its surface dotted with perfectly spherical, crystalline dewdrops that magnify the leaf's intricate textures and refract sunlight into miniature rainbows. Against a soft-focus backdrop, the ancient forest, breathing life, rises majestically with mossy tree trunks, and the forest floor is a tapestry of soft mosses and tiny wildflowers. The camera's steady, deliberate movement toward the snail enhances the dreamlike atmosphere of tranquility, culminating in an intimate extreme close-up that fills the frame with the subtle pearlescent texture of the translucent shell, capturing a moment of profound and magical stillness.", + "seed": 20297, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_12/index_2956_seed_20297_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": -0.02463445970788598, + "vlm_logprob": -0.02463445970788598, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.02463445970788598, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027935979887843132 + ], + "opposite_logprob_tokens": [ + -3.7779359817504883 + ], + "logprob_reward": -0.027935979887843132, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027935979887843132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1468755304813385, + 0.16239236295223236, + 0.058038122951984406, + 0.07748498022556305, + 0.11813219636678696, + 0.1439651995897293, + 0.11682086437940598, + 0.07143685221672058, + 0.05449247732758522, + 0.05036148428916931 + ], + "credit_map": [ + 0.019341468811035156, + 0.010150931775569916, + 0.024645201861858368, + 0.02010839805006981, + 0.022300006821751595, + 0.01642053946852684, + 0.015355060808360577, + 0.01855391263961792, + 0.021119633689522743, + 0.022335367277264595, + 0.02609771490097046, + 0.011785474605858326, + 0.01830943301320076, + 0.03032151237130165, + 0.014102539978921413, + 0.01832067035138607, + 0.010125501081347466, + 0.006102391984313726, + 0.008267981931567192, + 0.00398487551137805, + 0.005688669625669718, + 0.005460412707179785, + 0.003923015668988228, + 0.014485272578895092, + 0.010892046615481377, + 0.007866023108363152, + 0.012780143879354, + 0.009207206778228283, + 0.007072814274579287, + 0.010504290461540222, + 0.0068374029360711575, + 0.012325039133429527, + 0.010610227473080158, + 0.012737710028886795, + 0.013367990031838417, + 0.027262786403298378, + 0.01743396744132042, + 0.00894565787166357, + 0.012443955056369305, + 0.015329891815781593, + 0.014149337075650692, + 0.033498190343379974, + 0.011311114765703678, + 0.01342331524938345, + 0.015075337141752243, + 0.013290442526340485, + 0.008779974654316902, + 0.034437473863363266, + 0.009651236236095428, + 0.03969910368323326, + 0.013645253144204617, + 0.008618628606200218, + 0.019736358895897865, + 0.0069282143376767635, + 0.006714377086609602, + 0.011827686801552773, + 0.014743280597031116, + 0.0075601814314723015, + 0.005616521462798119, + 0.0083302678540349, + 0.009953459724783897, + 0.00824632216244936, + 0.005541102960705757, + 0.011445707641541958, + 0.007083924021571875, + 0.007815483957529068, + 0.009134157560765743, + 0.0060496218502521515, + 0.006499213166534901, + 0.0069585563614964485, + 0.0035634806845337152, + 0.007388038095086813, + 0.007291390094906092, + 0.007137160282582045, + 0.005861982237547636, + 0.007084656972438097, + 0.007182220928370953, + 0.005461398977786303, + 0.005848245229572058, + 0.004494423978030682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.027935979887843132 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015373327769339085 + ], + "opposite_logprob_tokens": [ + -6.501537322998047 + ], + "logprob_reward": -0.0015373327769339085, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015373327769339085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2149801403284073, + 0.32937607169151306, + 0.036834731698036194, + 0.05832308903336525, + 0.0701180100440979, + 0.11617564409971237, + 0.04215877130627632, + 0.025552082806825638, + 0.07989660650491714, + 0.02658488042652607 + ], + "credit_map": [ + 0.03098435141146183, + 0.01235007680952549, + 0.05017772316932678, + 0.020534835755825043, + 0.03312208876013756, + 0.01919933594763279, + 0.021789968013763428, + 0.026821766048669815, + 0.0389777347445488, + 0.04346618801355362, + 0.048363879323005676, + 0.019814038649201393, + 0.039267148822546005, + 0.06633603572845459, + 0.03582415357232094, + 0.0373268723487854, + 0.0067918444983661175, + 0.005501830950379372, + 0.006376252509653568, + 0.0025440407916903496, + 0.00321898702532053, + 0.004222302697598934, + 0.003231651149690151, + 0.00494781881570816, + 0.007303736638277769, + 0.0062132650054991245, + 0.010814766399562359, + 0.004707986488938332, + 0.00542472256347537, + 0.008390067145228386, + 0.004709803033620119, + 0.010758740827441216, + 0.009715791791677475, + 0.008085949346423149, + 0.009028159081935883, + 0.007485641166567802, + 0.008636794984340668, + 0.010937401093542576, + 0.00765962852165103, + 0.00856864731758833, + 0.012582753784954548, + 0.02283305861055851, + 0.009611399844288826, + 0.016778817400336266, + 0.011453844606876373, + 0.012059394270181656, + 0.008603989146649837, + 0.022252392023801804, + 0.003968239761888981, + 0.01305178739130497, + 0.002656386001035571, + 0.0022727015893906355, + 0.006376378238201141, + 0.004223354626446962, + 0.002909185364842415, + 0.0067007411271333694, + 0.0030720126815140247, + 0.00305464630946517, + 0.0024424123112112284, + 0.002797904657199979, + 0.004311898723244667, + 0.004231548402458429, + 0.0021613112185150385, + 0.003480349201709032, + 0.012022081762552261, + 0.009819017723202705, + 0.014465102925896645, + 0.009302817285060883, + 0.009409252554178238, + 0.011913812719285488, + 0.00587422214448452, + 0.007090298924595118, + 0.003908958751708269, + 0.0029079471714794636, + 0.003241682192310691, + 0.0037604826502501965, + 0.0028937901370227337, + 0.002880306448787451, + 0.004370269365608692, + 0.002621442312374711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0015373327769339085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075171530246735 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075171530246735, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075171530246735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15836568176746368, + 0.04818558320403099, + 0.03369913622736931, + 0.05974297225475311, + 0.1084042489528656, + 0.13373742997646332, + 0.15077847242355347, + 0.06749001145362854, + 0.1465587168931961, + 0.09303775429725647 + ], + "credit_map": [ + 0.016464080661535263, + 0.011705669574439526, + 0.030802961438894272, + 0.0115905636921525, + 0.023472130298614502, + 0.014030742458999157, + 0.02291259914636612, + 0.02738693170249462, + 0.004839982837438583, + 0.006324793212115765, + 0.006263758521527052, + 0.004474223125725985, + 0.006797163747251034, + 0.006410465110093355, + 0.006861218251287937, + 0.00621397839859128, + 0.005281928926706314, + 0.0033980642911046743, + 0.0048287976533174515, + 0.0038400599732995033, + 0.003432435216382146, + 0.00346752698533237, + 0.004274392034858465, + 0.005175929982215166, + 0.007803443353623152, + 0.004952576942741871, + 0.012102301232516766, + 0.00686050346121192, + 0.003754416713491082, + 0.008834456093609333, + 0.005293919704854488, + 0.010141356848180294, + 0.01664033718407154, + 0.01004735752940178, + 0.010282261297106743, + 0.017693646252155304, + 0.007979193702340126, + 0.012650661170482635, + 0.010056700557470322, + 0.0230540968477726, + 0.0120865348726511, + 0.0336248017847538, + 0.013612535782158375, + 0.011564197018742561, + 0.014286164194345474, + 0.009245460852980614, + 0.010029731318354607, + 0.02928799018263817, + 0.012685740366578102, + 0.031872984021902084, + 0.0125693678855896, + 0.013518002815544605, + 0.028800170868635178, + 0.017730195075273514, + 0.017722100019454956, + 0.0158799160271883, + 0.008463270962238312, + 0.009695363231003284, + 0.004679508972913027, + 0.005930597893893719, + 0.014011644758284092, + 0.01022681687027216, + 0.006332560442388058, + 0.008150247856974602, + 0.02014191821217537, + 0.023776115849614143, + 0.016066774725914, + 0.015485068783164024, + 0.021321814507246017, + 0.02004585973918438, + 0.012824548408389091, + 0.01689661294221878, + 0.014016871340572834, + 0.017099661752581596, + 0.007796426769345999, + 0.012646571733057499, + 0.016120869666337967, + 0.00804686825722456, + 0.010049629025161266, + 0.007260849699378014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.016075171530246735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999947875738144 + ], + "opposite_logprob_tokens": [ + -2.6949994564056396 + ], + "logprob_reward": -0.06999947875738144, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999947875738144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0646638497710228, + 0.02453264780342579, + 0.03624143451452255, + 0.13245972990989685, + 0.16187451779842377, + 0.0661439597606659, + 0.3179798722267151, + 0.02988377772271633, + 0.12820759415626526, + 0.03801266476511955 + ], + "credit_map": [ + 0.00895646307617426, + 0.0036886539310216904, + 0.010744831524789333, + 0.008357375860214233, + 0.007719690911471844, + 0.007274588104337454, + 0.010961060412228107, + 0.006961178965866566, + 0.0031106409151107073, + 0.0029086184222251177, + 0.0038259478751569986, + 0.0020350581035017967, + 0.002251791302114725, + 0.003721364540979266, + 0.002418610267341137, + 0.004260614980012178, + 0.004438863135874271, + 0.0039587742649018764, + 0.00501459464430809, + 0.0035985610447824, + 0.005559488665312529, + 0.003868928411975503, + 0.003801093203946948, + 0.006001128349453211, + 0.014292414300143719, + 0.012564495205879211, + 0.02471255324780941, + 0.01752498187124729, + 0.008855090476572514, + 0.016421999782323837, + 0.012880462221801281, + 0.02520773559808731, + 0.014004996977746487, + 0.016807524487376213, + 0.017319854348897934, + 0.0339086651802063, + 0.025785623118281364, + 0.015096088871359825, + 0.01605391874909401, + 0.02289785072207451, + 0.006725539918988943, + 0.007559328339993954, + 0.006693682633340359, + 0.014359465800225735, + 0.007714593317359686, + 0.005884657613933086, + 0.008407958783209324, + 0.008798731490969658, + 0.024714406579732895, + 0.1317165493965149, + 0.009278412908315659, + 0.009788598865270615, + 0.05682375282049179, + 0.015641268342733383, + 0.009479797445237637, + 0.06053707003593445, + 0.004036036320030689, + 0.004907998256385326, + 0.003171641146764159, + 0.004460214637219906, + 0.004451781511306763, + 0.0031016843859106302, + 0.0025971608702093363, + 0.0031572578009217978, + 0.02028796821832657, + 0.013556485064327717, + 0.025648538023233414, + 0.012381107546389103, + 0.011974207125604153, + 0.018127329647541046, + 0.0072737205773591995, + 0.018958240747451782, + 0.005471715237945318, + 0.006183011922985315, + 0.005104724783450365, + 0.004320981912314892, + 0.005831371992826462, + 0.0042617786675691605, + 0.004453551955521107, + 0.0023855273611843586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.06999947875738144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624335587024689 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624335587024689, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624335587024689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10895410180091858, + 0.0393453873693943, + 0.02222655527293682, + 0.03872852772474289, + 0.06780003756284714, + 0.26549744606018066, + 0.3839564323425293, + 0.030580395832657814, + 0.026952611282467842, + 0.015958430245518684 + ], + "credit_map": [ + 0.018889330327510834, + 0.00649639405310154, + 0.014794430695474148, + 0.012051061727106571, + 0.013641135767102242, + 0.00970615167170763, + 0.013710102066397667, + 0.019665511325001717, + 0.0063911001197993755, + 0.004714574199169874, + 0.005602851510047913, + 0.004390383139252663, + 0.005082765128463507, + 0.005931147839874029, + 0.0037461481988430023, + 0.0034864190965890884, + 0.003849722445011139, + 0.003290712833404541, + 0.0031395615078508854, + 0.002549038967117667, + 0.002314915880560875, + 0.0030830164905637503, + 0.0019463372882455587, + 0.002053252188488841, + 0.005004429258406162, + 0.004267713520675898, + 0.005250061396509409, + 0.004303935449570417, + 0.005354407709091902, + 0.005941449664533138, + 0.002777813235297799, + 0.005828720983117819, + 0.008356809616088867, + 0.005983362905681133, + 0.008657325990498066, + 0.007556022610515356, + 0.009110329672694206, + 0.013544044457376003, + 0.005805752240121365, + 0.008786393329501152, + 0.02787591516971588, + 0.046658486127853394, + 0.02292110025882721, + 0.04521455243229866, + 0.024637339636683464, + 0.02439090795814991, + 0.024793604388833046, + 0.049005575478076935, + 0.029739852994680405, + 0.16033799946308136, + 0.013420874252915382, + 0.010767646133899689, + 0.06903131306171417, + 0.018644263967871666, + 0.012733782641589642, + 0.06928074359893799, + 0.006447046063840389, + 0.004571259021759033, + 0.0028101718053221703, + 0.0029684300534427166, + 0.003996982704848051, + 0.004770799074321985, + 0.0023287504445761442, + 0.0026869599241763353, + 0.0034542204812169075, + 0.004641301464289427, + 0.003499854588881135, + 0.002680436475202441, + 0.003977303858846426, + 0.0033338083885610104, + 0.0023148031905293465, + 0.0030508856289088726, + 0.0019427721854299307, + 0.001682608388364315, + 0.0028213143814355135, + 0.0020224829204380512, + 0.001669520977884531, + 0.0017445015255361795, + 0.0024197078309953213, + 0.0016555222682654858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.007624335587024689 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In the gentle, Hayao Miyazaki animation style, the scene begins with a medium shot, immediately launching into a slow, wondrous advance toward a giant, vibrant green leaf shimmering within a lush, sun-dappled forest. Curled quietly at the leaf's center lies a tiny, fragile Partula gibba snail, its body contracted in slumber. As the camera glides closer with a quiet awe, the exquisite details of its miniature world are revealed; the snail's shell, a beautiful, translucent spiral composed of pale cream and soft amber hues, glows warmly as it catches the golden light filtering through the lofty canopy. The broad leaf it rests on is a landscape in itself, its surface dotted with perfectly spherical, crystalline dewdrops that magnify the leaf's intricate textures and refract sunlight into miniature rainbows. Against a soft-focus backdrop, the ancient forest, breathing life, rises majestically with mossy tree trunks, and the forest floor is a tapestry of soft mosses and tiny wildflowers. The camera's steady, deliberate movement toward the snail enhances the dreamlike atmosphere of tranquility, culminating in an intimate extreme close-up that fills the frame with the subtle pearlescent texture of the translucent shell, capturing a moment of profound and magical stillness.", + "seed": 20298, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_12/index_2956_seed_20298_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": -0.04213314345106482, + "vlm_logprob": -0.04213314345106482, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.04213314345106482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0821550264954567 + ], + "opposite_logprob_tokens": [ + -2.5821549892425537 + ], + "logprob_reward": -0.0821550264954567, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0821550264954567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17280428111553192, + 0.23388007283210754, + 0.10569659620523453, + 0.04731857776641846, + 0.03725191578269005, + 0.10261476039886475, + 0.18988613784313202, + 0.034968286752700806, + 0.019563116133213043, + 0.05601631850004196 + ], + "credit_map": [ + 0.014897403307259083, + 0.013236220926046371, + 0.033143430948257446, + 0.01657181791961193, + 0.01543359737843275, + 0.04207683727145195, + 0.019847052171826363, + 0.017597917467355728, + 0.0296090729534626, + 0.020037993788719177, + 0.04131392389535904, + 0.028622407466173172, + 0.017892908304929733, + 0.027371233329176903, + 0.02082381211221218, + 0.04820871725678444, + 0.00802211370319128, + 0.010663338005542755, + 0.016596125438809395, + 0.017674796283245087, + 0.01148344948887825, + 0.009597057476639748, + 0.014546838589012623, + 0.01711287535727024, + 0.005082202143967152, + 0.004058954771608114, + 0.014309296384453773, + 0.004299030173569918, + 0.003231940558180213, + 0.004635307937860489, + 0.005656717345118523, + 0.006045125424861908, + 0.008188221603631973, + 0.0034451319370418787, + 0.0055117481388151646, + 0.003990159370005131, + 0.00309567479416728, + 0.0031810549553483725, + 0.006107935681939125, + 0.0037319911643862724, + 0.025454433634877205, + 0.006325411610305309, + 0.00640382943674922, + 0.011734865605831146, + 0.0065577272325754166, + 0.0037244067061692476, + 0.033619556576013565, + 0.008794531226158142, + 0.014196020551025867, + 0.008564181625843048, + 0.07713228464126587, + 0.015313690528273582, + 0.01027029287070036, + 0.01081934105604887, + 0.012044837698340416, + 0.04154548421502113, + 0.004074883181601763, + 0.0058657145127654076, + 0.003680185880511999, + 0.004374501761049032, + 0.004230243619531393, + 0.004556106403470039, + 0.0034488297533243895, + 0.004737823270261288, + 0.0028967547696083784, + 0.0025582287926226854, + 0.002373466966673732, + 0.002312340075150132, + 0.0018411410273984075, + 0.001907783909700811, + 0.002163916826248169, + 0.003509484464302659, + 0.00558257894590497, + 0.010256116278469563, + 0.006318769417703152, + 0.005539679899811745, + 0.008290375582873821, + 0.0056077889166772366, + 0.004938212223351002, + 0.009482799097895622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.0821550264954567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028464775532484055 + ], + "opposite_logprob_tokens": [ + -5.8778462409973145 + ], + "logprob_reward": -0.0028464775532484055, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028464775532484055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13197700679302216, + 0.42697298526763916, + 0.2073618471622467, + 0.029751768335700035, + 0.05124444514513016, + 0.03727550059556961, + 0.04105060175061226, + 0.014828369952738285, + 0.011842135339975357, + 0.04769531264901161 + ], + "credit_map": [ + 0.011189299635589123, + 0.011132853105664253, + 0.02598690055310726, + 0.014817782677710056, + 0.0141359344124794, + 0.030540941283106804, + 0.014691400341689587, + 0.009481889195740223, + 0.055726952850818634, + 0.03680815175175667, + 0.07094686478376389, + 0.05202077329158783, + 0.033262815326452255, + 0.0581405907869339, + 0.03781505301594734, + 0.08225180208683014, + 0.015956878662109375, + 0.02148289605975151, + 0.0317370742559433, + 0.031001439318060875, + 0.021339206025004387, + 0.018483517691493034, + 0.0264506246894598, + 0.040910203009843826, + 0.004374274984002113, + 0.0025641871616244316, + 0.005739959888160229, + 0.0037852711975574493, + 0.0022085218224674463, + 0.00366003205999732, + 0.003396507352590561, + 0.004023014102131128, + 0.00837705098092556, + 0.005014859605580568, + 0.005384986288845539, + 0.00809426885098219, + 0.003878225339576602, + 0.0039649964310228825, + 0.010186852887272835, + 0.0063432059250772, + 0.005117760039865971, + 0.0018670994322746992, + 0.002971859183162451, + 0.007939107716083527, + 0.0036192312836647034, + 0.0019692708738148212, + 0.010096387006342411, + 0.0036947845946997404, + 0.010483584366738796, + 0.0027918496634811163, + 0.005604773294180632, + 0.0027394648641347885, + 0.0038613267242908478, + 0.0032643245067447424, + 0.003264585742726922, + 0.009040690958499908, + 0.0015989607200026512, + 0.002078697318211198, + 0.0019821077585220337, + 0.00199527177028358, + 0.0017585704335942864, + 0.001875563757494092, + 0.0016604650299996138, + 0.0018787331646308303, + 0.0013050626730546355, + 0.0020118067041039467, + 0.0016623077681288123, + 0.0013920271303504705, + 0.00157663703430444, + 0.0012520981254056096, + 0.0009247527341358364, + 0.001717442530207336, + 0.004656046163290739, + 0.007186909671872854, + 0.005879468750208616, + 0.005819936748594046, + 0.006021835375577211, + 0.005604267586022615, + 0.00492028659209609, + 0.0076065631583333015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0028464775532484055 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11118874698877335, + 0.21715746819972992, + 0.02374601736664772, + 0.02929084561765194, + 0.17509663105010986, + 0.11559467762708664, + 0.14552423357963562, + 0.07124077528715134, + 0.07686395198106766, + 0.034296680241823196 + ], + "credit_map": [ + 0.012533667497336864, + 0.01049895491451025, + 0.02026829682290554, + 0.016380125656723976, + 0.01164889708161354, + 0.020132867619395256, + 0.01043740101158619, + 0.009288537316024303, + 0.02269279956817627, + 0.015577816404402256, + 0.042096659541130066, + 0.028193127363920212, + 0.0176065806299448, + 0.026653317734599113, + 0.020489556714892387, + 0.04384761303663254, + 0.00285590928979218, + 0.003044995479285717, + 0.0039239865727722645, + 0.003281970974057913, + 0.0028445995412766933, + 0.001782499486580491, + 0.0032721604220569134, + 0.002739895833656192, + 0.004377344157546759, + 0.0026035781484097242, + 0.00524546205997467, + 0.0030449877958744764, + 0.0037437332794070244, + 0.002938860794529319, + 0.0032061105594038963, + 0.0041307685896754265, + 0.022740598767995834, + 0.012706095352768898, + 0.019659508019685745, + 0.03532121703028679, + 0.016241049394011497, + 0.010933193378150463, + 0.03916191682219505, + 0.01833305135369301, + 0.03151371702551842, + 0.006392421200871468, + 0.006650323513895273, + 0.013273201882839203, + 0.005437303334474564, + 0.00401197699829936, + 0.0380353182554245, + 0.010280419141054153, + 0.020135462284088135, + 0.008950084447860718, + 0.059244588017463684, + 0.012016416527330875, + 0.008537700399756432, + 0.00860441941767931, + 0.010577436536550522, + 0.017458131536841393, + 0.004776445683091879, + 0.012641871348023415, + 0.012221255339682102, + 0.008439169265329838, + 0.008678149431943893, + 0.006472765002399683, + 0.008737434633076191, + 0.009273688308894634, + 0.005801066290587187, + 0.010878287255764008, + 0.011984547600150108, + 0.01267238613218069, + 0.012473645620048046, + 0.005590144079178572, + 0.00668683834373951, + 0.010777035728096962, + 0.0034992843866348267, + 0.005407985765486956, + 0.00391315808519721, + 0.004323706030845642, + 0.00686920341104269, + 0.003268272615969181, + 0.0028203711844980717, + 0.004194700159132481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "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 video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291850358247757 + ], + "opposite_logprob_tokens": [ + -2.2379186153411865 + ], + "logprob_reward": -0.11291850358247757, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291850358247757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12267773598432541, + 0.12815426290035248, + 0.06559479236602783, + 0.06130295991897583, + 0.06295984238386154, + 0.0363755077123642, + 0.34740567207336426, + 0.05920341610908508, + 0.016428818926215172, + 0.09989701211452484 + ], + "credit_map": [ + 0.015697358176112175, + 0.014774478040635586, + 0.01619226299226284, + 0.020026106387376785, + 0.011492658406496048, + 0.01751035824418068, + 0.013536371290683746, + 0.01344815269112587, + 0.012429837137460709, + 0.009320195764303207, + 0.023457743227481842, + 0.017347140237689018, + 0.011660145595669746, + 0.017094099894165993, + 0.013342643156647682, + 0.023502467200160027, + 0.005640227813273668, + 0.006991252768784761, + 0.008282603695988655, + 0.00918117817491293, + 0.006924647372215986, + 0.007294897455722094, + 0.0083078034222126, + 0.012972177937626839, + 0.0053062718361616135, + 0.005725194234400988, + 0.014755257405340672, + 0.011069187894463539, + 0.005923210643231869, + 0.004195997957140207, + 0.008405747823417187, + 0.00592209305614233, + 0.01107056438922882, + 0.005916762165725231, + 0.007723663002252579, + 0.01171853020787239, + 0.004933859687298536, + 0.0038249932695180178, + 0.011157091706991196, + 0.006614370737224817, + 0.008218495175242424, + 0.002948973560705781, + 0.00469454238191247, + 0.00638629449531436, + 0.0022647043224424124, + 0.0016788322245702147, + 0.006673660594969988, + 0.003510006470605731, + 0.031442150473594666, + 0.018003182485699654, + 0.1317635327577591, + 0.027920013293623924, + 0.02998560480773449, + 0.021385150030255318, + 0.024036118760704994, + 0.06286992877721786, + 0.006060190033167601, + 0.008792119100689888, + 0.007075507193803787, + 0.006321705877780914, + 0.007677251473069191, + 0.006032287143170834, + 0.00620856462046504, + 0.011035791598260403, + 0.0015670915599912405, + 0.0024190174881368876, + 0.0024509155191481113, + 0.002318443264812231, + 0.002252620877698064, + 0.0013923017540946603, + 0.0020339167676866055, + 0.00199451157823205, + 0.009206138551235199, + 0.017223501577973366, + 0.011617192067205906, + 0.011120298877358437, + 0.013294325210154057, + 0.011498664505779743, + 0.009539109654724598, + 0.016397781670093536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.11291850358247757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a gentle Hayao Miyazaki animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a giant vibrant green leaf visible in a lush sun-dappled forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a tiny Partula gibba snail curled at the center of the leaf with a translucent spiral shell in pale cream and soft amber hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are perfectly spherical crystalline dewdrops visible on the leaf surface refracting sunlight into miniature rainbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow steady advance from a medium shot to an extreme close-up of the snail's shell?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094432573765516 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094432573765516, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094432573765516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0894995853304863, + 0.16776280105113983, + 0.06886766850948334, + 0.03876475244760513, + 0.11596434563398361, + 0.15029336512088776, + 0.2673494815826416, + 0.04016353189945221, + 0.012791785411536694, + 0.048542603850364685 + ], + "credit_map": [ + 0.006608882918953896, + 0.0047480324283242226, + 0.021378127858042717, + 0.00640829186886549, + 0.004691762384027243, + 0.027855252847075462, + 0.0075859203934669495, + 0.010223311372101307, + 0.021098127588629723, + 0.014373430982232094, + 0.024231018498539925, + 0.024604639038443565, + 0.015560231171548367, + 0.024474242702126503, + 0.017987724393606186, + 0.025433387607336044, + 0.005047355312854052, + 0.006703912280499935, + 0.011198452673852444, + 0.00936458632349968, + 0.00732828164473176, + 0.006140354089438915, + 0.008155398070812225, + 0.014929329976439476, + 0.006276432890444994, + 0.002584761241450906, + 0.010484589263796806, + 0.0034950380213558674, + 0.0025862297043204308, + 0.003618593793362379, + 0.004921521991491318, + 0.004797584842890501, + 0.01776333712041378, + 0.01690489612519741, + 0.014123614877462387, + 0.014973780140280724, + 0.010781093500554562, + 0.008965710178017616, + 0.017487118020653725, + 0.014964792877435684, + 0.025618305429816246, + 0.012063167989253998, + 0.01818131096661091, + 0.035472869873046875, + 0.020506558939814568, + 0.008392750285565853, + 0.019886519759893417, + 0.010171887464821339, + 0.03799578174948692, + 0.015042274259030819, + 0.1119581013917923, + 0.019612843170762062, + 0.013622124679386616, + 0.011577988043427467, + 0.01872994564473629, + 0.038810424506664276, + 0.005386227276176214, + 0.005102911032736301, + 0.004179190844297409, + 0.009179789572954178, + 0.003676590509712696, + 0.0033590784296393394, + 0.004837421700358391, + 0.0044423239305615425, + 0.001477727317251265, + 0.0019029552349820733, + 0.0019114805618301034, + 0.0016384704504162073, + 0.001559130265377462, + 0.0013926412211731076, + 0.0010473021538928151, + 0.001862077391706407, + 0.005280977115035057, + 0.0068867444060742855, + 0.005585774779319763, + 0.0061333635821938515, + 0.006733136251568794, + 0.0054026334546506405, + 0.005097772926092148, + 0.007422201335430145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.004094432573765516 + } + ], + "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/0000122/group_13/index_tensor([3897, 3897])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_tensor([3897, 3897])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..efe1096d3e2582c9b563758dfc6a21069cfa814a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_tensor([3897, 3897])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a display of photorealistic CGI cinematic realism, the scene opens with a low-angle medium close-up, focusing on a magnificent peacock engaging in a courtship ritual. The camera remains static, capturing the hypnotic movement in stunning detail. The peacock's massive train, perfectly fanned out, is a stunning semi-circle of iridescent feathers, each a jewel, shimmering with vibrant shades of blue, green, and bronze that shifts with the light. At the tip of each feather, the iconic eye-like pattern, or ocelli, is rendered with hyperreal clarity, its concentric rings of color seemingly glowing from within. The peacock's body, a deep blue, is deliberately tense, its entire feathers intentionally trembling; mesmerizing low-frequency vibrations ripple through the train, causing hundreds of \"eyes\" to flash in a hypnotic, wave-like motion. Nearby, a peacock, her feathers a subtle and earthy blend of speckled brown and cream, stands head bowed, observing the display with quiet and focused concentration. The camera then begins to slowly pull back, smoothly revealing the full context and grandeur of the moment. As the image widens, we see the two men in a sun-drenched forest clearing. Brilliant, warm sunlight penetrates the dense canopy of ancient trees, casting a sacred glow on the moss-covered ground, creating an atmosphere of powerful yet tranquil natural wonder.", + "seed": 26345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_3897_seed_26345_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": -0.04257800274062902, + "vlm_logprob": -0.04257800274062902, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.04257800274062902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0056894985027611256 + ], + "opposite_logprob_tokens": [ + -6.00568962097168 + ], + "logprob_reward": -0.0056894985027611256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0056894985027611256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021803125739097595, + 0.02637193538248539, + 0.011057933792471886, + 0.009317753836512566, + 0.023169714957475662, + 0.26486945152282715, + 0.5345832109451294, + 0.07298450917005539, + 0.022303253412246704, + 0.013539058156311512 + ], + "credit_map": [ + 0.0038902133237570524, + 0.002451537875458598, + 0.003218773053959012, + 0.0030803089030086994, + 0.002055458026006818, + 0.002356193959712982, + 0.0023162432480603456, + 0.0024343994446098804, + 0.004314874764531851, + 0.0027963430620729923, + 0.0037814911920577288, + 0.002405984792858362, + 0.0027464216109365225, + 0.0028927039820700884, + 0.0038460062351077795, + 0.00358811067417264, + 0.001124676549807191, + 0.000988320680335164, + 0.0012920981971547008, + 0.0009335592039860785, + 0.0017424836987629533, + 0.0016585077391937375, + 0.0024040667340159416, + 0.00091422104742378, + 0.001227305387146771, + 0.0009477134444750845, + 0.0014330324484035373, + 0.0016199531964957714, + 0.0009627221152186394, + 0.0010841533076018095, + 0.0012372072087600827, + 0.0008056675433181226, + 0.001948552904650569, + 0.0022271207999438047, + 0.004239404108375311, + 0.0025610653683543205, + 0.002878070343285799, + 0.004952848888933659, + 0.0025913098361343145, + 0.0017713445704430342, + 0.0238351970911026, + 0.008228415623307228, + 0.08382819592952728, + 0.009327957406640053, + 0.032069262117147446, + 0.0836603045463562, + 0.009514572098851204, + 0.014405573718249798, + 0.056626152247190475, + 0.01788523979485035, + 0.16127796471118927, + 0.01944238319993019, + 0.07023337483406067, + 0.15705516934394836, + 0.022454144433140755, + 0.029608864337205887, + 0.008202577941119671, + 0.0029615499079227448, + 0.021975863724946976, + 0.0026057607028633356, + 0.006819841451942921, + 0.022336246445775032, + 0.0032611871138215065, + 0.004821485839784145, + 0.0026504427660256624, + 0.0018077089916914701, + 0.0053872037678956985, + 0.0015759537927806377, + 0.0026518763042986393, + 0.004516731947660446, + 0.0022872677072882652, + 0.0014260697644203901, + 0.0022036454174667597, + 0.0017865360714495182, + 0.0032672174274921417, + 0.0012188578257337213, + 0.0010373997502028942, + 0.0011172061786055565, + 0.0017282575136050582, + 0.0011799391359090805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0056894985027611256 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001741322805173695 + ], + "opposite_logprob_tokens": [ + -6.376741409301758 + ], + "logprob_reward": -0.001741322805173695, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001741322805173695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07938360422849655, + 0.04330909252166748, + 0.014790271408855915, + 0.012198553420603275, + 0.09625726193189621, + 0.4647294580936432, + 0.11526734381914139, + 0.035397328436374664, + 0.12546363472938538, + 0.013203450478613377 + ], + "credit_map": [ + 0.01270113606005907, + 0.009198852814733982, + 0.01627553440630436, + 0.006837856490164995, + 0.015889810398221016, + 0.007090007420629263, + 0.005875276401638985, + 0.005515124648809433, + 0.007524585351347923, + 0.004420423414558172, + 0.00676700146868825, + 0.004645396955311298, + 0.005157014820724726, + 0.004135276190936565, + 0.006352403201162815, + 0.0043069892562925816, + 0.001982327550649643, + 0.0015272401506081223, + 0.0019903858192265034, + 0.0017985778395086527, + 0.001584756770171225, + 0.0011216303100809455, + 0.0031912450212985277, + 0.0015941074816510081, + 0.0011715117143467069, + 0.0012162075145170093, + 0.0016205098945647478, + 0.0018624591175466776, + 0.0011578680714592338, + 0.0011025640415027738, + 0.0027412623167037964, + 0.0013261702843010426, + 0.00727289030328393, + 0.004971779882907867, + 0.02840491756796837, + 0.004021608270704746, + 0.013363271951675415, + 0.029759488999843597, + 0.003921656869351864, + 0.004541653674095869, + 0.041517384350299835, + 0.014923419803380966, + 0.14700819551944733, + 0.017080577090382576, + 0.05672609433531761, + 0.14815972745418549, + 0.014706199988722801, + 0.024607857689261436, + 0.011886374093592167, + 0.005369070451706648, + 0.03154107183218002, + 0.005691327154636383, + 0.014728846028447151, + 0.030012818053364754, + 0.00931237917393446, + 0.00672545563429594, + 0.004007718991488218, + 0.001934607164002955, + 0.009904161095619202, + 0.0014525172300636768, + 0.0033143076580017805, + 0.010083810426294804, + 0.002089834539219737, + 0.002610368188470602, + 0.014333457686007023, + 0.007283353246748447, + 0.03556353226304054, + 0.005668878089636564, + 0.013957000337541103, + 0.03505262732505798, + 0.006898368243128061, + 0.006706406362354755, + 0.0017908585723489523, + 0.0018243276281282306, + 0.00167382275685668, + 0.0013788574142381549, + 0.0017278200248256326, + 0.001554065733216703, + 0.0019112261943519115, + 0.0013424709904938936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.001741322805173695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979080006480217 + ], + "opposite_logprob_tokens": [ + -3.5297908782958984 + ], + "logprob_reward": -0.02979080006480217, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979080006480217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13775764405727386, + 0.039961423724889755, + 0.015360782854259014, + 0.019283371046185493, + 0.11400450766086578, + 0.209334135055542, + 0.39180654287338257, + 0.030417220667004585, + 0.021736647933721542, + 0.02033773437142372 + ], + "credit_map": [ + 0.02169865556061268, + 0.01909429021179676, + 0.030701659619808197, + 0.008412974886596203, + 0.025206955149769783, + 0.01321671437472105, + 0.009738258086144924, + 0.009688140824437141, + 0.006904504727572203, + 0.00394204119220376, + 0.005889276973903179, + 0.006190708372741938, + 0.0032010837458074093, + 0.003828074550256133, + 0.006419747602194548, + 0.003585988888517022, + 0.002337824087589979, + 0.0015570430550724268, + 0.0020621749572455883, + 0.002608902519568801, + 0.0011916955700144172, + 0.0012431913055479527, + 0.0027108199428766966, + 0.001649130368605256, + 0.00411991449072957, + 0.0016720810672268271, + 0.0024396225344389677, + 0.003015726339071989, + 0.0015316299395635724, + 0.0014771637506783009, + 0.0027438802644610405, + 0.0022833531256765127, + 0.00905184168368578, + 0.007004726678133011, + 0.030234603211283684, + 0.006936666090041399, + 0.014846607111394405, + 0.0323534719645977, + 0.007372480351477861, + 0.006204107776284218, + 0.018742714077234268, + 0.007751397788524628, + 0.0659819021821022, + 0.007814429700374603, + 0.025507144629955292, + 0.06569882482290268, + 0.00691679073497653, + 0.010920926928520203, + 0.04136502742767334, + 0.013502269983291626, + 0.11712450534105301, + 0.014636236242949963, + 0.05153471976518631, + 0.11422546952962875, + 0.017843006178736687, + 0.02157529629766941, + 0.003227434353902936, + 0.001603560522198677, + 0.008218479342758656, + 0.0013864099746569991, + 0.0031206542626023293, + 0.008870933204889297, + 0.0018752500182017684, + 0.0021144982893019915, + 0.00246264785528183, + 0.002391872461885214, + 0.004205221310257912, + 0.0012578314635902643, + 0.0028773522935807705, + 0.004981741309165955, + 0.002241335576400161, + 0.001318645547144115, + 0.0031065023504197598, + 0.0023902570828795433, + 0.004273742903023958, + 0.002260726410895586, + 0.001962002832442522, + 0.0018534258706495166, + 0.0026880691293627024, + 0.0018030101200565696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.02979080006480217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865046218037605 + ], + "opposite_logprob_tokens": [ + -3.0486505031585693 + ], + "logprob_reward": -0.04865046218037605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865046218037605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006712738424539566, + 0.003658133093267679, + 0.006584748160094023, + 0.003548216773197055, + 0.009875847026705742, + 0.11404640227556229, + 0.8099471926689148, + 0.0035939489025622606, + 0.03657486289739609, + 0.0054578338749706745 + ], + "credit_map": [ + 0.001030679210089147, + 0.0005725065129809082, + 0.000922617909964174, + 0.0008332701399922371, + 0.0005838213255628943, + 0.001061891089193523, + 0.0008355200407095253, + 0.0008724323124624789, + 0.0007226686575450003, + 0.0002969565102830529, + 0.0005088894977234304, + 0.0004815966822206974, + 0.00039969093631953, + 0.000319252343615517, + 0.0005046171136200428, + 0.0004244614683557302, + 0.0006711528403684497, + 0.0006525706849060953, + 0.0008379307691939175, + 0.0006511539104394615, + 0.0010377317667007446, + 0.0008241560426540673, + 0.001373911276459694, + 0.0005361412186175585, + 0.0005318141775205731, + 0.00043129202094860375, + 0.0005105942254886031, + 0.0004993397742509842, + 0.0004052603617310524, + 0.0005064907600171864, + 0.00038773627602495253, + 0.00027568958466872573, + 0.0007815604913048446, + 0.000884914887137711, + 0.002225696574896574, + 0.0009408987243659794, + 0.0011876360513269901, + 0.0024980930611491203, + 0.0007399775204248726, + 0.0006170698907226324, + 0.010056065395474434, + 0.0038929395377635956, + 0.0355975441634655, + 0.004112476948648691, + 0.013780040666460991, + 0.036176495254039764, + 0.004227254539728165, + 0.0062035853043198586, + 0.08636149764060974, + 0.025807414203882217, + 0.24704675376415253, + 0.028042590245604515, + 0.10745269060134888, + 0.24232299625873566, + 0.027831263840198517, + 0.045082058757543564, + 0.000375077361240983, + 0.00029941866523586214, + 0.0007636736845597625, + 0.0002564314054325223, + 0.0003965540963690728, + 0.0007889938424341381, + 0.0004272728110663593, + 0.00028652718174271286, + 0.0039310031570494175, + 0.002487962367013097, + 0.009591792710125446, + 0.0021808145102113485, + 0.004294101614505053, + 0.009812073782086372, + 0.0022511505521833897, + 0.0020259693264961243, + 0.0009123529889620841, + 0.0004987849970348179, + 0.001611732062883675, + 0.0003717036743182689, + 0.0005687086959369481, + 0.000527353782672435, + 0.0003712139732670039, + 0.0005959838163107634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.04865046218037605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701793015003204 + ], + "opposite_logprob_tokens": [ + -2.1270179748535156 + ], + "logprob_reward": -0.12701793015003204, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701793015003204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012502784840762615, + 0.009374014101922512, + 0.009942583739757538, + 0.009834286756813526, + 0.026749204844236374, + 0.5196442008018494, + 0.3496203124523163, + 0.00895785540342331, + 0.03775355592370033, + 0.015621228143572807 + ], + "credit_map": [ + 0.002917151665315032, + 0.0010947687551379204, + 0.0024254927411675453, + 0.0014700752217322588, + 0.0008403242682106793, + 0.001093429746106267, + 0.001400592504069209, + 0.001260949531570077, + 0.002379327779635787, + 0.0006749257445335388, + 0.0013732908992096782, + 0.0011589467758312821, + 0.0007510829018428922, + 0.0006690627196803689, + 0.0015346325235441327, + 0.000832745514344424, + 0.0016340475995093584, + 0.0007715999381616712, + 0.0014279235620051622, + 0.001686303992755711, + 0.0009653535671532154, + 0.000819859211333096, + 0.0016609204467386007, + 0.00097657588776201, + 0.0012213807785883546, + 0.0009418496629223228, + 0.0014467433793470263, + 0.00217628781683743, + 0.0007821133476682007, + 0.0007542059756815434, + 0.00161700916942209, + 0.000894697557669133, + 0.0022569845896214247, + 0.0022102531511336565, + 0.006348360795527697, + 0.0020394364837557077, + 0.003047928214073181, + 0.007208083290606737, + 0.0014128091279417276, + 0.002225348260253668, + 0.04631664976477623, + 0.015435796231031418, + 0.1647508293390274, + 0.019332174211740494, + 0.06363507360219955, + 0.165644109249115, + 0.016554372385144234, + 0.027975155040621758, + 0.0369512215256691, + 0.012209980748593807, + 0.10461301356554031, + 0.012936783023178577, + 0.045650288462638855, + 0.10264416038990021, + 0.015307857654988766, + 0.019307022914290428, + 0.0012494090478867292, + 0.0006950470851734281, + 0.001872250926680863, + 0.0006693884497508407, + 0.0008968752808868885, + 0.0016906204400584102, + 0.0011002494720742106, + 0.0007840148755349219, + 0.004467878025025129, + 0.0017635658150538802, + 0.011223920620977879, + 0.0017166752368211746, + 0.004241835325956345, + 0.010156014934182167, + 0.0024733860045671463, + 0.0017102797282859683, + 0.002704794518649578, + 0.0019269933691248298, + 0.0032458524219691753, + 0.0012544747442007065, + 0.001690335338935256, + 0.0017114431830123067, + 0.0016607398865744472, + 0.001426595263183117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.12701793015003204 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a display of photorealistic CGI cinematic realism, the scene opens with a low-angle medium close-up, focusing on a magnificent peacock engaging in a courtship ritual. The camera remains static, capturing the hypnotic movement in stunning detail. The peacock's massive train, perfectly fanned out, is a stunning semi-circle of iridescent feathers, each a jewel, shimmering with vibrant shades of blue, green, and bronze that shifts with the light. At the tip of each feather, the iconic eye-like pattern, or ocelli, is rendered with hyperreal clarity, its concentric rings of color seemingly glowing from within. The peacock's body, a deep blue, is deliberately tense, its entire feathers intentionally trembling; mesmerizing low-frequency vibrations ripple through the train, causing hundreds of \"eyes\" to flash in a hypnotic, wave-like motion. Nearby, a peacock, her feathers a subtle and earthy blend of speckled brown and cream, stands head bowed, observing the display with quiet and focused concentration. The camera then begins to slowly pull back, smoothly revealing the full context and grandeur of the moment. As the image widens, we see the two men in a sun-drenched forest clearing. Brilliant, warm sunlight penetrates the dense canopy of ancient trees, casting a sacred glow on the moss-covered ground, creating an atmosphere of powerful yet tranquil natural wonder.", + "seed": 26346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_3897_seed_26346_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": -0.02903821775689721, + "vlm_logprob": -0.02903821775689721, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02903821775689721, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008161645382642746 + ], + "opposite_logprob_tokens": [ + -5.133161544799805 + ], + "logprob_reward": -0.008161645382642746, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008161645382642746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1761581003665924, + 0.08878811448812485, + 0.21646134555339813, + 0.07001779973506927, + 0.12157725542783737, + 0.026305172592401505, + 0.20723651349544525, + 0.05221929773688316, + 0.02497107908129692, + 0.016265328973531723 + ], + "credit_map": [ + 0.022004950791597366, + 0.026110034435987473, + 0.01842186599969864, + 0.020725209265947342, + 0.02145218662917614, + 0.034443702548742294, + 0.011042363941669464, + 0.021957771852612495, + 0.00813496857881546, + 0.007065359503030777, + 0.008712279610335827, + 0.018123995512723923, + 0.012947914190590382, + 0.02347368560731411, + 0.005550041329115629, + 0.004779865499585867, + 0.06100252643227577, + 0.014525889419019222, + 0.08218283206224442, + 0.015744103118777275, + 0.009876202791929245, + 0.016169680282473564, + 0.010470902547240257, + 0.00648919865489006, + 0.006900090258568525, + 0.010938877239823341, + 0.017931634560227394, + 0.005978452507406473, + 0.006587734911590815, + 0.007878857664763927, + 0.005511951167136431, + 0.008290200494229794, + 0.007923798635601997, + 0.025132721289992332, + 0.04111234098672867, + 0.005868698935955763, + 0.010118823498487473, + 0.013936866074800491, + 0.011382588185369968, + 0.006101406645029783, + 0.0030481386929750443, + 0.0033300737850368023, + 0.0056646522134542465, + 0.0023823664523661137, + 0.0037305292207747698, + 0.003764950903132558, + 0.002417232608422637, + 0.001967226853594184, + 0.032162487506866455, + 0.050323422998189926, + 0.02524021826684475, + 0.0164398942142725, + 0.019518492743372917, + 0.02540810965001583, + 0.015765290707349777, + 0.02237858809530735, + 0.006966682616621256, + 0.007949896156787872, + 0.006383770145475864, + 0.005442197900265455, + 0.011010240763425827, + 0.005880438257008791, + 0.005663165356963873, + 0.002922901650890708, + 0.0027543415781110525, + 0.0024278394412249327, + 0.003701187903061509, + 0.0031118327751755714, + 0.0028943901415914297, + 0.005043789744377136, + 0.0026833084411919117, + 0.002354384632781148, + 0.001833916874602437, + 0.0019408760126680136, + 0.0020046825520694256, + 0.0019431418040767312, + 0.002565849805250764, + 0.0027685530949383974, + 0.0016682471614331007, + 0.001540059456601739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.008161645382642746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009448830969631672 + ], + "opposite_logprob_tokens": [ + -7.000945091247559 + ], + "logprob_reward": -0.0009448830969631672, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009448830969631672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2067987620830536, + 0.1226167380809784, + 0.06331335008144379, + 0.04907558485865593, + 0.17722994089126587, + 0.054946012794971466, + 0.09741731733083725, + 0.1686856746673584, + 0.040370870381593704, + 0.019545800983905792 + ], + "credit_map": [ + 0.029326995834708214, + 0.03429098799824715, + 0.024204693734645844, + 0.025684893131256104, + 0.02088775672018528, + 0.03816869109869003, + 0.013168432749807835, + 0.021066315472126007, + 0.018364358693361282, + 0.010660777799785137, + 0.015847589820623398, + 0.012479190714657307, + 0.01749786175787449, + 0.030931759625673294, + 0.008900701999664307, + 0.0079345079138875, + 0.009873333387076855, + 0.006332889199256897, + 0.020659340545535088, + 0.008279448375105858, + 0.004466530866920948, + 0.005461059045046568, + 0.0042892936617136, + 0.003951455000787973, + 0.005841216072440147, + 0.008318420499563217, + 0.007096964865922928, + 0.005342261400073767, + 0.006476972252130508, + 0.005386420991271734, + 0.0041431039571762085, + 0.006470225285738707, + 0.010545887984335423, + 0.03853004425764084, + 0.059523507952690125, + 0.01020282506942749, + 0.014824140816926956, + 0.01979958452284336, + 0.015788746997714043, + 0.008015205152332783, + 0.007311677094548941, + 0.01045920792967081, + 0.011593442410230637, + 0.005340105388313532, + 0.005750064738094807, + 0.005197996273636818, + 0.00556078739464283, + 0.0037327376194298267, + 0.014593291096389294, + 0.019083214923739433, + 0.0167873352766037, + 0.007839687168598175, + 0.010401765815913677, + 0.01053521316498518, + 0.007690215017646551, + 0.010486597195267677, + 0.019182980060577393, + 0.029258662834763527, + 0.022552672773599625, + 0.018786322325468063, + 0.03316951543092728, + 0.019211607053875923, + 0.017410067841410637, + 0.00911384355276823, + 0.003759271465241909, + 0.00451684882864356, + 0.006059104111045599, + 0.004010253120213747, + 0.008360628969967365, + 0.005925151985138655, + 0.004270645324140787, + 0.0034689675085246563, + 0.002758405404165387, + 0.0023290489334613085, + 0.002226384589448571, + 0.0023990620393306017, + 0.0020430791191756725, + 0.003119345288723707, + 0.0025773742236196995, + 0.00209310045465827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0009448830969631672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493861436843872 + ], + "opposite_logprob_tokens": [ + -2.929938554763794 + ], + "logprob_reward": -0.05493861436843872, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493861436843872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07826266437768936, + 0.15345928072929382, + 0.07535633444786072, + 0.044839195907115936, + 0.09503809362649918, + 0.07208909839391708, + 0.09798626601696014, + 0.34715747833251953, + 0.021652307361364365, + 0.014159247279167175 + ], + "credit_map": [ + 0.00945967435836792, + 0.011204292066395283, + 0.007355689536780119, + 0.011058617383241653, + 0.011251688934862614, + 0.015159321017563343, + 0.004659432452172041, + 0.008113943971693516, + 0.010659299790859222, + 0.008411676622927189, + 0.02355828881263733, + 0.05044027045369148, + 0.011350365355610847, + 0.0332065224647522, + 0.00832296907901764, + 0.00750989792868495, + 0.015560180880129337, + 0.010768454521894455, + 0.020815877243876457, + 0.00696074403822422, + 0.0061148121021687984, + 0.0067487070336937904, + 0.003941921517252922, + 0.004445639904588461, + 0.0035626329481601715, + 0.004168154206126928, + 0.008605850860476494, + 0.0063938298262655735, + 0.004696758463978767, + 0.003992507699877024, + 0.00742619764059782, + 0.005993265192955732, + 0.004338128492236137, + 0.01685931906104088, + 0.028449755162000656, + 0.005985649302601814, + 0.008114499039947987, + 0.009936877526342869, + 0.009809383191168308, + 0.011544485576450825, + 0.008032658137381077, + 0.011598358862102032, + 0.009584259241819382, + 0.006305154412984848, + 0.01657281629741192, + 0.007709229830652475, + 0.007029436994343996, + 0.005257179029285908, + 0.014497912488877773, + 0.021323632448911667, + 0.012752371840178967, + 0.008167088031768799, + 0.010114319622516632, + 0.012094106525182724, + 0.008665098808705807, + 0.010371729731559753, + 0.031290654093027115, + 0.05682845786213875, + 0.040783192962408066, + 0.033318523317575455, + 0.059627339243888855, + 0.04272258281707764, + 0.027857452630996704, + 0.05472927913069725, + 0.0024990495294332504, + 0.002182386815547943, + 0.00395834119990468, + 0.0026749134995043278, + 0.0024896215181797743, + 0.0043859235011041164, + 0.0017670653760433197, + 0.0016950067365542054, + 0.0024513814132660627, + 0.0015747607685625553, + 0.001835967879742384, + 0.0017300096806138754, + 0.001818692428059876, + 0.0015531673561781645, + 0.0016301104333251715, + 0.0015651576686650515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.05493861436843872 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038092710077762604 + ], + "opposite_logprob_tokens": [ + -3.288092613220215 + ], + "logprob_reward": -0.038092710077762604, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038092710077762604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.067632757127285, + 0.3623656928539276, + 0.06995264440774918, + 0.026165347546339035, + 0.15840773284435272, + 0.06580479443073273, + 0.10825157910585403, + 0.11057620495557785, + 0.01210353709757328, + 0.018739746883511543 + ], + "credit_map": [ + 0.00773479463532567, + 0.009561286307871342, + 0.006744385696947575, + 0.010264528915286064, + 0.007235151715576649, + 0.012420820072293282, + 0.005007958505302668, + 0.008663823828101158, + 0.023418759927153587, + 0.017834525555372238, + 0.04384137690067291, + 0.12910215556621552, + 0.029570994898676872, + 0.0838966816663742, + 0.01792316697537899, + 0.016777990385890007, + 0.017906369641423225, + 0.007828262634575367, + 0.017719430848956108, + 0.006046826019883156, + 0.005723827518522739, + 0.005856527015566826, + 0.005890110973268747, + 0.0029812813736498356, + 0.0026079132221639156, + 0.003368682460859418, + 0.0051336451433598995, + 0.0030045926105231047, + 0.003185421461239457, + 0.0023078646045178175, + 0.0037591292057186365, + 0.0027980972081422806, + 0.010101930238306522, + 0.03235600143671036, + 0.05179410055279732, + 0.008837585337460041, + 0.014299985952675343, + 0.01782490499317646, + 0.014382749795913696, + 0.008810450322926044, + 0.0078616663813591, + 0.01006302610039711, + 0.009971877560019493, + 0.005575639195740223, + 0.012333396822214127, + 0.007508602924644947, + 0.005948209669440985, + 0.006542368326336145, + 0.01740647479891777, + 0.022381525486707687, + 0.017228467389941216, + 0.008727583102881908, + 0.011216532438993454, + 0.012511131353676319, + 0.007292976602911949, + 0.01148687768727541, + 0.012547009624540806, + 0.014257672242820263, + 0.011985724791884422, + 0.011082299053668976, + 0.023935675621032715, + 0.012232105247676373, + 0.013384807854890823, + 0.011150902137160301, + 0.00212414120323956, + 0.0012755137868225574, + 0.0013022837229073048, + 0.0012591208796948195, + 0.0023062280379235744, + 0.0019104619277641177, + 0.001003001700155437, + 0.0009227844420820475, + 0.0015219770139083266, + 0.0019465058576315641, + 0.002170934807509184, + 0.0028147657867521048, + 0.003259416203945875, + 0.0024223607033491135, + 0.002823632676154375, + 0.00178015255369246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.038092710077762604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305323585867882 + ], + "opposite_logprob_tokens": [ + -3.168053150177002 + ], + "logprob_reward": -0.04305323585867882, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305323585867882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08360712230205536, + 0.30239638686180115, + 0.05204318091273308, + 0.0304456427693367, + 0.035412464290857315, + 0.04557093232870102, + 0.08472386002540588, + 0.25361350178718567, + 0.08490420132875443, + 0.02728273905813694 + ], + "credit_map": [ + 0.010424277745187283, + 0.011277894489467144, + 0.00949318241328001, + 0.00962010771036148, + 0.009326406754553318, + 0.013471442274749279, + 0.006331769749522209, + 0.013662035576999187, + 0.019044438377022743, + 0.015358508564531803, + 0.042666658759117126, + 0.10957439243793488, + 0.022295329719781876, + 0.0650925263762474, + 0.01481602806597948, + 0.01354847103357315, + 0.008775459602475166, + 0.003948974888771772, + 0.017731893807649612, + 0.006745229475200176, + 0.0036713953595608473, + 0.004673024173825979, + 0.00386249553412199, + 0.0026347036473453045, + 0.0034883888438344, + 0.004474243614822626, + 0.004051699303090572, + 0.0037230788730084896, + 0.003375086933374405, + 0.0033287620171904564, + 0.0030029024928808212, + 0.005001476034522057, + 0.0036943131126463413, + 0.0052131046541035175, + 0.0077475314028561115, + 0.0027567369397729635, + 0.00403748219832778, + 0.005072733387351036, + 0.0036064654123038054, + 0.0032840941566973925, + 0.005091887898743153, + 0.0037650312297046185, + 0.007525826338678598, + 0.005806489381939173, + 0.008136754855513573, + 0.0057010166347026825, + 0.004338836297392845, + 0.005205083638429642, + 0.011434732005000114, + 0.016655299812555313, + 0.012902173213660717, + 0.008848193101584911, + 0.009734383784234524, + 0.009361155331134796, + 0.0076138353906571865, + 0.008174072951078415, + 0.028814977034926414, + 0.04338065907359123, + 0.0318840891122818, + 0.0286359004676342, + 0.05022505670785904, + 0.027849840000271797, + 0.02753121592104435, + 0.015291735529899597, + 0.011140966787934303, + 0.010207596234977245, + 0.01189790852367878, + 0.008924159221351147, + 0.01769460178911686, + 0.01170889288187027, + 0.006413368508219719, + 0.00691670086234808, + 0.0032672714442014694, + 0.0032086363062262535, + 0.004272450227290392, + 0.0032913798931986094, + 0.003042682074010372, + 0.004914833232760429, + 0.00261995499022305, + 0.002665526233613491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.04305323585867882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a display of photorealistic CGI cinematic realism, the scene opens with a low-angle medium close-up, focusing on a magnificent peacock engaging in a courtship ritual. The camera remains static, capturing the hypnotic movement in stunning detail. The peacock's massive train, perfectly fanned out, is a stunning semi-circle of iridescent feathers, each a jewel, shimmering with vibrant shades of blue, green, and bronze that shifts with the light. At the tip of each feather, the iconic eye-like pattern, or ocelli, is rendered with hyperreal clarity, its concentric rings of color seemingly glowing from within. The peacock's body, a deep blue, is deliberately tense, its entire feathers intentionally trembling; mesmerizing low-frequency vibrations ripple through the train, causing hundreds of \"eyes\" to flash in a hypnotic, wave-like motion. Nearby, a peacock, her feathers a subtle and earthy blend of speckled brown and cream, stands head bowed, observing the display with quiet and focused concentration. The camera then begins to slowly pull back, smoothly revealing the full context and grandeur of the moment. As the image widens, we see the two men in a sun-drenched forest clearing. Brilliant, warm sunlight penetrates the dense canopy of ancient trees, casting a sacred glow on the moss-covered ground, creating an atmosphere of powerful yet tranquil natural wonder.", + "seed": 26347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_3897_seed_26347_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": -0.0396261996589601, + "vlm_logprob": -0.0396261996589601, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.0396261996589601, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00915697030723095 + ], + "opposite_logprob_tokens": [ + -5.759157180786133 + ], + "logprob_reward": -0.00915697030723095, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00915697030723095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04579412192106247, + 0.03005812130868435, + 0.5263571739196777, + 0.025543296709656715, + 0.015986600890755653, + 0.03884189575910568, + 0.045681796967983246, + 0.06917519867420197, + 0.03606928512454033, + 0.1664925515651703 + ], + "credit_map": [ + 0.006724273320287466, + 0.007566622458398342, + 0.004914018791168928, + 0.0038797457236796618, + 0.007590074557811022, + 0.004191095475107431, + 0.0037835342809557915, + 0.007144750561565161, + 0.0031491280533373356, + 0.0038572438061237335, + 0.0034202905371785164, + 0.00500635476782918, + 0.006973010487854481, + 0.0024986660573631525, + 0.003150799311697483, + 0.0020026243291795254, + 0.17339001595973969, + 0.2230159491300583, + 0.019826510921120644, + 0.014499371871352196, + 0.02590765245258808, + 0.02176700532436371, + 0.025698846206068993, + 0.0222517941147089, + 0.0023945069406181574, + 0.005043709184974432, + 0.0035111100878566504, + 0.0030075625982135534, + 0.0026286777574568987, + 0.0029070330783724785, + 0.0031419754959642887, + 0.0029087194707244635, + 0.0023624212481081486, + 0.0014684622874483466, + 0.0012677639024332166, + 0.0013271748321130872, + 0.0008482330595143139, + 0.0022087341640144587, + 0.004354160279035568, + 0.002149648033082485, + 0.005191209260374308, + 0.006333223078399897, + 0.004134893883019686, + 0.005758302286267281, + 0.0035924941767007113, + 0.0035376683808863163, + 0.005835804156959057, + 0.004458297975361347, + 0.005779658909887075, + 0.004893098026514053, + 0.008567518554627895, + 0.005040932446718216, + 0.006412239279597998, + 0.004836228210479021, + 0.004763771779835224, + 0.005388344172388315, + 0.015323859639465809, + 0.010063697583973408, + 0.00644292589277029, + 0.006366811692714691, + 0.008296169340610504, + 0.012092246674001217, + 0.004430376458913088, + 0.00615910068154335, + 0.004680983256548643, + 0.004547506105154753, + 0.00408891448751092, + 0.00421455642208457, + 0.006650786381214857, + 0.0032489809673279524, + 0.0040360079146921635, + 0.004601546097546816, + 0.05296345800161362, + 0.030321264639496803, + 0.007549669593572617, + 0.006580678280442953, + 0.010129335336387157, + 0.04512520134449005, + 0.005198686383664608, + 0.008624237030744553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.00915697030723095 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001972635742276907 + ], + "opposite_logprob_tokens": [ + -6.251972675323486 + ], + "logprob_reward": -0.001972635742276907, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001972635742276907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07713194936513901, + 0.029743777588009834, + 0.4573614299297333, + 0.01799875497817993, + 0.0672205463051796, + 0.040455810725688934, + 0.1058829203248024, + 0.06430759280920029, + 0.030210765078663826, + 0.1096864864230156 + ], + "credit_map": [ + 0.012727625668048859, + 0.012444830499589443, + 0.007148042321205139, + 0.01120408158749342, + 0.007811492308974266, + 0.007057933136820793, + 0.0065210298635065556, + 0.012216905131936073, + 0.0037919373717159033, + 0.004091948736459017, + 0.004633540287613869, + 0.0037787016481161118, + 0.0030500555876642466, + 0.003208271460607648, + 0.004014728590846062, + 0.003174591576680541, + 0.15214502811431885, + 0.1918753683567047, + 0.018398363143205643, + 0.013123731128871441, + 0.023616209626197815, + 0.018439708277583122, + 0.020307866856455803, + 0.01945512741804123, + 0.0036085972096771, + 0.002103189704939723, + 0.0015990982064977288, + 0.0014163557207211852, + 0.0010761842131614685, + 0.002301410771906376, + 0.002652469789609313, + 0.003241450060158968, + 0.004079504869878292, + 0.004429090768098831, + 0.002994164824485779, + 0.0024643950164318085, + 0.0026412715669721365, + 0.004951374139636755, + 0.02352212555706501, + 0.02213861607015133, + 0.004246465861797333, + 0.009073610417544842, + 0.0040794191882014275, + 0.006908259354531765, + 0.005459569860249758, + 0.00330716953612864, + 0.0033306314144283533, + 0.004050681833177805, + 0.01463333796709776, + 0.01171248871833086, + 0.013625373132526875, + 0.01548116933554411, + 0.018527120351791382, + 0.013465847820043564, + 0.009878425858914852, + 0.008559152483940125, + 0.012501713819801807, + 0.0085831917822361, + 0.00687053008005023, + 0.006964217871427536, + 0.008874915540218353, + 0.009254349395632744, + 0.0061047873459756374, + 0.005153877194970846, + 0.004115196876227856, + 0.004366757348179817, + 0.003237262135371566, + 0.002862529596313834, + 0.0037594919558614492, + 0.0030733586754649878, + 0.0032809351105242968, + 0.005515232216566801, + 0.0321522019803524, + 0.020615383982658386, + 0.005968412850052118, + 0.005367088131606579, + 0.006367266643792391, + 0.0284572746604681, + 0.0033201577607542276, + 0.007438700180500746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.001972635742276907 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304216057062149 + ], + "opposite_logprob_tokens": [ + -3.1680421829223633 + ], + "logprob_reward": -0.04304216057062149, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304216057062149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016879038885235786, + 0.010877741500735283, + 0.5373592376708984, + 0.01269360352307558, + 0.0720234215259552, + 0.01896517723798752, + 0.028462212532758713, + 0.0383758470416069, + 0.023900005966424942, + 0.2404637336730957 + ], + "credit_map": [ + 0.0025955471210181713, + 0.0025373611133545637, + 0.0020383147057145834, + 0.0016859556781128049, + 0.0020334504079073668, + 0.00198394269682467, + 0.0014109936309978366, + 0.002593476325273514, + 0.0011708170641213655, + 0.002096883486956358, + 0.0016685229493305087, + 0.0011528043542057276, + 0.0014862261014059186, + 0.0010537273483350873, + 0.0013852925039827824, + 0.0008634685073047876, + 0.17781659960746765, + 0.22794626653194427, + 0.019842742010951042, + 0.014625491574406624, + 0.026420781388878822, + 0.02196301333606243, + 0.025802547112107277, + 0.022941790521144867, + 0.0009364159777760506, + 0.002482687821611762, + 0.0011607615742832422, + 0.0006591571145690978, + 0.0010357671417295933, + 0.0015504128532484174, + 0.00367009942419827, + 0.001198301208205521, + 0.01326113473623991, + 0.012029184959828854, + 0.00366925448179245, + 0.0026619741693139076, + 0.002334146061912179, + 0.010236731730401516, + 0.00790858082473278, + 0.019922420382499695, + 0.0017080414108932018, + 0.00312137883156538, + 0.0021598113235086203, + 0.003063674783334136, + 0.002754401648417115, + 0.0015820665284991264, + 0.0021409657783806324, + 0.002434839028865099, + 0.003017599694430828, + 0.0024154414422810078, + 0.005228530615568161, + 0.002404912607744336, + 0.0033264493104070425, + 0.0037058917805552483, + 0.002994731767103076, + 0.005368659272789955, + 0.005510537885129452, + 0.004038447048515081, + 0.0066780271008610725, + 0.0036760838702321053, + 0.004762664437294006, + 0.00491521880030632, + 0.0028326467145234346, + 0.005962228402495384, + 0.003074520267546177, + 0.003353814361616969, + 0.0022291098721325397, + 0.003314939560368657, + 0.004170290194451809, + 0.0021954381372779608, + 0.0031165480613708496, + 0.002445347374305129, + 0.07700405269861221, + 0.04445216432213783, + 0.01043248176574707, + 0.00956379808485508, + 0.014908401295542717, + 0.06549953669309616, + 0.006165563594549894, + 0.012437733821570873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.04304216057062149 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494583398103714 + ], + "opposite_logprob_tokens": [ + -2.929945945739746 + ], + "logprob_reward": -0.05494583398103714, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494583398103714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1021014079451561, + 0.015164844691753387, + 0.39552268385887146, + 0.023136069998145103, + 0.06460686773061752, + 0.02730684168636799, + 0.08156795799732208, + 0.04204360023140907, + 0.04461469128727913, + 0.20393501222133636 + ], + "credit_map": [ + 0.01258621271699667, + 0.018541624769568443, + 0.016041066497564316, + 0.0102442791685462, + 0.011667260900139809, + 0.011976187117397785, + 0.010447000153362751, + 0.010597780346870422, + 0.0018120716558769345, + 0.0024912916123867035, + 0.0016232543857768178, + 0.0016657582018524408, + 0.0023040350060909986, + 0.0017801237991079688, + 0.0019750206265598536, + 0.0015132902190089226, + 0.12814904749393463, + 0.16059835255146027, + 0.01873079128563404, + 0.013622287660837173, + 0.024076571688055992, + 0.016375957056879997, + 0.017171788960695267, + 0.01679791510105133, + 0.002177623799070716, + 0.0030512260273098946, + 0.0027200605254620314, + 0.0027412057388573885, + 0.0015549655072391033, + 0.0034264696296304464, + 0.00515207601711154, + 0.002312443917617202, + 0.008162974379956722, + 0.013326774351298809, + 0.00277117476798594, + 0.0019351798109710217, + 0.0022371856030076742, + 0.008841652423143387, + 0.021534286439418793, + 0.005797645542770624, + 0.0029219919815659523, + 0.005406217183917761, + 0.004047934431582689, + 0.004051567055284977, + 0.0031821313314139843, + 0.0026415036991238594, + 0.002293323865160346, + 0.0027621763292700052, + 0.013355779461562634, + 0.011450281366705894, + 0.012484288774430752, + 0.010172415524721146, + 0.00848717987537384, + 0.01004413329064846, + 0.010163869708776474, + 0.005410021636635065, + 0.0063049644231796265, + 0.005336370784789324, + 0.004486927296966314, + 0.004744588863104582, + 0.005667560268193483, + 0.005151563789695501, + 0.00594843365252018, + 0.004403193015605211, + 0.005361136980354786, + 0.007789865601807833, + 0.00362196983769536, + 0.005137855187058449, + 0.008024360053241253, + 0.0034122983925044537, + 0.004716687835752964, + 0.006550520192831755, + 0.06018616631627083, + 0.03582724556326866, + 0.01086721383035183, + 0.009672101587057114, + 0.013494727201759815, + 0.051537320017814636, + 0.009194394573569298, + 0.01315586268901825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.05494583398103714 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901339769363403 + ], + "opposite_logprob_tokens": [ + -2.4640133380889893 + ], + "logprob_reward": -0.08901339769363403, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901339769363403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03982501104474068, + 0.020245937630534172, + 0.07273255288600922, + 0.0287045706063509, + 0.060473132878541946, + 0.03652428090572357, + 0.030497737228870392, + 0.04277750849723816, + 0.07703866809606552, + 0.5911806225776672 + ], + "credit_map": [ + 0.006013617385178804, + 0.007098410744220018, + 0.004386194050312042, + 0.004165490157902241, + 0.004274753388017416, + 0.0040777381509542465, + 0.004663371481001377, + 0.005145433824509382, + 0.0027477035764604807, + 0.0025616148486733437, + 0.004056677687913179, + 0.0025481439661234617, + 0.0019809051882475615, + 0.001827346975915134, + 0.0022624891716986895, + 0.002261054003611207, + 0.02350361831486225, + 0.029617609456181526, + 0.0027685901150107384, + 0.002402219222858548, + 0.003563645062968135, + 0.0032386898528784513, + 0.0038713417015969753, + 0.0037668433506041765, + 0.0027503001037985086, + 0.003919546492397785, + 0.0025363340973854065, + 0.003473874181509018, + 0.0017083737766370177, + 0.004494210705161095, + 0.0076042418368160725, + 0.0022176899947226048, + 0.006963544059544802, + 0.010553108528256416, + 0.002717597410082817, + 0.0022292467765510082, + 0.0022658740635961294, + 0.0072516100481152534, + 0.006942353677004576, + 0.021549798548221588, + 0.003838291857391596, + 0.004451381973922253, + 0.005180719308555126, + 0.0059691788628697395, + 0.005846281535923481, + 0.003463238012045622, + 0.004086578264832497, + 0.0036886106245219707, + 0.0027816363144665956, + 0.0036664551589637995, + 0.004688667133450508, + 0.003244312945753336, + 0.003022297052666545, + 0.0031223511323332787, + 0.004029899835586548, + 0.005942117888480425, + 0.0064844293519854546, + 0.0071771470829844475, + 0.004483810625970364, + 0.0046622310765087605, + 0.008275755681097507, + 0.005293958354741335, + 0.003442451823502779, + 0.002957727061584592, + 0.010572953149676323, + 0.015458333306014538, + 0.006187802646309137, + 0.007123810239136219, + 0.01249928679317236, + 0.0054386514239013195, + 0.008413281291723251, + 0.0113445445895195, + 0.18180352449417114, + 0.103641577064991, + 0.029555775225162506, + 0.02683120220899582, + 0.040208540856838226, + 0.15298303961753845, + 0.023693351075053215, + 0.03246360644698143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.08901339769363403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a display of photorealistic CGI cinematic realism, the scene opens with a low-angle medium close-up, focusing on a magnificent peacock engaging in a courtship ritual. The camera remains static, capturing the hypnotic movement in stunning detail. The peacock's massive train, perfectly fanned out, is a stunning semi-circle of iridescent feathers, each a jewel, shimmering with vibrant shades of blue, green, and bronze that shifts with the light. At the tip of each feather, the iconic eye-like pattern, or ocelli, is rendered with hyperreal clarity, its concentric rings of color seemingly glowing from within. The peacock's body, a deep blue, is deliberately tense, its entire feathers intentionally trembling; mesmerizing low-frequency vibrations ripple through the train, causing hundreds of \"eyes\" to flash in a hypnotic, wave-like motion. Nearby, a peacock, her feathers a subtle and earthy blend of speckled brown and cream, stands head bowed, observing the display with quiet and focused concentration. The camera then begins to slowly pull back, smoothly revealing the full context and grandeur of the moment. As the image widens, we see the two men in a sun-drenched forest clearing. Brilliant, warm sunlight penetrates the dense canopy of ancient trees, casting a sacred glow on the moss-covered ground, creating an atmosphere of powerful yet tranquil natural wonder.", + "seed": 26348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_13/index_3897_seed_26348_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": -0.03774223106447607, + "vlm_logprob": -0.03774223106447607, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03774223106447607, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016554025933146477 + ], + "opposite_logprob_tokens": [ + -4.64155387878418 + ], + "logprob_reward": -0.016554025933146477, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.016554025933146477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.332339882850647, + 0.32154133915901184, + 0.04569978639483452, + 0.10763946920633316, + 0.01947679929435253, + 0.03052305430173874, + 0.01954018324613571, + 0.019050875678658485, + 0.03264779970049858, + 0.07154085487127304 + ], + "credit_map": [ + 0.052771154791116714, + 0.07513810694217682, + 0.02625877410173416, + 0.03219453990459442, + 0.04385939985513687, + 0.0429099015891552, + 0.03312491253018379, + 0.026083122938871384, + 0.03805515915155411, + 0.060537923127412796, + 0.01925831101834774, + 0.01636037416756153, + 0.036397796124219894, + 0.11681347340345383, + 0.017364749684929848, + 0.0167535450309515, + 0.009132030420005322, + 0.00950239785015583, + 0.0037724399007856846, + 0.006093563977628946, + 0.004500896669924259, + 0.004141455981880426, + 0.005081064533442259, + 0.0034759342670440674, + 0.027251053601503372, + 0.017822885885834694, + 0.008220027200877666, + 0.007404765114188194, + 0.009060371667146683, + 0.02630056068301201, + 0.005449435207992792, + 0.006130364257842302, + 0.0032531169708818197, + 0.002808269811794162, + 0.0017439385410398245, + 0.002484262455254793, + 0.0016771008959040046, + 0.001953649101778865, + 0.002274190541356802, + 0.0032822696957737207, + 0.004375765565782785, + 0.003938206005841494, + 0.004293160047382116, + 0.0027101628948003054, + 0.004597386810928583, + 0.0036093236412853003, + 0.0034638948272913694, + 0.003535154042765498, + 0.002456363756209612, + 0.002496008761227131, + 0.0016266689635813236, + 0.002754437969997525, + 0.0027621022891253233, + 0.0019109949935227633, + 0.00351848010905087, + 0.0020151252392679453, + 0.003755351295694709, + 0.0026021634694188833, + 0.0021405622828751802, + 0.0016071632271632552, + 0.0022289822809398174, + 0.0029666710179299116, + 0.0023129761684685946, + 0.0014370058197528124, + 0.008484352380037308, + 0.003130405442789197, + 0.0022164564579725266, + 0.0026871701702475548, + 0.003639721777290106, + 0.00711506512016058, + 0.002423099474981427, + 0.0029515274800360203, + 0.020386653020977974, + 0.011994878761470318, + 0.004099090117961168, + 0.008575387299060822, + 0.004886038601398468, + 0.007089444901794195, + 0.00910608284175396, + 0.005403278861194849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.016554025933146477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013711584033444524 + ], + "opposite_logprob_tokens": [ + -6.626371383666992 + ], + "logprob_reward": -0.0013711584033444524, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013711584033444524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20395030081272125, + 0.34611618518829346, + 0.03207836672663689, + 0.08175626397132874, + 0.017470354214310646, + 0.02141623944044113, + 0.04220147058367729, + 0.015789978206157684, + 0.07561288774013519, + 0.16360804438591003 + ], + "credit_map": [ + 0.031135374680161476, + 0.04793980345129967, + 0.014707530848681927, + 0.019819680601358414, + 0.029102232307195663, + 0.02719857543706894, + 0.02187349833548069, + 0.012173594906926155, + 0.07189348340034485, + 0.04320808872580528, + 0.018167883157730103, + 0.014588289894163609, + 0.03643845021724701, + 0.1307879239320755, + 0.01641581766307354, + 0.014616217464208603, + 0.008028554730117321, + 0.00788227841258049, + 0.002192036947235465, + 0.002919124672189355, + 0.002630384638905525, + 0.002895565005019307, + 0.0031259183306246996, + 0.00240450375713408, + 0.01565871201455593, + 0.009425305761396885, + 0.005148815922439098, + 0.00856082048267126, + 0.008953551761806011, + 0.021528175100684166, + 0.007434810046106577, + 0.0050460705533623695, + 0.002353264018893242, + 0.0022429528180509806, + 0.0014245721977204084, + 0.0021828117314726114, + 0.002965630730614066, + 0.0018119297455996275, + 0.002518173074349761, + 0.0019710210617631674, + 0.002031106734648347, + 0.0022304996382445097, + 0.0016488457331433892, + 0.0020479841623455286, + 0.004676914308220148, + 0.0030695218592882156, + 0.003959787078201771, + 0.0017515791114419699, + 0.004428059794008732, + 0.0045074885711073875, + 0.002792881801724434, + 0.007314582355320454, + 0.005383143667131662, + 0.005032488610595465, + 0.008616424165666103, + 0.0041264016181230545, + 0.00343573116697371, + 0.002463663462549448, + 0.001419678097590804, + 0.0018825953593477607, + 0.0013336571864783764, + 0.00251688901335001, + 0.0015811384655535221, + 0.001156625454314053, + 0.0257880799472332, + 0.009589516557753086, + 0.004109598696231842, + 0.004688347689807415, + 0.003991380333900452, + 0.01688108965754509, + 0.004316807258874178, + 0.006248061079531908, + 0.04365925118327141, + 0.027516676113009453, + 0.009880580008029938, + 0.020477160811424255, + 0.0112838726490736, + 0.017031148076057434, + 0.021924598142504692, + 0.011834751814603806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0013711584033444524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789064079523087 + ], + "opposite_logprob_tokens": [ + -3.5297889709472656 + ], + "logprob_reward": -0.029789064079523087, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789064079523087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07377371191978455, + 0.4684639573097229, + 0.06918071210384369, + 0.047253720462322235, + 0.01934032142162323, + 0.014573168940842152, + 0.04854842647910118, + 0.014853601343929768, + 0.23186102509498596, + 0.012151303701102734 + ], + "credit_map": [ + 0.009718173183500767, + 0.017372429370880127, + 0.004966499749571085, + 0.0067206500098109245, + 0.012587003409862518, + 0.009677557274699211, + 0.007781823631376028, + 0.004949579481035471, + 0.09841952472925186, + 0.06162898242473602, + 0.025444461032748222, + 0.02756715565919876, + 0.04570384323596954, + 0.16045235097408295, + 0.027239453047513962, + 0.022008251398801804, + 0.015683865174651146, + 0.016408761963248253, + 0.006001846864819527, + 0.007436293642967939, + 0.005425460170954466, + 0.0046964380890131, + 0.005545164458453655, + 0.007982891984283924, + 0.01191371213644743, + 0.005918948445469141, + 0.0037360964342951775, + 0.0032989438623189926, + 0.003984749782830477, + 0.012958663515746593, + 0.0026259049773216248, + 0.0028167101554572582, + 0.0027607653755694628, + 0.0019287324976176023, + 0.00266124471090734, + 0.0020600929856300354, + 0.0019492050632834435, + 0.0018873163498938084, + 0.0018667684635147452, + 0.004226199816912413, + 0.0017064742278307676, + 0.0016246538143604994, + 0.001142882159911096, + 0.001725205685943365, + 0.002454893197864294, + 0.0019024810753762722, + 0.002549824072048068, + 0.0014667566865682602, + 0.00422384450212121, + 0.004554893355816603, + 0.00338950683362782, + 0.008444301784038544, + 0.006696471478790045, + 0.006215409375727177, + 0.01011570356786251, + 0.004908300470560789, + 0.0024248757399618626, + 0.001878366689197719, + 0.001295713591389358, + 0.0015592261916026473, + 0.0016636601649224758, + 0.0019602354150265455, + 0.002359601901844144, + 0.0017119249096140265, + 0.08326643705368042, + 0.031158914789557457, + 0.011657512746751308, + 0.011259439401328564, + 0.009959196671843529, + 0.05443761497735977, + 0.010258764959871769, + 0.019863180816173553, + 0.0026058766525238752, + 0.001711533754132688, + 0.0009300642414018512, + 0.0018059328431263566, + 0.0011631270172074437, + 0.0011089950567111373, + 0.0017420684453099966, + 0.0010837079025804996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.029789064079523087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203797832131386 + ], + "opposite_logprob_tokens": [ + -2.812037944793701 + ], + "logprob_reward": -0.06203797832131386, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203797832131386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07796736806631088, + 0.2940005660057068, + 0.10565119236707687, + 0.046615540981292725, + 0.027339259162545204, + 0.03298501670360565, + 0.08005861937999725, + 0.05000012367963791, + 0.18774187564849854, + 0.09764041751623154 + ], + "credit_map": [ + 0.010617095045745373, + 0.011968761682510376, + 0.008211950771510601, + 0.008679837919771671, + 0.012230715714395046, + 0.009753423742949963, + 0.009191847406327724, + 0.007313728332519531, + 0.06249150633811951, + 0.0374499075114727, + 0.0147565808147192, + 0.012012519873678684, + 0.03078744374215603, + 0.11151082813739777, + 0.011932766065001488, + 0.013059010729193687, + 0.02111782133579254, + 0.01568758860230446, + 0.010846013203263283, + 0.013178309425711632, + 0.01006337720900774, + 0.012493320740759373, + 0.011006585322320461, + 0.011258178390562534, + 0.008603724651038647, + 0.0038281497545540333, + 0.006400196347385645, + 0.005618629045784473, + 0.0032887933775782585, + 0.00679807597771287, + 0.005164759233593941, + 0.006913211662322283, + 0.004806927405297756, + 0.0035907654091715813, + 0.0032220243010669947, + 0.003333489643409848, + 0.0026015799958258867, + 0.0025916520971804857, + 0.002743232063949108, + 0.004449587315320969, + 0.0029454920440912247, + 0.0031828824430704117, + 0.003050417872145772, + 0.0031674709171056747, + 0.006904772948473692, + 0.004668543580919504, + 0.006075229961425066, + 0.002990207402035594, + 0.0071237338706851006, + 0.007612831890583038, + 0.005274241790175438, + 0.012030542828142643, + 0.013751929625868797, + 0.009710635989904404, + 0.0167082529515028, + 0.007846453227102757, + 0.012343949638307095, + 0.008548658341169357, + 0.004163831472396851, + 0.005902399774640799, + 0.004614037927240133, + 0.0059143672697246075, + 0.00466283829882741, + 0.003850037232041359, + 0.06772946566343307, + 0.02303983084857464, + 0.00894208624958992, + 0.0079225217923522, + 0.007535875774919987, + 0.04759820178151131, + 0.007048238534480333, + 0.017925657331943512, + 0.024558229371905327, + 0.015030317939817905, + 0.006427519954741001, + 0.011889167129993439, + 0.007324523292481899, + 0.011162101291120052, + 0.013614085502922535, + 0.007634479086846113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.06203797832131386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a peacock with a massive train fanned out in a semi-circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the peacock's train display iridescent feathers with blue, green, and bronze colors and eye-like patterns (ocelli)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a second peacock nearby with speckled brown and cream feathers standing with its head bowed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the peacock's body trembling and the train feathers vibrating in a wave-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium close-up and then slowly pull back to reveal a sun-drenched forest clearing?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895892858505249 + ], + "opposite_logprob_tokens": [ + -2.5789589881896973 + ], + "logprob_reward": -0.07895892858505249, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895892858505249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07227716594934464, + 0.3164367079734802, + 0.07379808276891708, + 0.04967920482158661, + 0.01595556177198887, + 0.0183620173484087, + 0.04559788107872009, + 0.014696295373141766, + 0.1835607886314392, + 0.20963633060455322 + ], + "credit_map": [ + 0.010533616878092289, + 0.012055234052240849, + 0.007395675405859947, + 0.008826919831335545, + 0.008100618608295918, + 0.009250478819012642, + 0.008541597053408623, + 0.007573023438453674, + 0.06648948788642883, + 0.04033828526735306, + 0.01631135679781437, + 0.01467111799865961, + 0.03265846148133278, + 0.11709388345479965, + 0.015441137365996838, + 0.013432955369353294, + 0.020164519548416138, + 0.021628277376294136, + 0.005481599364429712, + 0.00702762370929122, + 0.006016078405082226, + 0.004501625429838896, + 0.004083981271833181, + 0.004894363693892956, + 0.010732639580965042, + 0.008875850588083267, + 0.003937381319701672, + 0.0038438879419118166, + 0.005178708117455244, + 0.010092843323946, + 0.003943800926208496, + 0.0030740872025489807, + 0.0023369856644421816, + 0.002441294025629759, + 0.0019887175876647234, + 0.0019143254030495882, + 0.0014056116342544556, + 0.001635697204619646, + 0.001771089038811624, + 0.0024618383031338453, + 0.0023970266338437796, + 0.0026265555061399937, + 0.0022507018875330687, + 0.0019732844084501266, + 0.002268248936161399, + 0.0020940459799021482, + 0.0019370389636605978, + 0.0028151129372417927, + 0.0047981636598706245, + 0.005297308787703514, + 0.002817022381350398, + 0.007057161536067724, + 0.007445480674505234, + 0.004836190026253462, + 0.009139872156083584, + 0.004206676036119461, + 0.003911613952368498, + 0.00261036423034966, + 0.0012363310670480132, + 0.0011880763340741396, + 0.001776576740667224, + 0.0019368858775123954, + 0.0009809338953346014, + 0.0010555111803114414, + 0.06481219828128815, + 0.021876201033592224, + 0.008934872224926949, + 0.008740496821701527, + 0.006945544853806496, + 0.046553995460271835, + 0.00848815031349659, + 0.01720930077135563, + 0.05587313324213028, + 0.035241905599832535, + 0.012798813171684742, + 0.02586282417178154, + 0.014623994939029217, + 0.02208220586180687, + 0.028032531961798668, + 0.015120889991521835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.07895892858505249 + } + ], + "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/0000122/group_14/index_tensor([6566, 6566])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_14/index_tensor([6566, 6566])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2f9d9fa03de77c24ab41aca4b63f30ed6f4f26a7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_14/index_tensor([6566, 6566])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, low-angle, static shot captures a scene immersed in a human-scale human experience. Inspired by Giger's style, a perfect, waxy cylinder of vibrant crimson lipstick extends with unsettling slowness from a biomechanical orifice. The opening is a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue, faintly pulsing in near-total darkness. From the lipstick's smoothly sculpted tip, a viscous black liquid begins to ooze, its surface shimmering with an oily, malevolent iridescence under a single, stark overhead light source. The liquid gathers weight, stretching into heavy droplets before separating and falling in deliberate slow motion, landing with a soft, wet impact on the shimmering chitinous surface below—a landscape of polished insectoid plates and skeletal ridges. The harsh light carves the entire image from the oppressive darkness, casting deep, sinister shadows that seem to wriggle and move with unseen, organic motion, creating an atmosphere of intensely tense and surreal beauty and horror.", + "seed": 26002, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_14/index_6566_seed_26002_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": -0.262431760225445, + "vlm_logprob": -0.262431760225445, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.262431760225445, + "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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027111297473311424 + ], + "opposite_logprob_tokens": [ + -4.527111530303955 + ], + "logprob_reward": -0.027111297473311424, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.027111297473311424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06304320693016052, + 0.02376730367541313, + 0.061317719519138336, + 0.04337051510810852, + 0.02573133073747158, + 0.03434726968407631, + 0.11012030392885208, + 0.5577165484428406, + 0.0437726154923439, + 0.0368131622672081 + ], + "credit_map": [ + 0.005460537038743496, + 0.0035561444237828255, + 0.00989577453583479, + 0.005661221221089363, + 0.006470782682299614, + 0.01192049216479063, + 0.011760734021663666, + 0.008317520841956139, + 0.001473316689953208, + 0.0011519994586706161, + 0.004964882042258978, + 0.0016561929369345307, + 0.003090908983722329, + 0.0030812835320830345, + 0.0023579546250402927, + 0.0059907641261816025, + 0.0064253476448357105, + 0.005147083196789026, + 0.007513334043323994, + 0.007478055078536272, + 0.006887126248329878, + 0.008507907390594482, + 0.010161584243178368, + 0.009197283536195755, + 0.005115854553878307, + 0.00323670101352036, + 0.006648203358054161, + 0.003575731534510851, + 0.006024123635143042, + 0.006632060278207064, + 0.006145192310214043, + 0.005992651451379061, + 0.002920447615906596, + 0.00252066133543849, + 0.004776212852448225, + 0.0019089807756245136, + 0.004352846182882786, + 0.0037136287428438663, + 0.002806720556691289, + 0.002731833141297102, + 0.003591173095628619, + 0.004016450140625238, + 0.005358620081096888, + 0.0030275334138423204, + 0.005505418870598078, + 0.004334310535341501, + 0.0047649359330534935, + 0.0037488287780433893, + 0.02130681276321411, + 0.005092531908303499, + 0.004591254983097315, + 0.0041309865191578865, + 0.019918691366910934, + 0.04139019548892975, + 0.006105923559516668, + 0.0075839124619960785, + 0.16147558391094208, + 0.0223719272762537, + 0.020449766889214516, + 0.020451152697205544, + 0.06492982059717178, + 0.22074899077415466, + 0.015258152037858963, + 0.03203115984797478, + 0.005405660718679428, + 0.004636767320334911, + 0.005327294114977121, + 0.004413269925862551, + 0.006408760789781809, + 0.005763646215200424, + 0.0046462006866931915, + 0.00717101339250803, + 0.004570939112454653, + 0.004964772611856461, + 0.002966667525470257, + 0.003973034210503101, + 0.006739452946931124, + 0.0051027750596404076, + 0.003847899381071329, + 0.004647621884942055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.027111297473311424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6934136748313904 + ], + "opposite_logprob_tokens": [ + -0.6934136748313904 + ], + "logprob_reward": -0.6934136748313904, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934136748313904, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10852893441915512, + 0.22970199584960938, + 0.05937942489981651, + 0.0422448068857193, + 0.03592512384057045, + 0.09929672628641129, + 0.07238899171352386, + 0.2662631869316101, + 0.06517041474580765, + 0.02110036462545395 + ], + "credit_map": [ + 0.007379310671240091, + 0.005632265470921993, + 0.019494449719786644, + 0.0103407958522439, + 0.011332083493471146, + 0.014395555481314659, + 0.020466018468141556, + 0.01948845572769642, + 0.011190852150321007, + 0.0055762347765266895, + 0.08149805665016174, + 0.0074875447899103165, + 0.006638854742050171, + 0.018643604591488838, + 0.00814727321267128, + 0.09051958471536636, + 0.006711383815854788, + 0.00470902631059289, + 0.0071524446830153465, + 0.006970696151256561, + 0.008289908058941364, + 0.0072409529238939285, + 0.011584186926484108, + 0.006720825098454952, + 0.004864828195422888, + 0.002203002106398344, + 0.005576366558670998, + 0.003568032756447792, + 0.00473408680409193, + 0.006654046475887299, + 0.006584159564226866, + 0.008060283958911896, + 0.004113590810447931, + 0.0038217136170715094, + 0.005320187658071518, + 0.0036802436225116253, + 0.004850259982049465, + 0.004531781654804945, + 0.005659020040184259, + 0.00394832668825984, + 0.01031851302832365, + 0.012173796072602272, + 0.012746981345117092, + 0.010133272036910057, + 0.013988972641527653, + 0.01114595029503107, + 0.016165629029273987, + 0.012623612768948078, + 0.007251814939081669, + 0.00702602369710803, + 0.005658970680087805, + 0.0039994241669774055, + 0.018698452040553093, + 0.018699346110224724, + 0.00393463671207428, + 0.0071203201077878475, + 0.07443416118621826, + 0.011592116206884384, + 0.009853589348495007, + 0.008477149531245232, + 0.03294889256358147, + 0.10371914505958557, + 0.008516127243638039, + 0.01672201417386532, + 0.00754641555249691, + 0.009960881434381008, + 0.006675513461232185, + 0.004471893887966871, + 0.013887431472539902, + 0.0076589034870266914, + 0.005522808060050011, + 0.009446565993130207, + 0.0031398185528814793, + 0.0023750828113406897, + 0.001829621265642345, + 0.001869976636953652, + 0.0036610267125070095, + 0.0031394914258271456, + 0.002295023063197732, + 0.0027903246227651834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.6934136748313904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00675099017098546 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.00675099017098546, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00675099017098546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18149927258491516, + 0.1760740578174591, + 0.06643106043338776, + 0.06620799750089645, + 0.05206594616174698, + 0.06567297875881195, + 0.12780340015888214, + 0.09741769731044769, + 0.0803191214799881, + 0.08650847524404526 + ], + "credit_map": [ + 0.01577925682067871, + 0.008696929551661015, + 0.035314422100782394, + 0.014536067843437195, + 0.02126307599246502, + 0.035540103912353516, + 0.027972115203738213, + 0.022397303953766823, + 0.010719086043536663, + 0.007335004862397909, + 0.04150557890534401, + 0.01570155844092369, + 0.01942208781838417, + 0.01927739568054676, + 0.02533586136996746, + 0.03677749261260033, + 0.010493463836610317, + 0.0044103204272687435, + 0.0071832966059446335, + 0.004817082546651363, + 0.007044701371341944, + 0.01201360672712326, + 0.009617016650736332, + 0.010851573199033737, + 0.009735874831676483, + 0.006281532812863588, + 0.009968486614525318, + 0.004333621822297573, + 0.008897185325622559, + 0.009915243834257126, + 0.007253939285874367, + 0.00982210785150528, + 0.0070674484595656395, + 0.007441327441483736, + 0.006284271366894245, + 0.003808962181210518, + 0.009286846965551376, + 0.005638713017106056, + 0.0067941732704639435, + 0.005744203459471464, + 0.009803622029721737, + 0.00900436844676733, + 0.006956429220736027, + 0.0044038742780685425, + 0.01171112060546875, + 0.00990193709731102, + 0.007135110907256603, + 0.006756514776498079, + 0.017899924889206886, + 0.008177812211215496, + 0.004877020604908466, + 0.005038605537265539, + 0.029029764235019684, + 0.04700387641787529, + 0.006737228017300367, + 0.009039168246090412, + 0.024350779131054878, + 0.00736704608425498, + 0.006368658505380154, + 0.0055784606374800205, + 0.012278280220925808, + 0.03065110556781292, + 0.0052834018133580685, + 0.005539966747164726, + 0.008126912638545036, + 0.005758046172559261, + 0.009231239557266235, + 0.006448105443269014, + 0.012739932164549828, + 0.009251927025616169, + 0.007804416585713625, + 0.020958542823791504, + 0.01198576856404543, + 0.009929545223712921, + 0.008933866396546364, + 0.009948494844138622, + 0.014622308313846588, + 0.012869623489677906, + 0.007455153856426477, + 0.01076371967792511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00675099017098546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.17986875772476196, + 0.23484569787979126, + 0.057638175785541534, + 0.04991311952471733, + 0.04561841860413551, + 0.042617324739694595, + 0.06301622092723846, + 0.18727155029773712, + 0.0716509148478508, + 0.06755978614091873 + ], + "credit_map": [ + 0.014041292481124401, + 0.007994728162884712, + 0.03656220808625221, + 0.012219395488500595, + 0.01808968372642994, + 0.03412243351340294, + 0.03135177865624428, + 0.025487223640084267, + 0.012124729342758656, + 0.006707946304231882, + 0.07463452965021133, + 0.009843871928751469, + 0.017178835347294807, + 0.022532710805535316, + 0.013784355483949184, + 0.07803872972726822, + 0.005976343061774969, + 0.005314757581800222, + 0.007459975313395262, + 0.006335256155580282, + 0.00732149463146925, + 0.007791117765009403, + 0.009767540730535984, + 0.007671690080314875, + 0.006439408753067255, + 0.004171147011220455, + 0.006932354066520929, + 0.0037334002554416656, + 0.006997422315180302, + 0.008906006813049316, + 0.0059861731715500355, + 0.006747206207364798, + 0.005663177464157343, + 0.005305022466927767, + 0.005802092142403126, + 0.003913580905646086, + 0.008370529860258102, + 0.006061788648366928, + 0.005467867944389582, + 0.005034359637647867, + 0.004896619822829962, + 0.006835214328020811, + 0.003864419413730502, + 0.002841278677806258, + 0.005165142938494682, + 0.006741686258465052, + 0.00641796737909317, + 0.005854997783899307, + 0.01620255783200264, + 0.004772770684212446, + 0.003015805035829544, + 0.00317340693436563, + 0.007599343545734882, + 0.01737661100924015, + 0.00527532072737813, + 0.00560040632262826, + 0.03787277266383171, + 0.009693564847111702, + 0.01644206792116165, + 0.011203885078430176, + 0.022692183032631874, + 0.06031816080212593, + 0.009024393744766712, + 0.020024513825774193, + 0.008305899798870087, + 0.007350690197199583, + 0.008674252778291702, + 0.006343877408653498, + 0.011207721196115017, + 0.009511594660580158, + 0.006934143137186766, + 0.013322739861905575, + 0.009968825615942478, + 0.005480288062244654, + 0.006715467199683189, + 0.00698945252224803, + 0.011315357871353626, + 0.012492825277149677, + 0.005764303728938103, + 0.008833265863358974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.008648676797747612 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762341618537903 + ], + "opposite_logprob_tokens": [ + -0.8262341618537903 + ], + "logprob_reward": -0.5762341618537903, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762341618537903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07705194503068924, + 0.07523537427186966, + 0.056073252111673355, + 0.05822842940688133, + 0.13902097940444946, + 0.043442729860544205, + 0.04424412548542023, + 0.40010350942611694, + 0.035751622170209885, + 0.0708480179309845 + ], + "credit_map": [ + 0.00560197327286005, + 0.004686963278800249, + 0.012384166941046715, + 0.006517833564430475, + 0.010202400386333466, + 0.010304798372089863, + 0.015128214843571186, + 0.012225589714944363, + 0.004332355223596096, + 0.0027153855189681053, + 0.021627821028232574, + 0.003546172985807061, + 0.005958896595984697, + 0.0076956613920629025, + 0.005607111379504204, + 0.023751968517899513, + 0.005266394931823015, + 0.003924731630831957, + 0.009568334557116032, + 0.005930393468588591, + 0.00853649340569973, + 0.008311895653605461, + 0.006890579126775265, + 0.007644429802894592, + 0.008153158240020275, + 0.004304112400859594, + 0.00902390107512474, + 0.004211661871522665, + 0.007533577270805836, + 0.0078587606549263, + 0.008563711307942867, + 0.008579548448324203, + 0.017416836693882942, + 0.019656654447317123, + 0.015089093707501888, + 0.010468262247741222, + 0.03139550983905792, + 0.019811498001217842, + 0.01473243162035942, + 0.010450686328113079, + 0.004632849246263504, + 0.006651267409324646, + 0.003804260166361928, + 0.0039465161971747875, + 0.005023206118494272, + 0.005459314677864313, + 0.005210472270846367, + 0.008714842610061169, + 0.009595191106200218, + 0.005112325772643089, + 0.0021510422229766846, + 0.0016888158861547709, + 0.013501539826393127, + 0.0060899583622813225, + 0.0026814094744622707, + 0.003423842368647456, + 0.11531615257263184, + 0.015143369324505329, + 0.01382087916135788, + 0.014994309283792973, + 0.046850450336933136, + 0.16027243435382843, + 0.010906518436968327, + 0.022799359634518623, + 0.0049369377084076405, + 0.004777982831001282, + 0.0037935415748506784, + 0.002688345964998007, + 0.0062222592532634735, + 0.004469634033739567, + 0.002970281057059765, + 0.005892638117074966, + 0.009721475653350353, + 0.010305498726665974, + 0.006480288226157427, + 0.0067664580419659615, + 0.013836606405675411, + 0.010450980626046658, + 0.004787900950759649, + 0.008498810231685638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.5762341618537903 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot captures a scene immersed in a human-scale human experience. Inspired by Giger's style, a perfect, waxy cylinder of vibrant crimson lipstick extends with unsettling slowness from a biomechanical orifice. The opening is a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue, faintly pulsing in near-total darkness. From the lipstick's smoothly sculpted tip, a viscous black liquid begins to ooze, its surface shimmering with an oily, malevolent iridescence under a single, stark overhead light source. The liquid gathers weight, stretching into heavy droplets before separating and falling in deliberate slow motion, landing with a soft, wet impact on the shimmering chitinous surface below—a landscape of polished insectoid plates and skeletal ridges. The harsh light carves the entire image from the oppressive darkness, casting deep, sinister shadows that seem to wriggle and move with unseen, organic motion, creating an atmosphere of intensely tense and surreal beauty and horror.", + "seed": 26003, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_14/index_6566_seed_26003_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": -0.36200249698013065, + "vlm_logprob": -0.36200249698013065, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.36200249698013065, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052489880472421646 + ], + "opposite_logprob_tokens": [ + -3.677489995956421 + ], + "logprob_reward": -0.052489880472421646, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.052489880472421646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18836306035518646, + 0.07373003661632538, + 0.055530477315187454, + 0.05222839489579201, + 0.09794176369905472, + 0.06076859310269356, + 0.09017159789800644, + 0.19265879690647125, + 0.12217497080564499, + 0.06643232703208923 + ], + "credit_map": [ + 0.017063478007912636, + 0.014909075573086739, + 0.02291703410446644, + 0.013468298129737377, + 0.0230342335999012, + 0.03370232135057449, + 0.035480257123708725, + 0.027788344770669937, + 0.00870540738105774, + 0.0072191208600997925, + 0.0056166257709264755, + 0.006594058126211166, + 0.006768534425646067, + 0.014763330109417439, + 0.014127027243375778, + 0.009935930371284485, + 0.009953559376299381, + 0.003965155687183142, + 0.008804001845419407, + 0.004127762280404568, + 0.00345635530538857, + 0.009941349737346172, + 0.008002174086868763, + 0.007280111778527498, + 0.007534130476415157, + 0.0052984305657446384, + 0.007723306771367788, + 0.003939182497560978, + 0.0038254582323133945, + 0.009532937780022621, + 0.007656255271285772, + 0.006718688178807497, + 0.012639174237847328, + 0.007749959826469421, + 0.01219411101192236, + 0.009216534905135632, + 0.008797425776720047, + 0.015300750732421875, + 0.01764223724603653, + 0.01440155878663063, + 0.007122353184968233, + 0.004076293203979731, + 0.006864587776362896, + 0.006465731654316187, + 0.005372423212975264, + 0.008879906497895718, + 0.014087098650634289, + 0.007900193333625793, + 0.008702052757143974, + 0.006171073764562607, + 0.007860221900045872, + 0.007032669615000486, + 0.012507811188697815, + 0.013759889639914036, + 0.020480122417211533, + 0.013657748699188232, + 0.024100081995129585, + 0.014802307821810246, + 0.01787686161696911, + 0.014895166270434856, + 0.019898895174264908, + 0.037010859698057175, + 0.036832619458436966, + 0.027241988107562065, + 0.013816150836646557, + 0.010368111543357372, + 0.014026069082319736, + 0.011716044507920742, + 0.012941494584083557, + 0.021523991599678993, + 0.01886730268597603, + 0.01891578920185566, + 0.007524058222770691, + 0.005945588927716017, + 0.005755877587944269, + 0.005900627002120018, + 0.00687501160427928, + 0.009990176185965538, + 0.014484187588095665, + 0.009956787340342999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.052489880472421646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578808069229126 + ], + "opposite_logprob_tokens": [ + -0.6328808069229126 + ], + "logprob_reward": -0.7578808069229126, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578808069229126, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14086371660232544, + 0.1195121705532074, + 0.07499601691961288, + 0.136647567152977, + 0.09026645869016647, + 0.0813179537653923, + 0.0465085543692112, + 0.06633856147527695, + 0.1242557242512703, + 0.11929327249526978 + ], + "credit_map": [ + 0.01317194476723671, + 0.011084256693720818, + 0.016495386138558388, + 0.012764892540872097, + 0.01917785219848156, + 0.02210276387631893, + 0.027222733944654465, + 0.018843892961740494, + 0.014285480603575706, + 0.009264974854886532, + 0.015534291975200176, + 0.01304017286747694, + 0.010122308507561684, + 0.019596753641963005, + 0.022577710449695587, + 0.01509048230946064, + 0.011846455745398998, + 0.005579452961683273, + 0.009042025543749332, + 0.006193523295223713, + 0.005178434308618307, + 0.014169231988489628, + 0.011769982054829597, + 0.011216908693313599, + 0.021493634209036827, + 0.010067255236208439, + 0.027706457301974297, + 0.01070378627628088, + 0.0075497752986848354, + 0.02164565771818161, + 0.018671797588467598, + 0.018809199333190918, + 0.01302493829280138, + 0.010224800556898117, + 0.007349912077188492, + 0.00803398061543703, + 0.008985359221696854, + 0.014598894864320755, + 0.01560095977038145, + 0.012447611428797245, + 0.013557301834225655, + 0.008502383716404438, + 0.013299095444381237, + 0.006899238098412752, + 0.006039280444383621, + 0.010217549279332161, + 0.0127748167142272, + 0.010028291493654251, + 0.005073771346360445, + 0.004345006309449673, + 0.007504693698137999, + 0.005674210377037525, + 0.003967856056988239, + 0.007005407474935055, + 0.006595842074602842, + 0.006341766566038132, + 0.006482687313109636, + 0.005463628098368645, + 0.0065863365307450294, + 0.005777458194643259, + 0.008552800863981247, + 0.012008776888251305, + 0.011377270333468914, + 0.010089606046676636, + 0.009735497646033764, + 0.009996481239795685, + 0.01189770270138979, + 0.01057275477796793, + 0.018746381625533104, + 0.024000531062483788, + 0.021971367299556732, + 0.017335010692477226, + 0.014964936301112175, + 0.010424429550766945, + 0.00910775177180767, + 0.012810376472771168, + 0.010132389143109322, + 0.01996288076043129, + 0.022335698828101158, + 0.019554803147912025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.7578808069229126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011100560426712036 + ], + "opposite_logprob_tokens": [ + -4.511100769042969 + ], + "logprob_reward": -0.011100560426712036, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011100560426712036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34123480319976807, + 0.12520506978034973, + 0.07653279602527618, + 0.06201334297657013, + 0.06345219910144806, + 0.07575053721666336, + 0.04055802896618843, + 0.08300961554050446, + 0.03999471664428711, + 0.09224890172481537 + ], + "credit_map": [ + 0.045653339475393295, + 0.026468010619282722, + 0.042937397956848145, + 0.02679937332868576, + 0.04557673633098602, + 0.06074044108390808, + 0.04277756065130234, + 0.05028194561600685, + 0.016858085989952087, + 0.008550936356186867, + 0.016262562945485115, + 0.009470473043620586, + 0.01288170088082552, + 0.021239999681711197, + 0.021682994440197945, + 0.01825830712914467, + 0.013886310160160065, + 0.006471300031989813, + 0.010087362490594387, + 0.005558840464800596, + 0.004262779373675585, + 0.014947809278964996, + 0.010402441024780273, + 0.010915948078036308, + 0.00818101316690445, + 0.006918853614479303, + 0.004738786723464727, + 0.007966886274516582, + 0.005041332449764013, + 0.009725989773869514, + 0.011355663649737835, + 0.008084822446107864, + 0.008145584724843502, + 0.005155859049409628, + 0.006484272889792919, + 0.004532939754426479, + 0.007944771088659763, + 0.011660229414701462, + 0.01080243568867445, + 0.008726106956601143, + 0.008485469967126846, + 0.005768326576799154, + 0.009466730989515781, + 0.008870954625308514, + 0.006263247691094875, + 0.01335509866476059, + 0.013524863868951797, + 0.010015846230089664, + 0.003926499281078577, + 0.0029371841810643673, + 0.008154134266078472, + 0.0035393740981817245, + 0.004675536882132292, + 0.006045136600732803, + 0.005947635509073734, + 0.005332526285201311, + 0.00791933387517929, + 0.006911012809723616, + 0.00704355351626873, + 0.00685016717761755, + 0.00935453176498413, + 0.015509211458265781, + 0.016681034117937088, + 0.012740771286189556, + 0.005235210061073303, + 0.0038675551768392324, + 0.004340076819062233, + 0.00337383639998734, + 0.0028222829569131136, + 0.006368553731590509, + 0.008038359694182873, + 0.005948840640485287, + 0.010332614183425903, + 0.007692092098295689, + 0.005753183737397194, + 0.00740903289988637, + 0.011828391812741756, + 0.018028629943728447, + 0.017842572182416916, + 0.013362384401261806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.011100560426712036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210931025445461 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210931025445461, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210931025445461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39375826716423035, + 0.10352795571088791, + 0.05733047053217888, + 0.04626622051000595, + 0.07906563580036163, + 0.07654371857643127, + 0.053929403424263, + 0.03973522037267685, + 0.05166015401482582, + 0.09818297624588013 + ], + "credit_map": [ + 0.05738500505685806, + 0.03288450837135315, + 0.02896701917052269, + 0.033505428582429886, + 0.04380182549357414, + 0.07784225791692734, + 0.05720318481326103, + 0.062168993055820465, + 0.015560953877866268, + 0.006067714653909206, + 0.019138004630804062, + 0.007665934041142464, + 0.008669395931065083, + 0.016107745468616486, + 0.016507940366864204, + 0.013810268603265285, + 0.010213970206677914, + 0.004435067065060139, + 0.006184990983456373, + 0.004299716558307409, + 0.0037924221251159906, + 0.011879123747348785, + 0.007787500042468309, + 0.008737679570913315, + 0.005667282268404961, + 0.003774139564484358, + 0.003969648852944374, + 0.004989283625036478, + 0.004264226648956537, + 0.007804156746715307, + 0.009473942220211029, + 0.006323540583252907, + 0.011631711386144161, + 0.006811229977756739, + 0.00815728958696127, + 0.007367010693997145, + 0.007100914139300585, + 0.012886440381407738, + 0.013725901022553444, + 0.011385142803192139, + 0.007648978382349014, + 0.005993838422000408, + 0.010780664160847664, + 0.010281232185661793, + 0.009757151827216148, + 0.011381791904568672, + 0.01122576929628849, + 0.009474294260144234, + 0.0065143052488565445, + 0.004322112537920475, + 0.00549927307292819, + 0.0037717767991125584, + 0.005170017946511507, + 0.010709981434047222, + 0.010127540677785873, + 0.007814393378794193, + 0.0038742420729249716, + 0.002718678442761302, + 0.004245859105139971, + 0.002718193456530571, + 0.004315529018640518, + 0.006574494764208794, + 0.009167049080133438, + 0.006121174897998571, + 0.006209490820765495, + 0.005369445309042931, + 0.006111152470111847, + 0.00438691908493638, + 0.004773373249918222, + 0.007964964024722576, + 0.009500421583652496, + 0.007344385609030724, + 0.013048476539552212, + 0.008496340364217758, + 0.006291182246059179, + 0.008855647407472134, + 0.01062606368213892, + 0.01998310163617134, + 0.015758655965328217, + 0.015123503282666206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.014210931025445461 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9743303060531616 + ], + "opposite_logprob_tokens": [ + -0.4743303060531616 + ], + "logprob_reward": -0.9743303060531616, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743303060531616, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1495734304189682, + 0.13896240293979645, + 0.08704235404729843, + 0.06906741857528687, + 0.05627154931426048, + 0.048881832510232925, + 0.07794535160064697, + 0.1654045581817627, + 0.053190432488918304, + 0.15366068482398987 + ], + "credit_map": [ + 0.01661212183535099, + 0.010936792008578777, + 0.02592068910598755, + 0.01364391390234232, + 0.01582316681742668, + 0.025847874581813812, + 0.019152354449033737, + 0.021636508405208588, + 0.014034705236554146, + 0.011436045169830322, + 0.009080040268599987, + 0.011560640297830105, + 0.015406100079417229, + 0.03366956114768982, + 0.02380269765853882, + 0.019972605630755424, + 0.015951357781887054, + 0.006718834862112999, + 0.016768502071499825, + 0.006546781398355961, + 0.005098041146993637, + 0.014273139648139477, + 0.010674293152987957, + 0.011011398397386074, + 0.011634932830929756, + 0.006044698413461447, + 0.012386677786707878, + 0.00537473801523447, + 0.004540943540632725, + 0.011406823992729187, + 0.009506994858384132, + 0.008171609602868557, + 0.0072330874390900135, + 0.004129157867282629, + 0.0058356765657663345, + 0.0038862847723066807, + 0.003988088574260473, + 0.008635982871055603, + 0.012041712179780006, + 0.010521559976041317, + 0.0054948157630860806, + 0.0044718775898218155, + 0.0031541467178612947, + 0.004346735309809446, + 0.005859791301190853, + 0.009766687639057636, + 0.008565749041736126, + 0.007222028449177742, + 0.009792149998247623, + 0.007653440814465284, + 0.00986618734896183, + 0.006077336147427559, + 0.005053555592894554, + 0.01343219168484211, + 0.013757256790995598, + 0.012313232757151127, + 0.014701969921588898, + 0.01315796747803688, + 0.017223555594682693, + 0.012460939586162567, + 0.01766302064061165, + 0.029233232140541077, + 0.03349332883954048, + 0.027470553293824196, + 0.007821272127330303, + 0.005108758341521025, + 0.008413939736783504, + 0.0044291336089372635, + 0.004571861587464809, + 0.008034471422433853, + 0.006988149136304855, + 0.007822846062481403, + 0.02123343199491501, + 0.013174579478800297, + 0.011014615185558796, + 0.02020602859556675, + 0.01340657938271761, + 0.025315875187516212, + 0.028762085363268852, + 0.02054748311638832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.9743303060531616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot captures a scene immersed in a human-scale human experience. Inspired by Giger's style, a perfect, waxy cylinder of vibrant crimson lipstick extends with unsettling slowness from a biomechanical orifice. The opening is a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue, faintly pulsing in near-total darkness. From the lipstick's smoothly sculpted tip, a viscous black liquid begins to ooze, its surface shimmering with an oily, malevolent iridescence under a single, stark overhead light source. The liquid gathers weight, stretching into heavy droplets before separating and falling in deliberate slow motion, landing with a soft, wet impact on the shimmering chitinous surface below—a landscape of polished insectoid plates and skeletal ridges. The harsh light carves the entire image from the oppressive darkness, casting deep, sinister shadows that seem to wriggle and move with unseen, organic motion, creating an atmosphere of intensely tense and surreal beauty and horror.", + "seed": 26004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_14/index_6566_seed_26004_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": -0.29815336801111697, + "vlm_logprob": -0.29815336801111697, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.29815336801111697, + "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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026801472529768944 + ], + "opposite_logprob_tokens": [ + -4.776801586151123 + ], + "logprob_reward": -0.026801472529768944, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.026801472529768944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16763027012348175, + 0.06495381891727448, + 0.06686155498027802, + 0.04173380136489868, + 0.0824660137295723, + 0.06430759280920029, + 0.06311003863811493, + 0.06196160987019539, + 0.19957862794399261, + 0.18739666044712067 + ], + "credit_map": [ + 0.02127057872712612, + 0.02301275171339512, + 0.025455789640545845, + 0.022810418158769608, + 0.021240876987576485, + 0.023606976494193077, + 0.015192779712378979, + 0.015040099620819092, + 0.006599316373467445, + 0.009909586980938911, + 0.010666636750102043, + 0.009358816780149937, + 0.010348225943744183, + 0.008647622540593147, + 0.004554394632577896, + 0.004869221709668636, + 0.007896032184362411, + 0.008286547847092152, + 0.01101057231426239, + 0.00903988629579544, + 0.008633339777588844, + 0.009030935354530811, + 0.005842853337526321, + 0.007121393922716379, + 0.004138047341257334, + 0.005758693441748619, + 0.008522224612534046, + 0.0058406926691532135, + 0.004314852878451347, + 0.005917943548411131, + 0.0030834635253995657, + 0.004157884046435356, + 0.007374376058578491, + 0.010465717874467373, + 0.011397737078368664, + 0.008733787573873997, + 0.01348850131034851, + 0.014241396449506283, + 0.008853249251842499, + 0.007911240682005882, + 0.00850104819983244, + 0.009354549460113049, + 0.009256250225007534, + 0.005901151802390814, + 0.010382228530943394, + 0.01048396248370409, + 0.004743323661386967, + 0.005685071460902691, + 0.007015352603048086, + 0.00745241018012166, + 0.008066206239163876, + 0.007538764737546444, + 0.012268532998859882, + 0.00910252332687378, + 0.004509367980062962, + 0.007156880106776953, + 0.007382669486105442, + 0.007622873410582542, + 0.007698710076510906, + 0.010296056047081947, + 0.009477381594479084, + 0.009707287885248661, + 0.0037187887355685234, + 0.006057841703295708, + 0.013123976066708565, + 0.01999540627002716, + 0.02151408977806568, + 0.045007430016994476, + 0.024596471339464188, + 0.03614156320691109, + 0.02170412801206112, + 0.017495568841695786, + 0.016741571947932243, + 0.015372071415185928, + 0.016589563339948654, + 0.03419218212366104, + 0.022315319627523422, + 0.030069895088672638, + 0.037049729377031326, + 0.015066336840391159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.026801472529768944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9743379950523376 + ], + "opposite_logprob_tokens": [ + -0.47433799505233765 + ], + "logprob_reward": -0.9743379950523376, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743379950523376, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10984180867671967, + 0.033318229019641876, + 0.028176702558994293, + 0.03233184665441513, + 0.06519362330436707, + 0.03696652501821518, + 0.09618741273880005, + 0.16037310659885406, + 0.25170278549194336, + 0.18590793013572693 + ], + "credit_map": [ + 0.011355034075677395, + 0.015615444630384445, + 0.014800267294049263, + 0.014638800173997879, + 0.02022874355316162, + 0.015768542885780334, + 0.009286642074584961, + 0.008148332126438618, + 0.0041090017184615135, + 0.0032645429018884897, + 0.005327265709638596, + 0.004285778850317001, + 0.0040137795731425285, + 0.004896296188235283, + 0.0033207382075488567, + 0.004100825637578964, + 0.0032865318935364485, + 0.0030364806298166513, + 0.004442702047526836, + 0.0036948896013200283, + 0.003730860771611333, + 0.0037152650766074657, + 0.0030762036330997944, + 0.003193769371137023, + 0.003030510153621435, + 0.0038003309164196253, + 0.00450081517919898, + 0.0032485872507095337, + 0.00635892990976572, + 0.004547980148345232, + 0.003731826087459922, + 0.00311286561191082, + 0.0051689669489860535, + 0.0067033106461167336, + 0.008489631116390228, + 0.006182752083986998, + 0.011607182212173939, + 0.010799561627209187, + 0.01011770311743021, + 0.006124519277364016, + 0.005063481163233519, + 0.004140704404562712, + 0.006490583531558514, + 0.004071446601301432, + 0.004250930156558752, + 0.005868821870535612, + 0.003054990665987134, + 0.004025568254292011, + 0.012397749349474907, + 0.014190592803061008, + 0.014131181873381138, + 0.012225345708429813, + 0.012760955840349197, + 0.014019213616847992, + 0.007654780056327581, + 0.00880760233849287, + 0.013244177214801311, + 0.021780595183372498, + 0.021058660000562668, + 0.015828821808099747, + 0.03963307663798332, + 0.02498760260641575, + 0.010230854153633118, + 0.01360931247472763, + 0.017497468739748, + 0.0204433873295784, + 0.021295102313160896, + 0.026284199208021164, + 0.08417527377605438, + 0.04262874647974968, + 0.02095218561589718, + 0.01842641644179821, + 0.016223903745412827, + 0.016085321083664894, + 0.015858059749007225, + 0.03058941289782524, + 0.035703353583812714, + 0.030981753021478653, + 0.028007183223962784, + 0.01245894469320774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.9743379950523376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1723846048116684, + 0.08469433337450027, + 0.12194247543811798, + 0.0417751669883728, + 0.0487329475581646, + 0.05286421999335289, + 0.06969314813613892, + 0.08782930672168732, + 0.19073837995529175, + 0.12934541702270508 + ], + "credit_map": [ + 0.020523810759186745, + 0.020822176709771156, + 0.03379172459244728, + 0.022027093917131424, + 0.01689295284450054, + 0.02594234049320221, + 0.016581924632191658, + 0.01580258086323738, + 0.006993693765252829, + 0.01265836600214243, + 0.012012362480163574, + 0.010078144259750843, + 0.01644706353545189, + 0.011308620683848858, + 0.008884456939995289, + 0.006311621516942978, + 0.011225610040128231, + 0.016407867893576622, + 0.02136738784611225, + 0.011661465279757977, + 0.021282432600855827, + 0.018995294347405434, + 0.010660708881914616, + 0.01034171599894762, + 0.004555650986731052, + 0.005215964745730162, + 0.007103248033672571, + 0.005790205672383308, + 0.005579859483987093, + 0.005341426935046911, + 0.004353052005171776, + 0.0038357609882950783, + 0.005571202374994755, + 0.0057541062124073505, + 0.00766631169244647, + 0.00585272628813982, + 0.007213221862912178, + 0.006888069678097963, + 0.005780302453786135, + 0.004007006995379925, + 0.006941035855561495, + 0.007091830484569073, + 0.00698647927492857, + 0.00538750272244215, + 0.009182724170386791, + 0.007407653611153364, + 0.005189346615225077, + 0.004677646327763796, + 0.007798843085765839, + 0.009497765451669693, + 0.008861941285431385, + 0.007527729030698538, + 0.01292119175195694, + 0.010706014931201935, + 0.006170372013002634, + 0.006209291517734528, + 0.007484516594558954, + 0.010409869253635406, + 0.007985561154782772, + 0.008081846870481968, + 0.02747996896505356, + 0.014525067061185837, + 0.005884109064936638, + 0.005978366360068321, + 0.019235573709011078, + 0.018588341772556305, + 0.022198479622602463, + 0.030418282374739647, + 0.031598612666130066, + 0.029606647789478302, + 0.01949586160480976, + 0.019596559926867485, + 0.010643025860190392, + 0.013036190532147884, + 0.014091008342802525, + 0.015538075007498264, + 0.024361222982406616, + 0.02327639050781727, + 0.017947949469089508, + 0.010451547801494598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.004111053422093391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089006438851357 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011089006438851357, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089006438851357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21131649613380432, + 0.07809672504663467, + 0.0647745355963707, + 0.07726407796144485, + 0.0551171638071537, + 0.06396225094795227, + 0.035996872931718826, + 0.10519421100616455, + 0.15718474984169006, + 0.15109287202358246 + ], + "credit_map": [ + 0.025179391726851463, + 0.023191845044493675, + 0.038505397737026215, + 0.0306791253387928, + 0.018870672211050987, + 0.03428521379828453, + 0.020482776686549187, + 0.020122066140174866, + 0.00868289079517126, + 0.00979282334446907, + 0.014187992550432682, + 0.01137026771903038, + 0.008666823618113995, + 0.011697709560394287, + 0.006667547393590212, + 0.007030673325061798, + 0.007417891174554825, + 0.006865324918180704, + 0.009727098979055882, + 0.00903019867837429, + 0.01002361997961998, + 0.009563824161887169, + 0.005479452665895224, + 0.0066671245731413364, + 0.007565825711935759, + 0.0098576620221138, + 0.012807984836399555, + 0.009821921586990356, + 0.010271826758980751, + 0.010940153151750565, + 0.007143697235733271, + 0.008855009451508522, + 0.006539632566273212, + 0.0055181789211928844, + 0.008283914998173714, + 0.006687472108751535, + 0.006060272920876741, + 0.00773653807118535, + 0.008436230942606926, + 0.0058549246750772, + 0.007765153888612986, + 0.0068029616959393024, + 0.010162348859012127, + 0.008094766177237034, + 0.008249207399785519, + 0.01046543288975954, + 0.005214487202465534, + 0.0072078946977853775, + 0.004268651828169823, + 0.004358161706477404, + 0.005781053565442562, + 0.003709250595420599, + 0.005310846492648125, + 0.0050339545123279095, + 0.0038214698433876038, + 0.0037134834565222263, + 0.010760742239654064, + 0.01231915783137083, + 0.010400181636214256, + 0.01221188623458147, + 0.025663234293460846, + 0.01623186096549034, + 0.008397480472922325, + 0.009209660813212395, + 0.01355757750570774, + 0.013480923138558865, + 0.020996859297156334, + 0.01951836608350277, + 0.03320883959531784, + 0.025304174050688744, + 0.017060786485671997, + 0.014057218097150326, + 0.011849849484860897, + 0.01553275715559721, + 0.019062258303165436, + 0.02067256160080433, + 0.026090536266565323, + 0.025635872036218643, + 0.019252950325608253, + 0.012996078468859196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.011089006438851357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47442731261253357 + ], + "opposite_logprob_tokens": [ + -0.974427342414856 + ], + "logprob_reward": -0.47442731261253357, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47442731261253357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10326103121042252, + 0.04755004122853279, + 0.05775774270296097, + 0.05570153519511223, + 0.09722612798213959, + 0.099403016269207, + 0.14429937303066254, + 0.17934240400791168, + 0.08984185755252838, + 0.1256168931722641 + ], + "credit_map": [ + 0.012798669748008251, + 0.011814329773187637, + 0.014536133967339993, + 0.010962381027638912, + 0.017937077209353447, + 0.014771933667361736, + 0.009015468880534172, + 0.011425037868320942, + 0.004574329126626253, + 0.00595715781673789, + 0.00909222848713398, + 0.006344130262732506, + 0.00632381159812212, + 0.006861768662929535, + 0.003900311654433608, + 0.004496306646615267, + 0.006548577919602394, + 0.006914699450135231, + 0.008693430572748184, + 0.007838607765734196, + 0.008187313564121723, + 0.008565232157707214, + 0.006262288894504309, + 0.004747588653117418, + 0.005687108263373375, + 0.007016296964138746, + 0.00912484060972929, + 0.009108682163059711, + 0.0054511637426912785, + 0.00806774664670229, + 0.005659080110490322, + 0.005586616229265928, + 0.007729555480182171, + 0.011543752625584602, + 0.01331468764692545, + 0.008927113376557827, + 0.013579349964857101, + 0.016669338569045067, + 0.016570476815104485, + 0.00889185257256031, + 0.009828425943851471, + 0.013443147763609886, + 0.01220108661800623, + 0.006456416100263596, + 0.02575315348803997, + 0.016278767958283424, + 0.008394156582653522, + 0.007047863211482763, + 0.012452633120119572, + 0.012776005081832409, + 0.018746700137853622, + 0.014684491790831089, + 0.03750326856970787, + 0.023805292323231697, + 0.012303279712796211, + 0.012027707882225513, + 0.014929411001503468, + 0.019942495971918106, + 0.02263845130801201, + 0.017212221398949623, + 0.05240023881196976, + 0.02847704291343689, + 0.012055945582687855, + 0.01168661005795002, + 0.0064036548137664795, + 0.007196617312729359, + 0.00882731657475233, + 0.013157323002815247, + 0.024305960163474083, + 0.017197543755173683, + 0.00542784109711647, + 0.007325598504394293, + 0.01111913938075304, + 0.017393967136740685, + 0.014321334660053253, + 0.017062820494174957, + 0.020132414996623993, + 0.019093265756964684, + 0.014172347262501717, + 0.012321600690484047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.47442731261253357 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot captures a scene immersed in a human-scale human experience. Inspired by Giger's style, a perfect, waxy cylinder of vibrant crimson lipstick extends with unsettling slowness from a biomechanical orifice. The opening is a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue, faintly pulsing in near-total darkness. From the lipstick's smoothly sculpted tip, a viscous black liquid begins to ooze, its surface shimmering with an oily, malevolent iridescence under a single, stark overhead light source. The liquid gathers weight, stretching into heavy droplets before separating and falling in deliberate slow motion, landing with a soft, wet impact on the shimmering chitinous surface below—a landscape of polished insectoid plates and skeletal ridges. The harsh light carves the entire image from the oppressive darkness, casting deep, sinister shadows that seem to wriggle and move with unseen, organic motion, creating an atmosphere of intensely tense and surreal beauty and horror.", + "seed": 26005, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_14/index_6566_seed_26005_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": -0.24687769780866803, + "vlm_logprob": -0.24687769780866803, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.24687769780866803, + "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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04079597815871239 + ], + "opposite_logprob_tokens": [ + -3.5407960414886475 + ], + "logprob_reward": -0.04079597815871239, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04079597815871239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4063846468925476, + 0.041394736617803574, + 0.06610140949487686, + 0.07770489901304245, + 0.15040171146392822, + 0.035797689110040665, + 0.059575922787189484, + 0.01591433212161064, + 0.11138512194156647, + 0.035339560359716415 + ], + "credit_map": [ + 0.06059593707323074, + 0.035126425325870514, + 0.05410822853446007, + 0.05622820928692818, + 0.055769018828868866, + 0.053286731243133545, + 0.057429227977991104, + 0.033840835094451904, + 0.009037842974066734, + 0.004563249181956053, + 0.005069315899163485, + 0.004963471554219723, + 0.00414434727281332, + 0.005244679283350706, + 0.00411383667960763, + 0.004257991909980774, + 0.009307853877544403, + 0.006336839869618416, + 0.015883775427937508, + 0.006889330223202705, + 0.005579374730587006, + 0.009880995377898216, + 0.007376509718596935, + 0.0048467242158949375, + 0.010632885619997978, + 0.007988330908119678, + 0.016523968428373337, + 0.010580006055533886, + 0.004620993509888649, + 0.011729220859706402, + 0.008082924410700798, + 0.007546571549028158, + 0.016151215881109238, + 0.005232676863670349, + 0.05399217829108238, + 0.00608546519652009, + 0.004284373950213194, + 0.047804657369852066, + 0.008029918186366558, + 0.008821211755275726, + 0.006959475576877594, + 0.0044720894657075405, + 0.0037101355846971273, + 0.004604840651154518, + 0.0036185933277010918, + 0.004390548914670944, + 0.005076698958873749, + 0.002965302439406514, + 0.012259980663657188, + 0.009741897694766521, + 0.007720596622675657, + 0.0065220920369029045, + 0.0038721358869224787, + 0.0068969521671533585, + 0.007043617777526379, + 0.00551864318549633, + 0.002021157182753086, + 0.0023232856765389442, + 0.003922165371477604, + 0.0014885040000081062, + 0.0007850578404031694, + 0.0012711076997220516, + 0.0029724605847150087, + 0.0011305941734462976, + 0.010538684204220772, + 0.004083551932126284, + 0.031364817172288895, + 0.004699654411524534, + 0.003926246892660856, + 0.007576187141239643, + 0.040813151746988297, + 0.008382822386920452, + 0.005818819161504507, + 0.0043047755025327206, + 0.0061245448887348175, + 0.0036489928606897593, + 0.0028989315032958984, + 0.0033908875193446875, + 0.006288597825914621, + 0.0028640085365623236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.04079597815871239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890608310699463 + ], + "opposite_logprob_tokens": [ + -1.0539060831069946 + ], + "logprob_reward": -0.42890608310699463, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42890608310699463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2675369679927826, + 0.049003224819898605, + 0.07187007367610931, + 0.059883683919906616, + 0.2006598562002182, + 0.026871951296925545, + 0.0505429282784462, + 0.047241706401109695, + 0.14843681454658508, + 0.07795274257659912 + ], + "credit_map": [ + 0.04053434729576111, + 0.030701633542776108, + 0.0514826625585556, + 0.032588716596364975, + 0.020183632150292397, + 0.034428730607032776, + 0.029604993760585785, + 0.02801225520670414, + 0.010297108441591263, + 0.005157427862286568, + 0.0048295422457158566, + 0.006724087987095118, + 0.004606358706951141, + 0.006192770320922136, + 0.005344075616449118, + 0.0058518522419035435, + 0.007199970539659262, + 0.0038907655980437994, + 0.022914879024028778, + 0.005122953560203314, + 0.004011805169284344, + 0.020354317501187325, + 0.004379760008305311, + 0.003995620645582676, + 0.011840173043310642, + 0.006604855880141258, + 0.011814451776444912, + 0.005451935343444347, + 0.00482968520373106, + 0.008266354911029339, + 0.006419049575924873, + 0.004657177720218897, + 0.02176937833428383, + 0.007175258360803127, + 0.06969980895519257, + 0.008585040457546711, + 0.006155485287308693, + 0.0645676702260971, + 0.010870419442653656, + 0.011836804449558258, + 0.0031325730960816145, + 0.0025749169290065765, + 0.005804385058581829, + 0.0030234113801270723, + 0.0019082255894318223, + 0.005424002651125193, + 0.0025363995227962732, + 0.0024680383503437042, + 0.010823705233633518, + 0.0060076359659433365, + 0.009240209124982357, + 0.0051062097772955894, + 0.003511686809360981, + 0.006285724230110645, + 0.00548208924010396, + 0.004085669759660959, + 0.01072346419095993, + 0.005066164303570986, + 0.007006978616118431, + 0.005693420767784119, + 0.0036173160187900066, + 0.004429560620337725, + 0.006342763546854258, + 0.0043620397336781025, + 0.021049417555332184, + 0.009531146846711636, + 0.04041394591331482, + 0.010162487626075745, + 0.006271274294704199, + 0.011289993301033974, + 0.03982744738459587, + 0.009891101159155369, + 0.012136436067521572, + 0.010702060535550117, + 0.009107641875743866, + 0.00967381987720728, + 0.006554850842803717, + 0.013310497626662254, + 0.011343658901751041, + 0.005123778712004423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.42890608310699463 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002206512028351426 + ], + "opposite_logprob_tokens": [ + -6.127206325531006 + ], + "logprob_reward": -0.002206512028351426, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002206512028351426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16452160477638245, + 0.12925375998020172, + 0.1585082709789276, + 0.024089615792036057, + 0.3370824158191681, + 0.03160376846790314, + 0.021189171820878983, + 0.01843920350074768, + 0.06691128015518188, + 0.04840083420276642 + ], + "credit_map": [ + 0.0249209925532341, + 0.0153757743537426, + 0.019571339711546898, + 0.02457929216325283, + 0.017130859196186066, + 0.024934014305472374, + 0.02303081937134266, + 0.014978528022766113, + 0.013791010715067387, + 0.011185713112354279, + 0.025131307542324066, + 0.017734654247760773, + 0.017331818118691444, + 0.021041782572865486, + 0.013270184397697449, + 0.009767316281795502, + 0.014686384238302708, + 0.007333966437727213, + 0.051480185240507126, + 0.011505352333188057, + 0.006624277215451002, + 0.050832364708185196, + 0.007379879709333181, + 0.00866587646305561, + 0.0031088890973478556, + 0.0026272612158209085, + 0.0024456409737467766, + 0.003041215008124709, + 0.0025297796819359064, + 0.002980193356052041, + 0.004472729284316301, + 0.00288390857167542, + 0.037662141025066376, + 0.011843197979032993, + 0.11725685745477676, + 0.014159660786390305, + 0.010382390581071377, + 0.10704468190670013, + 0.018870066851377487, + 0.019863460212945938, + 0.004630438517779112, + 0.004042019601911306, + 0.0029526439029723406, + 0.0037825549952685833, + 0.003945020027458668, + 0.0033940107095986605, + 0.005332215689122677, + 0.0035248706117272377, + 0.0027499902062118053, + 0.002620856976136565, + 0.003261592471972108, + 0.0027328012511134148, + 0.0021720086224377155, + 0.0026077558286488056, + 0.0032197418622672558, + 0.001824426930397749, + 0.0039754402823746204, + 0.0033364014234393835, + 0.00217896094545722, + 0.002204172546043992, + 0.0010905187809839845, + 0.0015788446180522442, + 0.002289022086188197, + 0.0017858454957604408, + 0.006041663698852062, + 0.0036781488452106714, + 0.01762593537569046, + 0.0035386732779443264, + 0.0025711613707244396, + 0.005127400625497103, + 0.023543670773506165, + 0.004784632474184036, + 0.006679882295429707, + 0.006172367371618748, + 0.009332778863608837, + 0.006012522149831057, + 0.0035927060525864363, + 0.004434189759194851, + 0.009032866917550564, + 0.003143526380881667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.002206512028351426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463838828727603 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.00463838828727603, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463838828727603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2669133245944977, + 0.040658436715602875, + 0.06365156173706055, + 0.019779853522777557, + 0.4934898614883423, + 0.032621242105960846, + 0.017005471512675285, + 0.01617211103439331, + 0.0149361751973629, + 0.03477201983332634 + ], + "credit_map": [ + 0.035820119082927704, + 0.023423032835125923, + 0.03117656148970127, + 0.03961002826690674, + 0.03473301976919174, + 0.0348777249455452, + 0.040446698665618896, + 0.026826102286577225, + 0.008018088527023792, + 0.00371175748296082, + 0.005227399058640003, + 0.00540328910574317, + 0.003981165122240782, + 0.00496634840965271, + 0.005318331532180309, + 0.0040320539847016335, + 0.006239479873329401, + 0.0037060510367155075, + 0.020653948187828064, + 0.005559070035815239, + 0.0029272797983139753, + 0.016787899658083916, + 0.004325210582464933, + 0.0034526127856224775, + 0.0030813859775662422, + 0.0018554521957412362, + 0.0037845089100301266, + 0.0025057990569621325, + 0.00131507171317935, + 0.0038444511592388153, + 0.0018139833118766546, + 0.0015791975893080235, + 0.05339127779006958, + 0.01753818430006504, + 0.17491121590137482, + 0.01955525018274784, + 0.01390106976032257, + 0.15913377702236176, + 0.02664508856832981, + 0.028413955122232437, + 0.004807293880730867, + 0.003290456021204591, + 0.007652109023183584, + 0.0027969370130449533, + 0.002174608176574111, + 0.005638028960675001, + 0.003552812384441495, + 0.0027089929208159447, + 0.002795029431581497, + 0.0023769999388605356, + 0.0022823598701506853, + 0.0019238145323470235, + 0.0014640951994806528, + 0.002225131494924426, + 0.002576082246378064, + 0.001361955888569355, + 0.0025034367572516203, + 0.0019624982960522175, + 0.003369782352820039, + 0.0015811269404366612, + 0.0009951511165127158, + 0.001332735293544829, + 0.003344586817547679, + 0.0010827919468283653, + 0.0014447805006057024, + 0.0008591460064053535, + 0.003711304860189557, + 0.000943024002481252, + 0.000571130309253931, + 0.0014179687714204192, + 0.004886332433670759, + 0.001102485810406506, + 0.004976161755621433, + 0.003832352114841342, + 0.0072633554227650166, + 0.003511689370498061, + 0.0027902538422495127, + 0.0032564420253038406, + 0.006567259784787893, + 0.0025745013263076544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.00463838828727603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a perfect, waxy cylinder of vibrant crimson lipstick extending from a biomechanical orifice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the orifice a grotesque amalgamation of cold, ridged metal and gleaming, fleshy tissue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single, stark overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578415274620056 + ], + "opposite_logprob_tokens": [ + -0.6328415274620056 + ], + "logprob_reward": -0.7578415274620056, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578415274620056, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3807883858680725, + 0.04395400732755661, + 0.15190668404102325, + 0.07508305460214615, + 0.12439686805009842, + 0.06257365643978119, + 0.09531319886445999, + 0.020383723080158234, + 0.028159713372588158, + 0.017440754920244217 + ], + "credit_map": [ + 0.04621788114309311, + 0.035107050091028214, + 0.0442294180393219, + 0.05475477874279022, + 0.054787274450063705, + 0.05197516083717346, + 0.06341880559921265, + 0.030298003926873207, + 0.006197637412697077, + 0.005589607637375593, + 0.0069659557193517685, + 0.005290219560265541, + 0.005926328711211681, + 0.005452537909150124, + 0.004667147062718868, + 0.0038645751774311066, + 0.013144833035767078, + 0.007395057938992977, + 0.05187491327524185, + 0.011314337141811848, + 0.007221931125968695, + 0.046307213604450226, + 0.007108982652425766, + 0.007539415266364813, + 0.008675103075802326, + 0.004511997103691101, + 0.024458179250359535, + 0.006455738563090563, + 0.003597977804020047, + 0.018812505528330803, + 0.005277655087411404, + 0.003293896559625864, + 0.015079728327691555, + 0.006680980324745178, + 0.04069660231471062, + 0.006851028650999069, + 0.005879296455532312, + 0.03283427283167839, + 0.0095962043851614, + 0.0067787510342895985, + 0.01152875181287527, + 0.0057223704643547535, + 0.011321303434669971, + 0.007277921307832003, + 0.004957345314323902, + 0.00823344849050045, + 0.008527562953531742, + 0.0050049470737576485, + 0.016413072124123573, + 0.014852805994451046, + 0.012030989862978458, + 0.0107145756483078, + 0.007897933013737202, + 0.01136730145663023, + 0.013588234782218933, + 0.008448290638625622, + 0.003193554235622287, + 0.003449689131230116, + 0.002263779053464532, + 0.0033276535104960203, + 0.001442630309611559, + 0.0022641278337687254, + 0.00255069718696177, + 0.0018915921682491899, + 0.0028516617603600025, + 0.0017970167100429535, + 0.007798357866704464, + 0.0018586302176117897, + 0.0010969258146360517, + 0.0020424469839781523, + 0.008711203001439571, + 0.002003470668569207, + 0.0038170262705534697, + 0.0018435027450323105, + 0.002189630875363946, + 0.0016750235809013247, + 0.0013639099197462201, + 0.0018327138386666775, + 0.0030350724700838327, + 0.0016838758019730449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.7578415274620056 + } + ], + "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/0000122/group_15/index_tensor([4898, 4898])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_15/index_tensor([4898, 4898])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..88377ae4b509a5b0f881c9443348412ba2ed6a99 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_15/index_tensor([4898, 4898])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot captures a scene rendered in the stark, high-contrast style of black-and-white ukiyo-e animation. From an unseen source below the frame, countless thick, fibrous sheets of wakashi-style paper rise in a powerful, swirling vortex. Each strip is meticulously crafted, featuring a rough, tactile texture with visible fibers and clean, perforated edges. They are embellished with intricate traditional motifs—some depicting Hokusai's waving waves, others the delicate lines of swirling clouds—all executed with the bold, calligraphic outlines characteristic of woodblock prints. As the papers tumble and flutter, they spiral upward, forming a mesmerizing, chaotic yet elegant vortex that dominates the center of the composition. Sharp, directed light pierces the darkness, catching the edges and flat surfaces of the paper, creating bright white highlights that contrast strikingly with the empty depths, inky darkness, and absolute blackness of the background, creating a hypnotic and powerful visual spectacle brimming with energy.", + "seed": 25139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_15/index_4898_seed_25139_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": -0.19447390628047287, + "vlm_logprob": -0.19447390628047287, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.19447390628047287, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.9227907657623291 + ], + "opposite_logprob_tokens": [ + -0.5477907657623291 + ], + "logprob_reward": -0.9227907657623291, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9227907657623291, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05038778856396675, + 0.07943572849035263, + 0.018863888457417488, + 0.010731526650488377, + 0.01147217396646738, + 0.025916239246726036, + 0.0848473384976387, + 0.6897807121276855, + 0.012067879550158978, + 0.016496630385518074 + ], + "credit_map": [ + 0.006438647862523794, + 0.006122767459601164, + 0.006944634020328522, + 0.0033763726241886616, + 0.0037041332107037306, + 0.012176449410617352, + 0.0032207383774220943, + 0.008404048159718513, + 0.00852944701910019, + 0.011575097218155861, + 0.021265016868710518, + 0.00699438713490963, + 0.007133654318749905, + 0.01522899977862835, + 0.004308764822781086, + 0.004400363191962242, + 0.0028593833558261395, + 0.0030577187426388264, + 0.003586395177990198, + 0.002187136560678482, + 0.002197951776906848, + 0.002036964986473322, + 0.0015243407106027007, + 0.0014139984268695116, + 0.0009900485165417194, + 0.0014349173288792372, + 0.001516341813839972, + 0.002057243837043643, + 0.001220830949023366, + 0.001329365884885192, + 0.0008701870683580637, + 0.0013125916011631489, + 0.0018315043998882174, + 0.001855748356319964, + 0.001614827779121697, + 0.0009308255393989384, + 0.0016263172728940845, + 0.0013020704500377178, + 0.0010806771460920572, + 0.0012302028480917215, + 0.0032997906673699617, + 0.0033289967104792595, + 0.0034406837075948715, + 0.0021053405944257975, + 0.005264598410576582, + 0.0038747675716876984, + 0.0019092941656708717, + 0.0026927676517516375, + 0.010677147656679153, + 0.011308196000754833, + 0.02600259706377983, + 0.007824143394827843, + 0.006863194052129984, + 0.00790442805737257, + 0.00876929983496666, + 0.005498331971466541, + 0.04249969869852066, + 0.024372512474656105, + 0.04475494846701622, + 0.25733649730682373, + 0.02292148768901825, + 0.18814679980278015, + 0.07284186780452728, + 0.03690692037343979, + 0.0010765540646389127, + 0.0012490673689171672, + 0.0008372210431843996, + 0.0011466533178463578, + 0.0023040813393890858, + 0.0020554650109261274, + 0.0017865481786429882, + 0.0016122900415211916, + 0.00273330882191658, + 0.001474014250561595, + 0.0018143628258258104, + 0.0016610667807981372, + 0.002835250459611416, + 0.0022669758182018995, + 0.002087437780573964, + 0.0016242127167060971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.9227907657623291 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.014250772073864937 + ], + "opposite_logprob_tokens": [ + -4.264250755310059 + ], + "logprob_reward": -0.014250772073864937, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014250772073864937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09312376379966736, + 0.07120754569768906, + 0.020782053470611572, + 0.026674909517169, + 0.04438823089003563, + 0.022076958790421486, + 0.14920412003993988, + 0.526710569858551, + 0.02989041619002819, + 0.015941448509693146 + ], + "credit_map": [ + 0.01274072378873825, + 0.012156829237937927, + 0.019254211336374283, + 0.006238885689526796, + 0.007411391474306583, + 0.01600687764585018, + 0.005630286876112223, + 0.013684548437595367, + 0.009269149973988533, + 0.015226495452225208, + 0.01147338468581438, + 0.005831676535308361, + 0.004093071911484003, + 0.009066332131624222, + 0.003991654142737389, + 0.012255778536200523, + 0.002541374647989869, + 0.0028525888919830322, + 0.004807481542229652, + 0.002234111074358225, + 0.002989483065903187, + 0.0020145925227552652, + 0.0019364425679668784, + 0.0014059790410101414, + 0.0032806703820824623, + 0.003613434499129653, + 0.002478760899975896, + 0.005211466923356056, + 0.003199446015059948, + 0.003468260867521167, + 0.0028926378581672907, + 0.0025302341673523188, + 0.005460840649902821, + 0.006545749027282, + 0.005767307709902525, + 0.004347293172031641, + 0.005790514871478081, + 0.004343537613749504, + 0.005227817688137293, + 0.006905166897922754, + 0.0026692405808717012, + 0.0026745053473860025, + 0.003315894864499569, + 0.00239654746837914, + 0.0035929169971495867, + 0.002627636305987835, + 0.0017537655076012015, + 0.0030464534647762775, + 0.014333120547235012, + 0.01633915677666664, + 0.04890856146812439, + 0.03427810221910477, + 0.008928229101002216, + 0.007104964926838875, + 0.014033777639269829, + 0.005278197582811117, + 0.057350803166627884, + 0.01941290870308876, + 0.03231251239776611, + 0.2081139087677002, + 0.014409482479095459, + 0.11506296694278717, + 0.053089819848537445, + 0.026958148926496506, + 0.0038455366156995296, + 0.003068558406084776, + 0.0019444954814389348, + 0.003765741130337119, + 0.004431576002389193, + 0.005013319198042154, + 0.002853676676750183, + 0.004967513959854841, + 0.0028385340701788664, + 0.0010927903931587934, + 0.0016279794508591294, + 0.0012512587709352374, + 0.0033176287543028593, + 0.0021565298084169626, + 0.0020892515312880278, + 0.0015674765454605222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.014250772073864937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.01254853792488575 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.01254853792488575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254853792488575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20935791730880737, + 0.06203293427824974, + 0.033487722277641296, + 0.016011834144592285, + 0.03130694106221199, + 0.036897461861371994, + 0.19844387471675873, + 0.3556884527206421, + 0.020880278199911118, + 0.03589257225394249 + ], + "credit_map": [ + 0.010891989804804325, + 0.012506918981671333, + 0.032309237867593765, + 0.005380410235375166, + 0.007037756033241749, + 0.04581293836236, + 0.006528941448777914, + 0.0888897180557251, + 0.0057326313108205795, + 0.01265905424952507, + 0.015238732099533081, + 0.004803474526852369, + 0.004313332960009575, + 0.00967903807759285, + 0.0032335000578314066, + 0.0063731735572218895, + 0.004772798158228397, + 0.004998786840587854, + 0.004947284236550331, + 0.004512866027653217, + 0.0034769917838275433, + 0.004098684526979923, + 0.002822746755555272, + 0.0038575618527829647, + 0.001876683789305389, + 0.00216243346221745, + 0.002066875109449029, + 0.0027825243305414915, + 0.0017124818405136466, + 0.0022361434530466795, + 0.0017419570358470082, + 0.0014327351236715913, + 0.004067947622388601, + 0.003430950455367565, + 0.00343207735568285, + 0.0023472660686820745, + 0.0053003085777163506, + 0.003270503832027316, + 0.0036015817895531654, + 0.00585630489513278, + 0.00529405428096652, + 0.004250658676028252, + 0.005036004353314638, + 0.0034439207520335913, + 0.006718249525874853, + 0.004742467310279608, + 0.0027566631324589252, + 0.004655445460230112, + 0.017328066751360893, + 0.021228056401014328, + 0.05899330973625183, + 0.05108208581805229, + 0.013080403208732605, + 0.009946483187377453, + 0.019961658865213394, + 0.006823815405368805, + 0.03393237665295601, + 0.011953824199736118, + 0.017977969720959663, + 0.09240899980068207, + 0.015132645145058632, + 0.12451587617397308, + 0.0407014898955822, + 0.0190652497112751, + 0.002438540104776621, + 0.0024031621869653463, + 0.0016558613860979676, + 0.0020752125419676304, + 0.0035488642752170563, + 0.003625721437856555, + 0.002300360007211566, + 0.0028325573075562716, + 0.007871284149587154, + 0.004048551898449659, + 0.003631215775385499, + 0.003172324039041996, + 0.006086465436965227, + 0.003806982422247529, + 0.0037050156388431787, + 0.003570731496438384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01254853792488575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.020580556243658066 + ], + "opposite_logprob_tokens": [ + -3.895580530166626 + ], + "logprob_reward": -0.020580556243658066, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580556243658066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05273305997252464, + 0.053791265934705734, + 0.011962969787418842, + 0.00536282267421484, + 0.027294237166643143, + 0.028276721015572548, + 0.062439098954200745, + 0.7387136816978455, + 0.00923518929630518, + 0.01019099447876215 + ], + "credit_map": [ + 0.003352375468239188, + 0.0042992038652300835, + 0.007573081646114588, + 0.0021936120465397835, + 0.0022955460008233786, + 0.013742905110120773, + 0.0020330213010311127, + 0.017243312671780586, + 0.005468954332172871, + 0.013843894004821777, + 0.01122080534696579, + 0.004223334603011608, + 0.003074699081480503, + 0.0069800373166799545, + 0.0027381035033613443, + 0.006241433322429657, + 0.0014818438794463873, + 0.0015391111373901367, + 0.0026112431660294533, + 0.0011031413450837135, + 0.0016308851772919297, + 0.001491038710810244, + 0.0011364369420334697, + 0.0009692682651802897, + 0.0004987847642041743, + 0.0008434616611339152, + 0.0006023583700880408, + 0.0009982276242226362, + 0.0006287505384534597, + 0.0006306211580522358, + 0.0006500673480331898, + 0.0005105509189888835, + 0.002934287767857313, + 0.0035535525530576706, + 0.003012804314494133, + 0.0025370384100824594, + 0.004152202047407627, + 0.0027410860639065504, + 0.004082840401679277, + 0.004280421882867813, + 0.003899218048900366, + 0.004400968551635742, + 0.0034980790223926306, + 0.0023679693695157766, + 0.0067713153548538685, + 0.0033482750877738, + 0.0018534476403146982, + 0.002137446776032448, + 0.006414604838937521, + 0.0065820529125630856, + 0.020805126056075096, + 0.014344817027449608, + 0.0036919733975082636, + 0.003142824163660407, + 0.005477097816765308, + 0.001980600645765662, + 0.08443667739629745, + 0.02830093540251255, + 0.04731685668230057, + 0.2936430871486664, + 0.019345704466104507, + 0.15515466034412384, + 0.07300686836242676, + 0.03750885650515556, + 0.0010247220052406192, + 0.0009804638102650642, + 0.0006847006734460592, + 0.001012478838674724, + 0.0013431147672235966, + 0.0016188973095268011, + 0.0008822462987154722, + 0.0016885646618902683, + 0.0018367518205195665, + 0.0012985765933990479, + 0.0011586664477363229, + 0.0010809616651386023, + 0.0015048052882775664, + 0.0011721240589395165, + 0.0011454059276729822, + 0.0009937019785866141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.020580556243658066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.002198899397626519 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198899397626519, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198899397626519, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049770787358284, + 0.025532782077789307, + 0.04503418877720833, + 0.011876707896590233, + 0.023382531479001045, + 0.07642006129026413, + 0.041691847145557404, + 0.6716476678848267, + 0.01360343862324953, + 0.04103996604681015 + ], + "credit_map": [ + 0.0069766114465892315, + 0.008090433664619923, + 0.008413733914494514, + 0.0036432803608477116, + 0.0044794450514018536, + 0.008365239016711712, + 0.0032181439455598593, + 0.006583901587873697, + 0.002404589904472232, + 0.00662181805819273, + 0.0046133194118738174, + 0.0018352949991822243, + 0.0022097835317254066, + 0.003824882907792926, + 0.002242541406303644, + 0.00178055209107697, + 0.006277893669903278, + 0.0054803453385829926, + 0.0065669044852256775, + 0.004765847697854042, + 0.006042021792382002, + 0.007968337275087833, + 0.0036879603285342455, + 0.004244879353791475, + 0.0014755369629710913, + 0.001211710274219513, + 0.0014760990161448717, + 0.0026626617182046175, + 0.0013223992427811027, + 0.0014526781160384417, + 0.0009282482787966728, + 0.0013473746366798878, + 0.003510846057906747, + 0.00244231428951025, + 0.003232181305065751, + 0.0022072307765483856, + 0.0030193135607987642, + 0.0027534877881407738, + 0.002647427609190345, + 0.003569730557501316, + 0.009896162897348404, + 0.011531948111951351, + 0.011035766452550888, + 0.005734340287744999, + 0.016230886802077293, + 0.009364798665046692, + 0.0050614927895367146, + 0.007564669009298086, + 0.004977411590516567, + 0.004556083120405674, + 0.0047019547782838345, + 0.009736545383930206, + 0.004595489241182804, + 0.004355547949671745, + 0.005330638960003853, + 0.003438178449869156, + 0.15632009506225586, + 0.03448085859417915, + 0.031235728412866592, + 0.05021621659398079, + 0.03276217728853226, + 0.25542357563972473, + 0.07590863108634949, + 0.03530031442642212, + 0.0015422001015394926, + 0.001671023666858673, + 0.0011509531177580357, + 0.0014333039289340377, + 0.0020864983089268208, + 0.0018284410471096635, + 0.0014212654205039144, + 0.0024697526823729277, + 0.008316903375089169, + 0.004517024382948875, + 0.004369756206870079, + 0.003789337119087577, + 0.007554049603641033, + 0.004540450405329466, + 0.004170683212578297, + 0.0037817605771124363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.002198899397626519 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot captures a scene rendered in the stark, high-contrast style of black-and-white ukiyo-e animation. From an unseen source below the frame, countless thick, fibrous sheets of wakashi-style paper rise in a powerful, swirling vortex. Each strip is meticulously crafted, featuring a rough, tactile texture with visible fibers and clean, perforated edges. They are embellished with intricate traditional motifs—some depicting Hokusai's waving waves, others the delicate lines of swirling clouds—all executed with the bold, calligraphic outlines characteristic of woodblock prints. As the papers tumble and flutter, they spiral upward, forming a mesmerizing, chaotic yet elegant vortex that dominates the center of the composition. Sharp, directed light pierces the darkness, catching the edges and flat surfaces of the paper, creating bright white highlights that contrast strikingly with the empty depths, inky darkness, and absolute blackness of the background, creating a hypnotic and powerful visual spectacle brimming with energy.", + "seed": 25140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_15/index_4898_seed_25140_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": -0.17766580851748587, + "vlm_logprob": -0.17766580851748587, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.17766580851748587, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.8575851321220398 + ], + "opposite_logprob_tokens": [ + -0.6075851321220398 + ], + "logprob_reward": -0.8575851321220398, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8575851321220398, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1470516175031662, + 0.10508935898542404, + 0.07327315211296082, + 0.1069255992770195, + 0.22810058295726776, + 0.12009426206350327, + 0.04861103370785713, + 0.12484773993492126, + 0.014249485917389393, + 0.031757161021232605 + ], + "credit_map": [ + 0.02096128650009632, + 0.027109220623970032, + 0.026564063504338264, + 0.018626026809215546, + 0.014278138056397438, + 0.013261622749269009, + 0.011819400824606419, + 0.014431841671466827, + 0.010016964748501778, + 0.007042113225907087, + 0.013149022124707699, + 0.009146216325461864, + 0.011333253234624863, + 0.020546814426779747, + 0.017706817016005516, + 0.01614815928041935, + 0.00839290302246809, + 0.008127513341605663, + 0.011461257934570312, + 0.007140050176531076, + 0.010042168200016022, + 0.010400258004665375, + 0.00799958873540163, + 0.00970941036939621, + 0.012839212082326412, + 0.0074263992719352245, + 0.010724145919084549, + 0.009277202188968658, + 0.011155580170452595, + 0.023965902626514435, + 0.009801375679671764, + 0.021735772490501404, + 0.03672663867473602, + 0.022160017862915993, + 0.03893790394067764, + 0.018035179004073143, + 0.016301875934004784, + 0.028463199734687805, + 0.027549469843506813, + 0.03992629051208496, + 0.0048197731375694275, + 0.003844327526167035, + 0.013894256204366684, + 0.004013245925307274, + 0.003602269571274519, + 0.047851912677288055, + 0.006816104520112276, + 0.035252366214990616, + 0.004959144629538059, + 0.0033414196223020554, + 0.00392587436363101, + 0.004615440033376217, + 0.008931676857173443, + 0.008740375749766827, + 0.006369483191519976, + 0.007727616000920534, + 0.00998922809958458, + 0.015312687493860722, + 0.014321125112473965, + 0.012216060422360897, + 0.026536108925938606, + 0.013962644152343273, + 0.015368247404694557, + 0.017141640186309814, + 0.0016394075937569141, + 0.0025460398755967617, + 0.0013395071728155017, + 0.0021339652594178915, + 0.001565241371281445, + 0.002170370891690254, + 0.001246497849933803, + 0.0016084554372355342, + 0.0040329028852283955, + 0.004141152370721102, + 0.0031742968130856752, + 0.0034220858942717314, + 0.003374992636963725, + 0.003226082306355238, + 0.0059011406265199184, + 0.004484505858272314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.8575851321220398 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.011127321980893612 + ], + "opposite_logprob_tokens": [ + -4.511127471923828 + ], + "logprob_reward": -0.011127321980893612, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011127321980893612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20993086695671082, + 0.03789522498846054, + 0.11723333597183228, + 0.04196310415863991, + 0.03177359327673912, + 0.3999035954475403, + 0.032492078840732574, + 0.03091459348797798, + 0.028346549719572067, + 0.06954705715179443 + ], + "credit_map": [ + 0.02757401205599308, + 0.043942805379629135, + 0.030650107190012932, + 0.021472828462719917, + 0.022714750841259956, + 0.0206892229616642, + 0.01709909550845623, + 0.025788061320781708, + 0.01016279961913824, + 0.002649649977684021, + 0.004222912713885307, + 0.003236873773857951, + 0.00402624998241663, + 0.0046921540051698685, + 0.004137760493904352, + 0.00476682698354125, + 0.01567835547029972, + 0.015236063860356808, + 0.016478752717375755, + 0.00780002074316144, + 0.015362023375928402, + 0.019658192992210388, + 0.01360565796494484, + 0.013414276763796806, + 0.005940630100667477, + 0.003828268963843584, + 0.006257795728743076, + 0.004989159759134054, + 0.005543114617466927, + 0.006667104084044695, + 0.0037386559415608644, + 0.0049983784556388855, + 0.004300461150705814, + 0.0027012433856725693, + 0.004057294223457575, + 0.003911581356078386, + 0.0030305737163871527, + 0.002778974361717701, + 0.0052304333075881, + 0.00576303293928504, + 0.008956770412623882, + 0.008272230625152588, + 0.033286549150943756, + 0.007514690048992634, + 0.009593005292117596, + 0.17290377616882324, + 0.02089347504079342, + 0.13848312199115753, + 0.0025698132812976837, + 0.0023524572607129812, + 0.0029695930425077677, + 0.0025666523724794388, + 0.0048262025229632854, + 0.007194270379841328, + 0.004166526719927788, + 0.005846563261002302, + 0.0028006145730614662, + 0.004169194493442774, + 0.0041896505281329155, + 0.003218705765902996, + 0.00515835452824831, + 0.003190142335370183, + 0.0037664955016225576, + 0.00442143902182579, + 0.0037279778625816107, + 0.00376027449965477, + 0.00285469857044518, + 0.002929788315668702, + 0.004224217962473631, + 0.004618930630385876, + 0.0026524383574724197, + 0.00357822491787374, + 0.011896392330527306, + 0.00956256128847599, + 0.011516518890857697, + 0.007592498324811459, + 0.0055361464619636536, + 0.005178268067538738, + 0.008984214626252651, + 0.009280463680624962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.011127321980893612 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.0076385317370295525 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.0076385317370295525, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076385317370295525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2906877100467682, + 0.04416089877486229, + 0.05952068418264389, + 0.1311434507369995, + 0.0798846110701561, + 0.20493640005588531, + 0.02567882649600506, + 0.0829358845949173, + 0.010554498992860317, + 0.0704970434308052 + ], + "credit_map": [ + 0.03163446858525276, + 0.04628484696149826, + 0.0504462756216526, + 0.0287324208766222, + 0.03559614717960358, + 0.02998093143105507, + 0.036345064640045166, + 0.03166753798723221, + 0.0077505954541265965, + 0.0030027765315026045, + 0.0049622803926467896, + 0.0047615040093660355, + 0.0046255989000201225, + 0.0057078092359006405, + 0.007275234442204237, + 0.0060750944539904594, + 0.006339586339890957, + 0.006180629599839449, + 0.008454049937427044, + 0.00516301766037941, + 0.009140211157500744, + 0.010172639973461628, + 0.007446975912898779, + 0.006623569875955582, + 0.01917952299118042, + 0.0091174878180027, + 0.008238056674599648, + 0.008923955261707306, + 0.008403057232499123, + 0.04274222254753113, + 0.008872302249073982, + 0.025666840374469757, + 0.013350313529372215, + 0.007947264239192009, + 0.014465602114796638, + 0.006758043542504311, + 0.005797976162284613, + 0.009490806609392166, + 0.009920751675963402, + 0.01215384528040886, + 0.0037544909864664078, + 0.006190907210111618, + 0.0060515557415783405, + 0.004451337270438671, + 0.007680883165448904, + 0.08626141399145126, + 0.011713031679391861, + 0.07883276045322418, + 0.0023459892254322767, + 0.0026616831310093403, + 0.0018736664205789566, + 0.0021677748300135136, + 0.002631772542372346, + 0.004450291395187378, + 0.00490083871409297, + 0.004646809305995703, + 0.007242094725370407, + 0.009971654042601585, + 0.008978004567325115, + 0.006755848880857229, + 0.019745148718357086, + 0.009058531373739243, + 0.009909378364682198, + 0.011275217868387699, + 0.0015807049348950386, + 0.0015445440076291561, + 0.001105677685700357, + 0.0011114340741187334, + 0.0011806307593360543, + 0.0017004563705995679, + 0.0009275255142711103, + 0.001403524773195386, + 0.009143324568867683, + 0.010187179781496525, + 0.008885621093213558, + 0.007601127028465271, + 0.00770882423967123, + 0.00956194568425417, + 0.008563203737139702, + 0.008845816366374493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0076385317370295525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.00978010892868042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.00978010892868042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978010892868042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24978739023208618, + 0.07799094915390015, + 0.07367000728845596, + 0.15029458701610565, + 0.04641639441251755, + 0.0966828465461731, + 0.13303899765014648, + 0.06382244825363159, + 0.02368154190480709, + 0.0846148431301117 + ], + "credit_map": [ + 0.03904655948281288, + 0.04183906689286232, + 0.0673070102930069, + 0.015849454328417778, + 0.02018028125166893, + 0.027014359831809998, + 0.01944209262728691, + 0.019108563661575317, + 0.010135388001799583, + 0.0050481087528169155, + 0.010260246694087982, + 0.006621731910854578, + 0.008443229831755161, + 0.012556459754705429, + 0.016101963818073273, + 0.008823814801871777, + 0.010018201544880867, + 0.011237725615501404, + 0.009147951379418373, + 0.0063409097492694855, + 0.01061909832060337, + 0.008608792908489704, + 0.009143846109509468, + 0.00855348166078329, + 0.0202520452439785, + 0.015578225255012512, + 0.014042302034795284, + 0.01374910119920969, + 0.017707595601677895, + 0.03463452681899071, + 0.014958626590669155, + 0.019372163340449333, + 0.00746952323243022, + 0.004128228407353163, + 0.00629303976893425, + 0.005762223154306412, + 0.004346766974776983, + 0.003742812667042017, + 0.006609294563531876, + 0.008064505644142628, + 0.003471585689112544, + 0.00362685089930892, + 0.014219610951840878, + 0.005510207265615463, + 0.004869971890002489, + 0.035829272121191025, + 0.005521331913769245, + 0.02363402768969536, + 0.00969913974404335, + 0.00954032689332962, + 0.011967543512582779, + 0.01058085449039936, + 0.016489380970597267, + 0.032503921538591385, + 0.01642724499106407, + 0.025830578058958054, + 0.005408314522355795, + 0.007070496212691069, + 0.008016190491616726, + 0.004024587105959654, + 0.015433129854500294, + 0.007658730261027813, + 0.006689909845590591, + 0.009521088562905788, + 0.004254046827554703, + 0.0025392170064151287, + 0.003406299976631999, + 0.0029582250863313675, + 0.002182853175327182, + 0.003806863445788622, + 0.0018381265690550208, + 0.0026959103997796774, + 0.009303001686930656, + 0.013229010626673698, + 0.009926127269864082, + 0.010943785309791565, + 0.010072861798107624, + 0.006535833701491356, + 0.01499138679355383, + 0.009612832218408585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.00978010892868042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.002197947818785906 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.002197947818785906, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197947818785906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16194626688957214, + 0.03442266583442688, + 0.06044641137123108, + 0.07285120338201523, + 0.12528091669082642, + 0.2542797327041626, + 0.046050433069467545, + 0.16255128383636475, + 0.016297675669193268, + 0.06587349623441696 + ], + "credit_map": [ + 0.024912912398576736, + 0.03398270532488823, + 0.03110426850616932, + 0.01587538793683052, + 0.01428303960710764, + 0.015143980272114277, + 0.012992673553526402, + 0.013651295565068722, + 0.005368795711547136, + 0.002648154739290476, + 0.005406391341239214, + 0.0031635514460504055, + 0.002878729719668627, + 0.0043251668103039265, + 0.0061654276214540005, + 0.00446644751355052, + 0.005886680446565151, + 0.006606690119951963, + 0.0073140207678079605, + 0.00849942397326231, + 0.008584970608353615, + 0.009422698058187962, + 0.006954454816877842, + 0.007177469786256552, + 0.009819109924137592, + 0.004648052621632814, + 0.00524035794660449, + 0.00528696458786726, + 0.006072141695767641, + 0.02396990731358528, + 0.004597163293510675, + 0.013217498548328876, + 0.016638195142149925, + 0.015614164061844349, + 0.021492809057235718, + 0.010909552685916424, + 0.010428068228065968, + 0.01408495381474495, + 0.01642533205449581, + 0.0196878332644701, + 0.005460174288600683, + 0.006718121934682131, + 0.006294519640505314, + 0.0051012891344726086, + 0.006543013732880354, + 0.11016614735126495, + 0.014047031290829182, + 0.09994940459728241, + 0.003740208223462105, + 0.002750811632722616, + 0.0035943996626883745, + 0.003710318822413683, + 0.007577159442007542, + 0.009846840985119343, + 0.006172143388539553, + 0.008658546954393387, + 0.013372644782066345, + 0.01799304597079754, + 0.01737581193447113, + 0.014313819818198681, + 0.038315549492836, + 0.01852289028465748, + 0.020116405561566353, + 0.02254110388457775, + 0.001791950548067689, + 0.0018647843971848488, + 0.0016149728326126933, + 0.0027562431059777737, + 0.0018368142191320658, + 0.0032988230232149363, + 0.0013460819609463215, + 0.0017880036029964685, + 0.011578257195651531, + 0.010530667379498482, + 0.008126686327159405, + 0.005847541615366936, + 0.006633484270423651, + 0.005633203312754631, + 0.007280927151441574, + 0.010242728516459465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.002197947818785906 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot captures a scene rendered in the stark, high-contrast style of black-and-white ukiyo-e animation. From an unseen source below the frame, countless thick, fibrous sheets of wakashi-style paper rise in a powerful, swirling vortex. Each strip is meticulously crafted, featuring a rough, tactile texture with visible fibers and clean, perforated edges. They are embellished with intricate traditional motifs—some depicting Hokusai's waving waves, others the delicate lines of swirling clouds—all executed with the bold, calligraphic outlines characteristic of woodblock prints. As the papers tumble and flutter, they spiral upward, forming a mesmerizing, chaotic yet elegant vortex that dominates the center of the composition. Sharp, directed light pierces the darkness, catching the edges and flat surfaces of the paper, creating bright white highlights that contrast strikingly with the empty depths, inky darkness, and absolute blackness of the background, creating a hypnotic and powerful visual spectacle brimming with energy.", + "seed": 25141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_15/index_4898_seed_25141_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": -0.32884335052222013, + "vlm_logprob": -0.32884335052222013, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.32884335052222013, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -1.6193338632583618 + ], + "opposite_logprob_tokens": [ + -0.24433381855487823 + ], + "logprob_reward": -1.6193338632583618, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6193338632583618, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.036338914185762405, + 0.035825636237859726, + 0.037217702716588974, + 0.25514915585517883, + 0.15246373414993286, + 0.08878610283136368, + 0.0728733167052269, + 0.048283010721206665, + 0.2546544671058655, + 0.018407927826046944 + ], + "credit_map": [ + 0.006355784833431244, + 0.005078090820461512, + 0.0035644781310111284, + 0.003562365425750613, + 0.005010912660509348, + 0.0045821787789464, + 0.003877764567732811, + 0.00430733896791935, + 0.003986385650932789, + 0.005986204370856285, + 0.004866224247962236, + 0.004867675714194775, + 0.004503670148551464, + 0.005784689914435148, + 0.003231648588553071, + 0.0025991362053900957, + 0.0036611538380384445, + 0.0033951387740671635, + 0.002577204257249832, + 0.004484040196985006, + 0.004974513780325651, + 0.005869449116289616, + 0.0054428610019385815, + 0.006813338957726955, + 0.032013632357120514, + 0.0448518767952919, + 0.021459609270095825, + 0.02910514734685421, + 0.029757032170891762, + 0.042140159755945206, + 0.021889522671699524, + 0.03393219783902168, + 0.019620368257164955, + 0.02074630931019783, + 0.018778491765260696, + 0.01792183518409729, + 0.014050091616809368, + 0.02543475478887558, + 0.007931596599519253, + 0.027980288490653038, + 0.010124066844582558, + 0.008831411600112915, + 0.008205076679587364, + 0.009694482199847698, + 0.010271226055920124, + 0.015303199179470539, + 0.01121150329709053, + 0.015145138837397099, + 0.013580542989075184, + 0.005117596127092838, + 0.011473453603684902, + 0.013905370607972145, + 0.0038421391509473324, + 0.005739493761211634, + 0.00940422061830759, + 0.009810498915612698, + 0.006014582701027393, + 0.004653762560337782, + 0.006859966553747654, + 0.007432363927364349, + 0.0045107100158929825, + 0.003820850746706128, + 0.006558967288583517, + 0.008431804366409779, + 0.020525289699435234, + 0.022024022415280342, + 0.05868523567914963, + 0.025766178965568542, + 0.012642593123018742, + 0.0229191817343235, + 0.029705174267292023, + 0.06238679587841034, + 0.0025870988611131907, + 0.0015346015570685267, + 0.002917597070336342, + 0.0022919620387256145, + 0.0015588151291012764, + 0.0025260294787585735, + 0.0026834821328520775, + 0.002308339811861515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -1.6193338632583618 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.006003205664455891 + ], + "opposite_logprob_tokens": [ + -5.131003379821777 + ], + "logprob_reward": -0.006003205664455891, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006003205664455891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04372011125087738, + 0.29597264528274536, + 0.021158114075660706, + 0.13883085548877716, + 0.12169864028692245, + 0.07528641074895859, + 0.07880205661058426, + 0.019056392833590508, + 0.17594189941883087, + 0.029532866552472115 + ], + "credit_map": [ + 0.006894199177622795, + 0.006414945237338543, + 0.004888829775154591, + 0.004784666933119297, + 0.004653637297451496, + 0.005213955417275429, + 0.005143843125551939, + 0.005726035684347153, + 0.028747854754328728, + 0.031376857310533524, + 0.11130424588918686, + 0.0155910924077034, + 0.0177862998098135, + 0.022260820493102074, + 0.010303749702870846, + 0.05860171839594841, + 0.0017983377911150455, + 0.002267932053655386, + 0.0018834418151527643, + 0.002765452489256859, + 0.0030056769028306007, + 0.00345323933288455, + 0.003315928392112255, + 0.002668104600161314, + 0.017218392342329025, + 0.020999295637011528, + 0.013647234998643398, + 0.015220481902360916, + 0.01698356308043003, + 0.022841615602374077, + 0.013007529079914093, + 0.01891273818910122, + 0.016708675771951675, + 0.01638825610280037, + 0.013935797847807407, + 0.01372024230659008, + 0.011605489067733288, + 0.02053876593708992, + 0.0061524296179413795, + 0.02264898270368576, + 0.008231810294091702, + 0.007250953931361437, + 0.006828352343291044, + 0.008094646967947483, + 0.009448250755667686, + 0.01282215490937233, + 0.009646912105381489, + 0.012963321059942245, + 0.014277005568146706, + 0.005303983110934496, + 0.011570165865123272, + 0.015757907181978226, + 0.004182963166385889, + 0.006224106531590223, + 0.011272544972598553, + 0.010213378816843033, + 0.0016129391733556986, + 0.0019528089324012399, + 0.0031488779932260513, + 0.0033698761835694313, + 0.0018047092016786337, + 0.0015268998686224222, + 0.0025306097231805325, + 0.003109671175479889, + 0.014614627696573734, + 0.014637596905231476, + 0.0378597192466259, + 0.018270323053002357, + 0.00916243251413107, + 0.01685764640569687, + 0.024329565465450287, + 0.04020998626947403, + 0.004915208555757999, + 0.0020504153799265623, + 0.005237109959125519, + 0.00290776789188385, + 0.001591252745129168, + 0.003933955915272236, + 0.004152196459472179, + 0.004744959529489279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.006003205664455891 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.007636165712028742 + ], + "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": [ + 0.06589696556329727, + 0.31623363494873047, + 0.023049911484122276, + 0.1721220165491104, + 0.09374960511922836, + 0.0757160410284996, + 0.047236211597919464, + 0.023577943444252014, + 0.15676744282245636, + 0.025650186464190483 + ], + "credit_map": [ + 0.009171398356556892, + 0.010146832093596458, + 0.006069398485124111, + 0.0076705170795321465, + 0.011254904791712761, + 0.007135116029530764, + 0.005062325857579708, + 0.009386471472680569, + 0.031151974573731422, + 0.03411359712481499, + 0.11925734579563141, + 0.016561370342969894, + 0.01825106330215931, + 0.023640815168619156, + 0.010792065411806107, + 0.06246541813015938, + 0.0022801614832133055, + 0.002598434453830123, + 0.0019343396415933967, + 0.002825109986588359, + 0.0027838884852826595, + 0.0037172355223447084, + 0.003285844111815095, + 0.0036248962860554457, + 0.02210542932152748, + 0.03037390485405922, + 0.013711757026612759, + 0.020519109442830086, + 0.019208485260605812, + 0.027258964255452156, + 0.013557073660194874, + 0.02538730762898922, + 0.012335396371781826, + 0.012314784340560436, + 0.010259519331157207, + 0.009852353483438492, + 0.009546511806547642, + 0.01618652604520321, + 0.005017413292080164, + 0.018237099051475525, + 0.008693990297615528, + 0.00750014279037714, + 0.006937643047422171, + 0.008436287753283978, + 0.008214809000492096, + 0.013371884822845459, + 0.009001619182527065, + 0.013559667393565178, + 0.006229662336409092, + 0.0034946571104228497, + 0.006817733868956566, + 0.009695492684841156, + 0.0031969882547855377, + 0.0024487576447427273, + 0.009094907902181149, + 0.006258011795580387, + 0.0020917693618685007, + 0.0021254466846585274, + 0.0038846477400511503, + 0.004752239212393761, + 0.0018770035821944475, + 0.0015833735233172774, + 0.0034459796734154224, + 0.0038174851797521114, + 0.013339394703507423, + 0.012902004644274712, + 0.03345711901783943, + 0.01615726947784424, + 0.008367015048861504, + 0.015322861261665821, + 0.02192535065114498, + 0.03529644012451172, + 0.004814603365957737, + 0.0015937736025080085, + 0.004740084521472454, + 0.0023593599908053875, + 0.001262766309082508, + 0.003487658454105258, + 0.003556018229573965, + 0.003835920477285981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "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 video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.007627056445926428 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627056445926428, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627056445926428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07990560680627823, + 0.3255268335342407, + 0.031989797949790955, + 0.15602822601795197, + 0.06492272764444351, + 0.0689845159649849, + 0.07350695878267288, + 0.0910576581954956, + 0.06185613200068474, + 0.04622150957584381 + ], + "credit_map": [ + 0.008728096261620522, + 0.011718750931322575, + 0.00629142951220274, + 0.007912511005997658, + 0.01753544621169567, + 0.007515811361372471, + 0.005311562214046717, + 0.01489200722426176, + 0.0321088470518589, + 0.035160984843969345, + 0.12346744537353516, + 0.016025329008698463, + 0.02006632648408413, + 0.022970449179410934, + 0.011934812180697918, + 0.06379267573356628, + 0.0030674515292048454, + 0.0029811454005539417, + 0.0026050684973597527, + 0.0036519644781947136, + 0.00401087524369359, + 0.00646031042560935, + 0.004403376951813698, + 0.004809611476957798, + 0.019068997353315353, + 0.028498444706201553, + 0.013711463660001755, + 0.019168294966220856, + 0.018157167360186577, + 0.025723276659846306, + 0.01271274033933878, + 0.018987862393260002, + 0.007642826531082392, + 0.009099212475121021, + 0.007298919837921858, + 0.005885601043701172, + 0.007282106671482325, + 0.011169178411364555, + 0.004505124408751726, + 0.012039760127663612, + 0.007801819127053022, + 0.007857079617679119, + 0.007011753972619772, + 0.0079565504565835, + 0.00861266441643238, + 0.010079007595777512, + 0.008579094894230366, + 0.01108655333518982, + 0.00972210243344307, + 0.00646755937486887, + 0.009806614369153976, + 0.014229711145162582, + 0.005114613100886345, + 0.005119378212839365, + 0.013625334948301315, + 0.009421654045581818, + 0.008759650401771069, + 0.007709719240665436, + 0.01354976836591959, + 0.017866963520646095, + 0.006192958448082209, + 0.005361347924917936, + 0.014395560137927532, + 0.01722169667482376, + 0.00539977615699172, + 0.00401116069406271, + 0.009974133223295212, + 0.008458904922008514, + 0.003975603263825178, + 0.006134210620075464, + 0.012532232329249382, + 0.011370114982128143, + 0.005825893022119999, + 0.004836469888687134, + 0.0076143997721374035, + 0.005545602645725012, + 0.0033110948279500008, + 0.005104654002934694, + 0.0069870492443442345, + 0.006996351294219494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.007627056445926428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.003616461530327797 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.003616461530327797, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616461530327797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08376503735780716, + 0.1405758261680603, + 0.03278415650129318, + 0.21661795675754547, + 0.12774555385112762, + 0.06769220530986786, + 0.05433354154229164, + 0.05937028303742409, + 0.14065229892730713, + 0.07646311819553375 + ], + "credit_map": [ + 0.008277962915599346, + 0.011585481464862823, + 0.00567044410854578, + 0.007631318643689156, + 0.019268296658992767, + 0.006728434935212135, + 0.008016576990485191, + 0.016586516052484512, + 0.013992845080792904, + 0.01562163420021534, + 0.05226480960845947, + 0.007276618853211403, + 0.008645463734865189, + 0.01002716924995184, + 0.004989531822502613, + 0.02775774896144867, + 0.003009669715538621, + 0.00393716199323535, + 0.002668769331648946, + 0.0036669964902102947, + 0.004240207374095917, + 0.006921498104929924, + 0.0036155430134385824, + 0.004724309779703617, + 0.027146881446242332, + 0.038085244596004486, + 0.017755631357431412, + 0.024953311309218407, + 0.024855170398950577, + 0.035665929317474365, + 0.018201671540737152, + 0.029954105615615845, + 0.01715647056698799, + 0.016719719395041466, + 0.015723861753940582, + 0.015284884721040726, + 0.01172195840626955, + 0.021034354344010353, + 0.0064681824296712875, + 0.02363610826432705, + 0.007046558428555727, + 0.007150961086153984, + 0.007839009165763855, + 0.007966868579387665, + 0.007593742571771145, + 0.01110814604908228, + 0.008932108990848064, + 0.010054806247353554, + 0.00576474703848362, + 0.00478817243129015, + 0.005959228612482548, + 0.007913562469184399, + 0.005245102569460869, + 0.004692587535828352, + 0.011922886595129967, + 0.008047252893447876, + 0.007095839362591505, + 0.0053360010497272015, + 0.008131331764161587, + 0.009619842283427715, + 0.006366802379488945, + 0.006142554804682732, + 0.0074072955176234245, + 0.009270616807043552, + 0.010587343946099281, + 0.010387425310909748, + 0.03317806497216225, + 0.014233039692044258, + 0.006705374922603369, + 0.01219032984226942, + 0.018712062388658524, + 0.034658659249544144, + 0.010995655320584774, + 0.006663901265710592, + 0.01332426443696022, + 0.009276477620005608, + 0.004971674643456936, + 0.007734654471278191, + 0.010941319167613983, + 0.012555171735584736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.003616461530327797 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot captures a scene rendered in the stark, high-contrast style of black-and-white ukiyo-e animation. From an unseen source below the frame, countless thick, fibrous sheets of wakashi-style paper rise in a powerful, swirling vortex. Each strip is meticulously crafted, featuring a rough, tactile texture with visible fibers and clean, perforated edges. They are embellished with intricate traditional motifs—some depicting Hokusai's waving waves, others the delicate lines of swirling clouds—all executed with the bold, calligraphic outlines characteristic of woodblock prints. As the papers tumble and flutter, they spiral upward, forming a mesmerizing, chaotic yet elegant vortex that dominates the center of the composition. Sharp, directed light pierces the darkness, catching the edges and flat surfaces of the paper, creating bright white highlights that contrast strikingly with the empty depths, inky darkness, and absolute blackness of the background, creating a hypnotic and powerful visual spectacle brimming with energy.", + "seed": 25142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_15/index_4898_seed_25142_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": -0.1137091416399926, + "vlm_logprob": -0.1137091416399926, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.1137091416399926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.5483261942863464 + ], + "opposite_logprob_tokens": [ + -0.9233261942863464 + ], + "logprob_reward": -0.5483261942863464, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5483261942863464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06952842324972153, + 0.48520126938819885, + 0.05696258693933487, + 0.048748522996902466, + 0.03693370893597603, + 0.08496545255184174, + 0.03031059354543686, + 0.06921618431806564, + 0.08328913897275925, + 0.03484411537647247 + ], + "credit_map": [ + 0.008109906688332558, + 0.007708974182605743, + 0.011850481852889061, + 0.008661091327667236, + 0.007562906015664339, + 0.008593279868364334, + 0.008064975030720234, + 0.00897680688649416, + 0.022958379238843918, + 0.009472828358411789, + 0.21036337316036224, + 0.03885764256119728, + 0.012886757031083107, + 0.01836010068655014, + 0.06156449392437935, + 0.11073771119117737, + 0.008883921429514885, + 0.007762886583805084, + 0.005003588739782572, + 0.007182848174124956, + 0.008346937596797943, + 0.005429270677268505, + 0.005673081614077091, + 0.008680053986608982, + 0.004221626091748476, + 0.00395900197327137, + 0.006478422321379185, + 0.0045278132893145084, + 0.007197702303528786, + 0.01026216521859169, + 0.005288031417876482, + 0.006813759915530682, + 0.009558011777698994, + 0.0028479101601988077, + 0.005526505410671234, + 0.002713986672461033, + 0.003959743306040764, + 0.004402015823870897, + 0.0025353282690048218, + 0.005390209145843983, + 0.009935034438967705, + 0.009354385547339916, + 0.007444129791110754, + 0.0062603880651295185, + 0.007290740963071585, + 0.01949901320040226, + 0.011996399611234665, + 0.013185359537601471, + 0.0034369630739092827, + 0.004571783822029829, + 0.0029990163166075945, + 0.003214299911633134, + 0.003055384149774909, + 0.0049370345659554005, + 0.004583404399454594, + 0.003512707306072116, + 0.006581388879567385, + 0.007557913661003113, + 0.007123746443539858, + 0.008422981947660446, + 0.009874296374619007, + 0.00990153569728136, + 0.010116137564182281, + 0.009638184681534767, + 0.013264203444123268, + 0.008012835867702961, + 0.012078053317964077, + 0.013430098071694374, + 0.004684646148234606, + 0.014291156083345413, + 0.006949242670089006, + 0.01057890523225069, + 0.004881423432379961, + 0.004224769305437803, + 0.005064483266323805, + 0.004301021341234446, + 0.00309598445892334, + 0.0048864190466701984, + 0.004088773857802153, + 0.0043012406677007675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.5483261942863464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.0032406931277364492 + ], + "opposite_logprob_tokens": [ + -5.753240585327148 + ], + "logprob_reward": -0.0032406931277364492, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032406931277364492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05941033735871315, + 0.6053712964057922, + 0.027757678180933, + 0.03094319812953472, + 0.02867533080279827, + 0.06192091107368469, + 0.062175579369068146, + 0.026478193700313568, + 0.024794653058052063, + 0.07247284054756165 + ], + "credit_map": [ + 0.005457120016217232, + 0.003942642826586962, + 0.010134752839803696, + 0.008777838200330734, + 0.004853608086705208, + 0.010026087984442711, + 0.007470596581697464, + 0.008747684769332409, + 0.028562646359205246, + 0.011230502277612686, + 0.26438459753990173, + 0.04761840030550957, + 0.015926459804177284, + 0.022196635603904724, + 0.07873167097568512, + 0.13672038912773132, + 0.002890476258471608, + 0.0036171802785247564, + 0.0023048894945532084, + 0.002793773775920272, + 0.004705897532403469, + 0.0057836310006678104, + 0.0022010253742337227, + 0.0034608046989887953, + 0.002321117790415883, + 0.00180942309089005, + 0.004555330611765385, + 0.0025224171113222837, + 0.003411786165088415, + 0.00823457632213831, + 0.0033450222108513117, + 0.004743524827063084, + 0.005124666728079319, + 0.001837837160564959, + 0.0030230919364839792, + 0.0027239848859608173, + 0.002845972077921033, + 0.00710239028558135, + 0.002420533914119005, + 0.003596854628995061, + 0.006304668262600899, + 0.008076374419033527, + 0.006031375378370285, + 0.004615188576281071, + 0.005944062024354935, + 0.013792671263217926, + 0.007918925024569035, + 0.009237646125257015, + 0.01132154744118452, + 0.007999351248145103, + 0.008348241448402405, + 0.006358705926686525, + 0.005689552519470453, + 0.009229307994246483, + 0.006709932815283537, + 0.006518939044326544, + 0.002373696304857731, + 0.0038835369050502777, + 0.0017390005523338914, + 0.002660028636455536, + 0.0037842029705643654, + 0.0035615793894976377, + 0.004846008028835058, + 0.003630139399319887, + 0.0024370630271732807, + 0.0019988329149782658, + 0.0025547444820404053, + 0.0028365550097078085, + 0.001792949391528964, + 0.007804578635841608, + 0.0018967738142237067, + 0.003473154501989484, + 0.011563066393136978, + 0.007729257456958294, + 0.011420096270740032, + 0.008825578726828098, + 0.0054211788810789585, + 0.009688789024949074, + 0.008569472469389439, + 0.009255398996174335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.0032406931277364492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.007635101210325956 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.007635101210325956, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635101210325956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037493135780096054, + 0.7967606782913208, + 0.01819179765880108, + 0.024540701881051064, + 0.027416681870818138, + 0.026473956182599068, + 0.012886060401797295, + 0.024630559608340263, + 0.02235337160527706, + 0.00925308745354414 + ], + "credit_map": [ + 0.0034037125296890736, + 0.0034265201538801193, + 0.005421943496912718, + 0.004811548627912998, + 0.0037705202121287584, + 0.006578768137842417, + 0.004759245552122593, + 0.005320876836776733, + 0.03742039576172829, + 0.014735789969563484, + 0.3474600911140442, + 0.06249270588159561, + 0.02063254825770855, + 0.02938833273947239, + 0.10343392938375473, + 0.18119683861732483, + 0.0032735085114836693, + 0.0024536976125091314, + 0.001749801798723638, + 0.0017218725988641381, + 0.0029394798912107944, + 0.0027459205593913794, + 0.001097226981073618, + 0.002210289938375354, + 0.0018076233100146055, + 0.0020177452825009823, + 0.004145325161516666, + 0.0026610244531184435, + 0.0030925204046070576, + 0.0033714526798576117, + 0.0036156962160021067, + 0.00382931437343359, + 0.006595642305910587, + 0.001900693867355585, + 0.0038849622942507267, + 0.002469286322593689, + 0.00235931808128953, + 0.0042763929814100266, + 0.002354783471673727, + 0.0035756020806729794, + 0.003078646259382367, + 0.0027994683478027582, + 0.002130829729139805, + 0.0014675036072731018, + 0.002256442094221711, + 0.006996918004006147, + 0.0037564581725746393, + 0.0039876908995211124, + 0.001746942987665534, + 0.0014287478988990188, + 0.0012641354696825147, + 0.0012258072383701801, + 0.0011284664506092668, + 0.0032364479266107082, + 0.0012582074850797653, + 0.001597305410541594, + 0.0022164771798998117, + 0.004662718158215284, + 0.0015400736592710018, + 0.0019754301756620407, + 0.004857028368860483, + 0.003540136618539691, + 0.002592594362795353, + 0.0032461010850965977, + 0.0024954532273113728, + 0.0016132215969264507, + 0.002230813028290868, + 0.0022984568495303392, + 0.0015345800202339888, + 0.00701585179194808, + 0.0018972534453496337, + 0.0032677422277629375, + 0.002223582472652197, + 0.0007467115065082908, + 0.0014909510500729084, + 0.0008989766938611865, + 0.0005603444296866655, + 0.0013773804530501366, + 0.0007651688647456467, + 0.0011899712262675166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.007635101210325956 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.005250236950814724 + ], + "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": [ + 0.025106776505708694, + 0.7996075749397278, + 0.03947950899600983, + 0.03375408053398132, + 0.018463896587491035, + 0.023282138630747795, + 0.016177689656615257, + 0.015121047385036945, + 0.02056613564491272, + 0.008441178128123283 + ], + "credit_map": [ + 0.002299429150298238, + 0.002388174179941416, + 0.004068539012223482, + 0.003319038078188896, + 0.0023066543508321047, + 0.003489000489935279, + 0.0037923739291727543, + 0.0034435673151165247, + 0.03761224076151848, + 0.014717667363584042, + 0.3484039604663849, + 0.06286680698394775, + 0.020929601043462753, + 0.030297936871647835, + 0.1043233796954155, + 0.18045593798160553, + 0.004982035141438246, + 0.005892840214073658, + 0.003118830965831876, + 0.0037584970705211163, + 0.0065558897331357, + 0.0076703643426299095, + 0.0024488605558872223, + 0.005052190274000168, + 0.002512498991563916, + 0.002498254645615816, + 0.004458686336874962, + 0.0030872870702296495, + 0.0037867496721446514, + 0.00773881608620286, + 0.004281098488718271, + 0.005390692502260208, + 0.004242324270308018, + 0.001396492007188499, + 0.0028933468274772167, + 0.0016596439527347684, + 0.0014014827320352197, + 0.0030286230612546206, + 0.0012648631818592548, + 0.002577119506895542, + 0.0023196637630462646, + 0.0024154616985470057, + 0.001979622757062316, + 0.0014882044633850455, + 0.00199456000700593, + 0.00544031010940671, + 0.0037849810905754566, + 0.0038593353237956762, + 0.0029758564196527004, + 0.0021413571666926146, + 0.0021183271892368793, + 0.001782740349881351, + 0.0013767416821792722, + 0.002644486492499709, + 0.0016673057107254863, + 0.0014708754606544971, + 0.0013795157428830862, + 0.002756245667114854, + 0.0010913574369624257, + 0.0013424389762803912, + 0.002844384405761957, + 0.0019880610052496195, + 0.0017713910201564431, + 0.0019476525485515594, + 0.002282977569848299, + 0.0016016366425901651, + 0.0019426140934228897, + 0.0017965490696951747, + 0.0014873790787532926, + 0.006957957986742258, + 0.0015336415963247418, + 0.0029633785597980022, + 0.0013529672287404537, + 0.00098734674975276, + 0.0013446124503389, + 0.0009738763910718262, + 0.0006411916692741215, + 0.0010820388561114669, + 0.0009849111083894968, + 0.0010742334416136146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "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 video shot as a steady medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a stark, high-contrast black-and-white ukiyo-e animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, fibrous sheets of paper rising in a swirling vortex from below the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper strips feature rough textures, perforated edges, and traditional motifs like waves or clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there sharp, directed light creating bright white highlights against an inky black 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": [ + -0.0040934826247394085 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.0040934826247394085, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040934826247394085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05250287428498268, + 0.6296276450157166, + 0.04158281534910202, + 0.0616963766515255, + 0.05365241691470146, + 0.022784249857068062, + 0.04464719071984291, + 0.04231076315045357, + 0.03387969732284546, + 0.01731591671705246 + ], + "credit_map": [ + 0.00530967628583312, + 0.004300582222640514, + 0.007560423109680414, + 0.007238395046442747, + 0.005103750620037317, + 0.00940881296992302, + 0.006251182407140732, + 0.007330052554607391, + 0.0291785579174757, + 0.012194340117275715, + 0.27437835931777954, + 0.05012710392475128, + 0.015162764117121696, + 0.023434394970536232, + 0.07983332872390747, + 0.145318865776062, + 0.005214199423789978, + 0.004549788776785135, + 0.0032791525591164827, + 0.003084172960370779, + 0.005033715628087521, + 0.011517428793013096, + 0.0035291046369820833, + 0.00537525350227952, + 0.005949230864644051, + 0.004253329709172249, + 0.0074373153038322926, + 0.007677625399082899, + 0.006720412522554398, + 0.0119917718693614, + 0.007964554242789745, + 0.009702140465378761, + 0.009418383240699768, + 0.0041645183227956295, + 0.006789794657379389, + 0.004298267886042595, + 0.005426935385912657, + 0.011803199537098408, + 0.00441849697381258, + 0.007332819979637861, + 0.0023371584247797728, + 0.0020125398878008127, + 0.002046918496489525, + 0.00161497772205621, + 0.0018375621875748038, + 0.005687115713953972, + 0.003436926519498229, + 0.003811052767559886, + 0.005372183863073587, + 0.005537489429116249, + 0.003479979932308197, + 0.003808800596743822, + 0.003916488029062748, + 0.010680112987756729, + 0.00545841921120882, + 0.00639371806755662, + 0.0038918128702789545, + 0.006463538855314255, + 0.0028840864542871714, + 0.004498359747231007, + 0.005350245628505945, + 0.0057912953197956085, + 0.007280478719621897, + 0.006150944624096155, + 0.0032386397942900658, + 0.0025049736723303795, + 0.003455663798376918, + 0.0032963156700134277, + 0.002148631028831005, + 0.011447661556303501, + 0.002731883665546775, + 0.005055926740169525, + 0.002766517922282219, + 0.0015723617980256677, + 0.0029962388798594475, + 0.0018978510051965714, + 0.0012468366185203195, + 0.001803161925636232, + 0.002587683033198118, + 0.002445265417918563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.0040934826247394085 + } + ], + "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/0000122/group_2/index_tensor([3378, 3378])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_tensor([3378, 3378])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3719c71d394982ab23cb6727f1c1bf98148f9c04 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_tensor([3378, 3378])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward-and-slightly-closed shot takes us deep into a pristine jungle, where a lone explorer stands before a majestic ancient stone gate. He wears a weathered, multi-pocket khaki jacket, its cuffs and hem stained with dirt, a dark gray T-shirt, durable dark overalls, and muddy hiking boots. He appears determined yet slightly exhausted. The gate before him is massive and majestic, carved from a single block of dark rock. Its surface is covered with intricate relief carvings eroded by time. Thick, dark green vines coil around it like giant pythons, almost blending it into the surrounding forest. In the damp, misty air, the explorer, hesitant yet resolute, slowly raises his right hand, clenching it into a fist. Before striking the cold, mossy stone, his fist pauses in mid-air. Dappled sunlight penetrates the dense treetops overhead, casting beams of light and illuminating the dust particles floating in the air. The entire picture is dominated by dark green and earth tones, creating a mysterious, solemn and suspenseful atmosphere, as if a long-buried secret is about to be revealed.", + "seed": 3212, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_3378_seed_3212_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": -0.07162754335440695, + "vlm_logprob": -0.07162754335440695, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.07162754335440695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23816873133182526 + ], + "opposite_logprob_tokens": [ + -1.738168716430664 + ], + "logprob_reward": -0.23816873133182526, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.23816873133182526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02046222798526287, + 0.009465438313782215, + 0.009858811274170876, + 0.014057443477213383, + 0.05061478912830353, + 0.00973710510879755, + 0.01832129806280136, + 0.012740286067128181, + 0.006936823949217796, + 0.8478057384490967 + ], + "credit_map": [ + 0.0023654643446207047, + 0.003827400505542755, + 0.003181898733600974, + 0.0026629185304045677, + 0.001785026746802032, + 0.002394386101514101, + 0.0019473378779366612, + 0.0022977960761636496, + 0.0009021254372783005, + 0.001273808185942471, + 0.0015148189850151539, + 0.0011192434467375278, + 0.0012554398272186518, + 0.0011334243463352323, + 0.0009465999901294708, + 0.0013199785025790334, + 0.0010120873339474201, + 0.001227125059813261, + 0.0014637998538091779, + 0.0009978924645110965, + 0.0010587377473711967, + 0.0016218641540035605, + 0.0008898890228010714, + 0.0015874163946136832, + 0.001368817174807191, + 0.0019862339831888676, + 0.0015865284949541092, + 0.0015095333801582456, + 0.0020196023397147655, + 0.001980625791475177, + 0.0014613153180107474, + 0.0021447869949042797, + 0.005279132165014744, + 0.005819248501211405, + 0.0035773420240730047, + 0.0038563874550163746, + 0.004493009299039841, + 0.012543512508273125, + 0.0036865361034870148, + 0.01135962177067995, + 0.0013100843643769622, + 0.0014755106531083584, + 0.0012306718854233623, + 0.001470153802074492, + 0.0008188174106180668, + 0.0014434009790420532, + 0.000619928992819041, + 0.0013685368467122316, + 0.0017776216845959425, + 0.003021531505510211, + 0.002050406765192747, + 0.00204725144430995, + 0.003299988806247711, + 0.002679686527699232, + 0.0013421188341453671, + 0.0021026916801929474, + 0.0007585707935504615, + 0.0015255511971190572, + 0.002013435121625662, + 0.0017918568337336183, + 0.0013364861952140927, + 0.0015934566035866737, + 0.0019811801612377167, + 0.0017397502670064569, + 0.0004535836342256516, + 0.0007010961417108774, + 0.000787726603448391, + 0.0007406209479086101, + 0.0004728613421320915, + 0.0017101714620366693, + 0.0006296432111412287, + 0.0014411205193027854, + 0.021827982738614082, + 0.01244139950722456, + 0.012775934301316738, + 0.019460024312138557, + 0.02707245759665966, + 0.37040337920188904, + 0.035717323422431946, + 0.3481072783470154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.23816873133182526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812025114893913 + ], + "opposite_logprob_tokens": [ + -3.2881202697753906 + ], + "logprob_reward": -0.03812025114893913, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812025114893913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011632951907813549, + 0.007135552819818258, + 0.006503487471491098, + 0.01234771590679884, + 0.09351170063018799, + 0.0038940240629017353, + 0.0054204463958740234, + 0.006162847857922316, + 0.020178303122520447, + 0.8332129716873169 + ], + "credit_map": [ + 0.001545794541016221, + 0.0014674809062853456, + 0.001885272329673171, + 0.0014564415905624628, + 0.001396140898577869, + 0.0014155013486742973, + 0.0010459567420184612, + 0.0014203640166670084, + 0.0008468936430290341, + 0.0008122528088279068, + 0.0012595630250871181, + 0.0009954910492524505, + 0.0007352124666795135, + 0.000938866869546473, + 0.0006443450693041086, + 0.0009029279462993145, + 0.0007083471864461899, + 0.0008145786705426872, + 0.001587758189998567, + 0.0006440689321607351, + 0.0005175094702281058, + 0.0007421207264997065, + 0.0005916377413086593, + 0.0008974668453447521, + 0.0010562732350081205, + 0.0023723726626485586, + 0.0009570901747792959, + 0.001858582254499197, + 0.0012140534818172455, + 0.002061291364952922, + 0.0009341593249700963, + 0.0018938943976536393, + 0.008465574122965336, + 0.010047179646790028, + 0.00570204621180892, + 0.007368830498307943, + 0.007748074363917112, + 0.025249058380723, + 0.006517975125461817, + 0.022412968799471855, + 0.0005574130918830633, + 0.0005335776950232685, + 0.00046441954327747226, + 0.00048300749040208757, + 0.0003264706174377352, + 0.0007230208138935268, + 0.00020591887005139142, + 0.000600195664446801, + 0.0005717151216231287, + 0.000844508467707783, + 0.0004963368992321193, + 0.0007267557666637003, + 0.0007517196354456246, + 0.0007867988897487521, + 0.0005585091421380639, + 0.0006841023568995297, + 0.00037070270627737045, + 0.0008797328337095678, + 0.0007475042948499322, + 0.0007488270057365298, + 0.0007227631867863238, + 0.001126862596720457, + 0.0005597890703938901, + 0.001006665755994618, + 0.0013198622036725283, + 0.002550950041040778, + 0.0015521638561040163, + 0.002174069406464696, + 0.0014151677023619413, + 0.003923570271581411, + 0.0030678187031298876, + 0.004174701403826475, + 0.02127653732895851, + 0.01204301230609417, + 0.012451366521418095, + 0.019393213093280792, + 0.026222601532936096, + 0.3639759421348572, + 0.03505781665444374, + 0.34279245138168335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.03812025114893913 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495903640985489 + ], + "opposite_logprob_tokens": [ + -2.9299590587615967 + ], + "logprob_reward": -0.05495903640985489, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495903640985489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03140535578131676, + 0.029876552522182465, + 0.016184180974960327, + 0.013707838952541351, + 0.03149384632706642, + 0.014195761643350124, + 0.014276152476668358, + 0.0033792953472584486, + 0.00897976290434599, + 0.8365011811256409 + ], + "credit_map": [ + 0.0038650985807180405, + 0.002881887136027217, + 0.006142735946923494, + 0.004996840842068195, + 0.003093909239396453, + 0.003574708476662636, + 0.0029057702049613, + 0.0039444067515432835, + 0.0029598381370306015, + 0.00461675226688385, + 0.0072332206182181835, + 0.002951430855318904, + 0.002615127479657531, + 0.0031619591172784567, + 0.0027288910932838917, + 0.0036093336530029774, + 0.0015766301658004522, + 0.0019342204323038459, + 0.003763853572309017, + 0.0018850858323276043, + 0.0013315172400325537, + 0.002356355544179678, + 0.0013509620912373066, + 0.001985556911677122, + 0.0011903790291398764, + 0.002095871837809682, + 0.0016509788110852242, + 0.0020070108585059643, + 0.001523011364042759, + 0.0022023709025233984, + 0.0011189107317477465, + 0.00191930599976331, + 0.0027763275429606438, + 0.0052413628436625, + 0.0032096151262521744, + 0.0035296163987368345, + 0.0024911288637667894, + 0.005911209620535374, + 0.0021510529331862926, + 0.0061835371889173985, + 0.0013168565928936005, + 0.002228938741609454, + 0.0016996451886370778, + 0.0022250255569815636, + 0.001157483202405274, + 0.0022281226702034473, + 0.0010290842037647963, + 0.0023106071166694164, + 0.0011798443738371134, + 0.00248821172863245, + 0.0015345995780080557, + 0.0019416685681790113, + 0.002509722951799631, + 0.0019277794053778052, + 0.0009601900237612426, + 0.0017341359052807093, + 0.00024408416356891394, + 0.00026799182523973286, + 0.00032852721051312983, + 0.00030473238439299166, + 0.00040512013947591186, + 0.0005194925470277667, + 0.0006530653918161988, + 0.0006562818307429552, + 0.0005533883231692016, + 0.0005775425815954804, + 0.000653041759505868, + 0.000698730000294745, + 0.0006044395850040019, + 0.0024815641809254885, + 0.0008637403952889144, + 0.002547317184507847, + 0.0215167086571455, + 0.012456313706934452, + 0.012593641877174377, + 0.01918984204530716, + 0.026343313977122307, + 0.36519160866737366, + 0.0351746529340744, + 0.34403514862060547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.05495903640985489 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08738071471452713, + 0.05858174338936806, + 0.020849164575338364, + 0.055267926305532455, + 0.08065967261791229, + 0.018096525222063065, + 0.01154446043074131, + 0.01542381290346384, + 0.01674622856080532, + 0.6354497075080872 + ], + "credit_map": [ + 0.010351353324949741, + 0.007138610351830721, + 0.015191051177680492, + 0.010975685901939869, + 0.010683592408895493, + 0.011251499876379967, + 0.009743379428982735, + 0.012045544572174549, + 0.005742356181144714, + 0.008415907621383667, + 0.010922768153250217, + 0.006711938418447971, + 0.007779194973409176, + 0.006912035867571831, + 0.005449701100587845, + 0.006647838279604912, + 0.0018265631515532732, + 0.0034997984766960144, + 0.0037435332778841257, + 0.002139496151357889, + 0.0023799072951078415, + 0.0023441878147423267, + 0.0020775138400495052, + 0.002838162938132882, + 0.006348912604153156, + 0.01007638219743967, + 0.0056742047891020775, + 0.0071627129800617695, + 0.0068961395882070065, + 0.007496244739741087, + 0.00450867647305131, + 0.00710465619340539, + 0.007378190290182829, + 0.011460776440799236, + 0.006758939940482378, + 0.009211204946041107, + 0.008368590846657753, + 0.017742305994033813, + 0.006152534391731024, + 0.013587134890258312, + 0.0020449198782444, + 0.002465151948854327, + 0.002241085283458233, + 0.002479751594364643, + 0.0014195690164342523, + 0.002977181226015091, + 0.0014934136997908354, + 0.0029754533898085356, + 0.001144647365435958, + 0.0014934379141777754, + 0.0017090047476813197, + 0.001398048480041325, + 0.0016263903817161918, + 0.0016080668428912759, + 0.001136349281296134, + 0.0014285154175013304, + 0.0007649203180335462, + 0.000947915599681437, + 0.001245957682840526, + 0.0010598944500088692, + 0.0019866982474923134, + 0.002540681743994355, + 0.0041619837284088135, + 0.0027157615404576063, + 0.00141051784157753, + 0.0013247053138911724, + 0.0017945148283615708, + 0.0017017330974340439, + 0.001144076231867075, + 0.00329032214358449, + 0.002806758973747492, + 0.0032736009452492, + 0.01644093357026577, + 0.00953875295817852, + 0.009877758100628853, + 0.01444970816373825, + 0.020075758919119835, + 0.27718937397003174, + 0.026850303635001183, + 0.26102712750434875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0036140859592705965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327561192214489 + ], + "opposite_logprob_tokens": [ + -3.77327561378479 + ], + "logprob_reward": -0.02327561192214489, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327561192214489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04801033064723015, + 0.024005679413676262, + 0.03874654322862625, + 0.05002700537443161, + 0.08630840480327606, + 0.024929869920015335, + 0.0345769077539444, + 0.09122990816831589, + 0.08089838922023773, + 0.5212669968605042 + ], + "credit_map": [ + 0.005148620344698429, + 0.00717118987813592, + 0.007972417399287224, + 0.006699478719383478, + 0.004846906289458275, + 0.005456874147057533, + 0.004786095581948757, + 0.005928754806518555, + 0.003147430019453168, + 0.0028965824749320745, + 0.00513041578233242, + 0.002942936262115836, + 0.002268698997795582, + 0.002489306963980198, + 0.002461713505908847, + 0.002668597036972642, + 0.0036880415864288807, + 0.004948650021106005, + 0.007694515865296125, + 0.00432157376781106, + 0.0035192095674574375, + 0.006276490166783333, + 0.0029473910108208656, + 0.0053506772965192795, + 0.004549968522042036, + 0.008208987303078175, + 0.005091909784823656, + 0.007801393512636423, + 0.006263750605285168, + 0.007949716411530972, + 0.003564954036846757, + 0.006596332881599665, + 0.008392340503633022, + 0.008472763933241367, + 0.005532245151698589, + 0.00503377802670002, + 0.0064640226773917675, + 0.02425144612789154, + 0.005779408849775791, + 0.022382404655218124, + 0.002969514112919569, + 0.003852652385830879, + 0.00325917755253613, + 0.004485868848860264, + 0.002160061150789261, + 0.003541788784787059, + 0.0015874849632382393, + 0.0030733251478523016, + 0.0026711924001574516, + 0.006545279640704393, + 0.003878504503518343, + 0.004220390226691961, + 0.005008995067328215, + 0.005278236232697964, + 0.002240918343886733, + 0.00473339669406414, + 0.003759497543796897, + 0.00828532688319683, + 0.008853836916387081, + 0.008448408916592598, + 0.012582282535731792, + 0.012929917313158512, + 0.025413215160369873, + 0.010957431979477406, + 0.005910153966397047, + 0.005711971782147884, + 0.009024222381412983, + 0.009263028390705585, + 0.0076467255130410194, + 0.010589119046926498, + 0.02197067067027092, + 0.010782505385577679, + 0.013442009687423706, + 0.009021429345011711, + 0.009440326131880283, + 0.012271535582840443, + 0.017346523702144623, + 0.22550351917743683, + 0.0223646592348814, + 0.2118769735097885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.02327561192214489 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-and-slightly-closed shot takes us deep into a pristine jungle, where a lone explorer stands before a majestic ancient stone gate. He wears a weathered, multi-pocket khaki jacket, its cuffs and hem stained with dirt, a dark gray T-shirt, durable dark overalls, and muddy hiking boots. He appears determined yet slightly exhausted. The gate before him is massive and majestic, carved from a single block of dark rock. Its surface is covered with intricate relief carvings eroded by time. Thick, dark green vines coil around it like giant pythons, almost blending it into the surrounding forest. In the damp, misty air, the explorer, hesitant yet resolute, slowly raises his right hand, clenching it into a fist. Before striking the cold, mossy stone, his fist pauses in mid-air. Dappled sunlight penetrates the dense treetops overhead, casting beams of light and illuminating the dust particles floating in the air. The entire picture is dominated by dark green and earth tones, creating a mysterious, solemn and suspenseful atmosphere, as if a long-buried secret is about to be revealed.", + "seed": 3213, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_3378_seed_3213_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": -0.04000826901756227, + "vlm_logprob": -0.04000826901756227, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.04000826901756227, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11797226965427399 + ], + "opposite_logprob_tokens": [ + -2.3679723739624023 + ], + "logprob_reward": -0.11797226965427399, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11797226965427399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21600258350372314, + 0.1472143530845642, + 0.07494811713695526, + 0.0487077459692955, + 0.06196208670735359, + 0.04283501207828522, + 0.07817839831113815, + 0.0979669988155365, + 0.1132134348154068, + 0.11897128075361252 + ], + "credit_map": [ + 0.04543684050440788, + 0.06262809783220291, + 0.014955013990402222, + 0.020826544612646103, + 0.012755345553159714, + 0.012275153771042824, + 0.023355580866336823, + 0.023769991472363472, + 0.017836859449744225, + 0.016388321295380592, + 0.028384443372488022, + 0.018213508650660515, + 0.013677638955414295, + 0.011185429990291595, + 0.025745240971446037, + 0.0157829150557518, + 0.009127345867455006, + 0.00727277435362339, + 0.008702343329787254, + 0.009771320037543774, + 0.00932732131332159, + 0.008886689320206642, + 0.008411934599280357, + 0.013448391109704971, + 0.007359321694821119, + 0.006990180350840092, + 0.005557730793952942, + 0.008564332500100136, + 0.005294718313962221, + 0.005786871071904898, + 0.003194817341864109, + 0.005959774367511272, + 0.005062640178948641, + 0.007373149506747723, + 0.00872899778187275, + 0.011624014936387539, + 0.006982250139117241, + 0.007888046093285084, + 0.0053727938793599606, + 0.008930196985602379, + 0.004164473153650761, + 0.005266773048788309, + 0.005248851142823696, + 0.006028539501130581, + 0.006209523882716894, + 0.005141630303114653, + 0.004949048161506653, + 0.005826178006827831, + 0.00794966146349907, + 0.008398487232625484, + 0.00894279032945633, + 0.007233736105263233, + 0.016462810337543488, + 0.006960731465369463, + 0.009921216405928135, + 0.012308970093727112, + 0.011992552317678928, + 0.010096858255565166, + 0.013135844841599464, + 0.012455562129616737, + 0.01621377468109131, + 0.010777417570352554, + 0.014253214001655579, + 0.009041773155331612, + 0.014172639697790146, + 0.010552974417805672, + 0.014377471059560776, + 0.01209022756665945, + 0.019099950790405273, + 0.015420672483742237, + 0.016250981017947197, + 0.011248511262238026, + 0.02059468813240528, + 0.0098585719242692, + 0.02139333076775074, + 0.01424581278115511, + 0.013975617475807667, + 0.012359559535980225, + 0.01776750385761261, + 0.008776185102760792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.11797226965427399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033707667142152786 + ], + "opposite_logprob_tokens": [ + -3.408707618713379 + ], + "logprob_reward": -0.033707667142152786, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033707667142152786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17478494346141815, + 0.23003678023815155, + 0.04623367264866829, + 0.08317374438047409, + 0.030058598145842552, + 0.01979796402156353, + 0.13213412463665009, + 0.1388544738292694, + 0.04885609447956085, + 0.09606961905956268 + ], + "credit_map": [ + 0.04334304854273796, + 0.05180603638291359, + 0.00947272777557373, + 0.015406833030283451, + 0.009303788654506207, + 0.014556089416146278, + 0.01826857030391693, + 0.012627855874598026, + 0.023660050705075264, + 0.026598133146762848, + 0.028562892228364944, + 0.03261849284172058, + 0.02170390635728836, + 0.016539720818400383, + 0.05224036052823067, + 0.02811320871114731, + 0.005950275342911482, + 0.0041913543827831745, + 0.008906693197786808, + 0.006976451724767685, + 0.003985704854130745, + 0.00490740267559886, + 0.004091695416718721, + 0.007224096450954676, + 0.009621751494705677, + 0.010484965518116951, + 0.008672215044498444, + 0.013380132615566254, + 0.012218656949698925, + 0.00907085556536913, + 0.01087285578250885, + 0.008852311410009861, + 0.0031478379387408495, + 0.002659131772816181, + 0.0037034261040389538, + 0.004089830908924341, + 0.003066839650273323, + 0.002456900430843234, + 0.004559301305562258, + 0.006375329568982124, + 0.0018548076041042805, + 0.0026761111803352833, + 0.002306394511833787, + 0.0028820442967116833, + 0.0025243228301405907, + 0.002569290343672037, + 0.0021686158142983913, + 0.0028163765091449022, + 0.011918492615222931, + 0.008749919012188911, + 0.019485076889395714, + 0.017239069566130638, + 0.014912574551999569, + 0.009307322092354298, + 0.02448238618671894, + 0.026039285585284233, + 0.01580304466187954, + 0.01171167939901352, + 0.016399435698986053, + 0.014803956262767315, + 0.026787012815475464, + 0.013400087133049965, + 0.026370657607913017, + 0.013578598387539387, + 0.006084277760237455, + 0.0036643296480178833, + 0.006737215910106897, + 0.006648529786616564, + 0.006366813089698553, + 0.0065815565176308155, + 0.00649906974285841, + 0.006274297367781401, + 0.015057471580803394, + 0.00827102456241846, + 0.015839124098420143, + 0.009527908638119698, + 0.015063615515828133, + 0.008326577953994274, + 0.014463365077972412, + 0.009520535357296467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.033707667142152786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107108686119318 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.01107108686119318, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107108686119318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2863747477531433, + 0.14196011424064636, + 0.07368933409452438, + 0.04046212509274483, + 0.060328204184770584, + 0.029793450608849525, + 0.07317425310611725, + 0.0576019287109375, + 0.06964870542287827, + 0.16696715354919434 + ], + "credit_map": [ + 0.04853815212845802, + 0.05975479260087013, + 0.028584493324160576, + 0.02685040608048439, + 0.035988833755254745, + 0.02434472180902958, + 0.040996212512254715, + 0.021317122504115105, + 0.01638263836503029, + 0.016956385225057602, + 0.02005642279982567, + 0.01981297694146633, + 0.012861954048275948, + 0.010839059017598629, + 0.02863493375480175, + 0.01641574688255787, + 0.009076133370399475, + 0.006934093777090311, + 0.011765713803470135, + 0.012530462816357613, + 0.008106947876513004, + 0.005890514701604843, + 0.00809394009411335, + 0.011291529051959515, + 0.0036106870975345373, + 0.006084871478378773, + 0.003929393365979195, + 0.005671881604939699, + 0.005899151787161827, + 0.005044003948569298, + 0.0032600145787000656, + 0.006962119601666927, + 0.00573534332215786, + 0.007011896930634975, + 0.0070190103724598885, + 0.00880389753729105, + 0.007252277806401253, + 0.007356690708547831, + 0.008878608234226704, + 0.008270478807389736, + 0.003033579559996724, + 0.004410686902701855, + 0.0035731156822293997, + 0.004590347874909639, + 0.003356531262397766, + 0.004224284086376429, + 0.003093897132202983, + 0.003511006711050868, + 0.007168565411120653, + 0.009804107248783112, + 0.008357838727533817, + 0.010281910188496113, + 0.013742809183895588, + 0.005534294527024031, + 0.009993573650717735, + 0.0082911541685462, + 0.0058231414295732975, + 0.004523118492215872, + 0.007498421240597963, + 0.00914281327277422, + 0.007780561689287424, + 0.01041927095502615, + 0.007746270392090082, + 0.0046683321706950665, + 0.007188538089394569, + 0.005536193959414959, + 0.007626009173691273, + 0.0081420186907053, + 0.012192659080028534, + 0.009786984883248806, + 0.01210528239607811, + 0.007071018684655428, + 0.017614109441637993, + 0.015240672044456005, + 0.021628623828291893, + 0.025089172646403313, + 0.02338748797774315, + 0.017773034051060677, + 0.031152760609984398, + 0.01508127897977829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01107108686119318 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003611353924497962 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.003611353924497962, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003611353924497962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5729563236236572, + 0.11739082634449005, + 0.03963000699877739, + 0.04080156609416008, + 0.03469089791178703, + 0.023226946592330933, + 0.0382133312523365, + 0.0228626299649477, + 0.03806836158037186, + 0.07215911895036697 + ], + "credit_map": [ + 0.14548319578170776, + 0.18948552012443542, + 0.04230315238237381, + 0.04704820364713669, + 0.027906358242034912, + 0.03767380490899086, + 0.04728434234857559, + 0.035771749913692474, + 0.023480387404561043, + 0.013040227815508842, + 0.021044006571173668, + 0.019466986879706383, + 0.0057920231483876705, + 0.009519286453723907, + 0.00905375275760889, + 0.015994152054190636, + 0.004641904029995203, + 0.005000559613108635, + 0.0053552729077637196, + 0.006597225554287434, + 0.004253895953297615, + 0.0040861209854483604, + 0.004637831822037697, + 0.005057197064161301, + 0.005575256422162056, + 0.006133218761533499, + 0.0046439277939498425, + 0.006899541709572077, + 0.005175498779863119, + 0.002998440060764551, + 0.003941832110285759, + 0.005433850456029177, + 0.003505486296489835, + 0.004308176692575216, + 0.0040628304705023766, + 0.005667558405548334, + 0.0045114909298717976, + 0.0033073981758207083, + 0.005034398753196001, + 0.004293556325137615, + 0.002366496017202735, + 0.003415516810491681, + 0.00288976333104074, + 0.003927325364202261, + 0.003272687317803502, + 0.002268357900902629, + 0.0018718527862802148, + 0.0032149474136531353, + 0.004149090964347124, + 0.004664548672735691, + 0.005608928389847279, + 0.005499245133250952, + 0.0038906694389879704, + 0.003753305645659566, + 0.005559440702199936, + 0.00508810393512249, + 0.0028108658734709024, + 0.002654177835211158, + 0.003012335393577814, + 0.003383200615644455, + 0.002721717581152916, + 0.003172209719195962, + 0.0028347496408969164, + 0.002273371908813715, + 0.005199488252401352, + 0.0029855130705982447, + 0.005261653568595648, + 0.0053348722867667675, + 0.004371709655970335, + 0.004613474942743778, + 0.005204672459512949, + 0.005096975713968277, + 0.014670309610664845, + 0.004575131461024284, + 0.014738007448613644, + 0.007147878408432007, + 0.009663629345595837, + 0.005393806844949722, + 0.010735685005784035, + 0.0052346657030284405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.003611353924497962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033678967505693436 + ], + "opposite_logprob_tokens": [ + -3.4086790084838867 + ], + "logprob_reward": -0.033678967505693436, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033678967505693436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2547794580459595, + 0.08692872524261475, + 0.08197085559368134, + 0.04362918809056282, + 0.06684128195047379, + 0.10172845423221588, + 0.08906125277280807, + 0.14666855335235596, + 0.053487956523895264, + 0.07490432262420654 + ], + "credit_map": [ + 0.06165764108300209, + 0.06278472393751144, + 0.01760479249060154, + 0.020818930119276047, + 0.023318607360124588, + 0.016668228432536125, + 0.03360434249043465, + 0.018322180956602097, + 0.008809631690382957, + 0.011499977670609951, + 0.012628341093659401, + 0.01003717165440321, + 0.01084474753588438, + 0.008431186899542809, + 0.016748685389757156, + 0.007928984239697456, + 0.008607250638306141, + 0.008483805693686008, + 0.009895737282931805, + 0.018100561574101448, + 0.008183029480278492, + 0.009989523328840733, + 0.00886266864836216, + 0.009848278015851974, + 0.005251743830740452, + 0.005840406287461519, + 0.003759708721190691, + 0.00528590427711606, + 0.007307569961994886, + 0.006022049114108086, + 0.0037572423461824656, + 0.006404564715921879, + 0.005716120358556509, + 0.010012865997850895, + 0.008927893824875355, + 0.010641509667038918, + 0.007518026977777481, + 0.007991212420165539, + 0.00766453193500638, + 0.008369119837880135, + 0.008002177812159061, + 0.01193702407181263, + 0.01026914268732071, + 0.013520068489015102, + 0.019182562828063965, + 0.010998229496181011, + 0.015086783096194267, + 0.012732472270727158, + 0.008477509021759033, + 0.00873994454741478, + 0.011216516606509686, + 0.00926148984581232, + 0.013640143908560276, + 0.008710809983313084, + 0.019598742946982384, + 0.009416098706424236, + 0.016860229894518852, + 0.012220007367432117, + 0.01817585714161396, + 0.016245098784565926, + 0.02653243951499462, + 0.011484198272228241, + 0.029215434566140175, + 0.01593528874218464, + 0.0062012081034481525, + 0.005331241060048342, + 0.006855680141597986, + 0.00634941877797246, + 0.007595376577228308, + 0.005315064452588558, + 0.009236703626811504, + 0.006603264715522528, + 0.011078083887696266, + 0.0051668556407094, + 0.012203611433506012, + 0.008090249262750149, + 0.008803177624940872, + 0.0057902866974473, + 0.015165839344263077, + 0.008606219664216042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.033678967505693436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-and-slightly-closed shot takes us deep into a pristine jungle, where a lone explorer stands before a majestic ancient stone gate. He wears a weathered, multi-pocket khaki jacket, its cuffs and hem stained with dirt, a dark gray T-shirt, durable dark overalls, and muddy hiking boots. He appears determined yet slightly exhausted. The gate before him is massive and majestic, carved from a single block of dark rock. Its surface is covered with intricate relief carvings eroded by time. Thick, dark green vines coil around it like giant pythons, almost blending it into the surrounding forest. In the damp, misty air, the explorer, hesitant yet resolute, slowly raises his right hand, clenching it into a fist. Before striking the cold, mossy stone, his fist pauses in mid-air. Dappled sunlight penetrates the dense treetops overhead, casting beams of light and illuminating the dust particles floating in the air. The entire picture is dominated by dark green and earth tones, creating a mysterious, solemn and suspenseful atmosphere, as if a long-buried secret is about to be revealed.", + "seed": 3214, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_3378_seed_3214_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": -0.06721695363521576, + "vlm_logprob": -0.06721695363521576, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.06721695363521576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19213812053203583 + ], + "opposite_logprob_tokens": [ + -1.9421380758285522 + ], + "logprob_reward": -0.19213812053203583, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.19213812053203583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15790587663650513, + 0.10810065269470215, + 0.12195911258459091, + 0.1751355081796646, + 0.07898680120706558, + 0.15568919479846954, + 0.11276786774396896, + 0.0329166054725647, + 0.028637604787945747, + 0.02790079265832901 + ], + "credit_map": [ + 0.015217581763863564, + 0.020657077431678772, + 0.022419923916459084, + 0.025461925193667412, + 0.017000677064061165, + 0.02052115648984909, + 0.018037889152765274, + 0.018589656800031662, + 0.013103135861456394, + 0.018087295815348625, + 0.00993309821933508, + 0.020176423713564873, + 0.007858053781092167, + 0.023569725453853607, + 0.006616073660552502, + 0.0087568499147892, + 0.0076835425570607185, + 0.019210271537303925, + 0.020587479695677757, + 0.012924714013934135, + 0.013241933658719063, + 0.0278099924325943, + 0.00925244390964508, + 0.011248734779655933, + 0.009677501395344734, + 0.015513253398239613, + 0.025024428963661194, + 0.022855490446090698, + 0.023371851071715355, + 0.04131442308425903, + 0.015524613671004772, + 0.021853934973478317, + 0.00488653127104044, + 0.00947575643658638, + 0.010457721538841724, + 0.010216454043984413, + 0.008940842002630234, + 0.014962567947804928, + 0.004207317717373371, + 0.015839610248804092, + 0.006445619277656078, + 0.013445318676531315, + 0.010769380256533623, + 0.02358710765838623, + 0.016727164387702942, + 0.03859566152095795, + 0.010842126794159412, + 0.035276804119348526, + 0.008216854184865952, + 0.011976317502558231, + 0.008798942901194096, + 0.02164945937693119, + 0.018169136717915535, + 0.0168770644813776, + 0.014218764379620552, + 0.012861327268183231, + 0.0021993869449943304, + 0.00263174157589674, + 0.002919408492743969, + 0.006624769419431686, + 0.0028719771653413773, + 0.005625621415674686, + 0.004234527703374624, + 0.005809175316244364, + 0.0016762504819780588, + 0.0020280613098293543, + 0.0022898458410054445, + 0.0034928233362734318, + 0.0038595215883105993, + 0.004840676207095385, + 0.004290856886655092, + 0.006159567274153233, + 0.002384909661486745, + 0.0030671542044728994, + 0.0025958996266126633, + 0.00518745556473732, + 0.0026447921991348267, + 0.0034811843652278185, + 0.005092871841043234, + 0.0034465258941054344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.19213812053203583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014217865653336048 + ], + "opposite_logprob_tokens": [ + -4.264217853546143 + ], + "logprob_reward": -0.014217865653336048, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014217865653336048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1671181172132492, + 0.15462125837802887, + 0.16050976514816284, + 0.1177164688706398, + 0.06403110921382904, + 0.05517587065696716, + 0.068714439868927, + 0.115083709359169, + 0.0482252836227417, + 0.048803944140672684 + ], + "credit_map": [ + 0.0179753415286541, + 0.027411753311753273, + 0.01994195394217968, + 0.025484297424554825, + 0.014913414604961872, + 0.027655966579914093, + 0.017019610852003098, + 0.016715770587325096, + 0.017449956387281418, + 0.029649432748556137, + 0.01760505698621273, + 0.026745013892650604, + 0.012934060767292976, + 0.021928895264863968, + 0.011412817053496838, + 0.016896028071641922, + 0.010182034224271774, + 0.023319652304053307, + 0.029835781082510948, + 0.017202379181981087, + 0.014583675190806389, + 0.041547778993844986, + 0.010397501289844513, + 0.013440951704978943, + 0.006807129830121994, + 0.014434851706027985, + 0.014424611814320087, + 0.015808280557394028, + 0.02277720905840397, + 0.016337236389517784, + 0.013792069628834724, + 0.013335078954696655, + 0.003663289826363325, + 0.009233798831701279, + 0.008239315822720528, + 0.010137541219592094, + 0.00863979384303093, + 0.007678892929106951, + 0.004995537456125021, + 0.011442938819527626, + 0.004627852700650692, + 0.00543274125084281, + 0.005575955379754305, + 0.010370015166699886, + 0.00765171879902482, + 0.008168286643922329, + 0.008219498209655285, + 0.005129803903400898, + 0.005279162898659706, + 0.007558329962193966, + 0.004946615546941757, + 0.010294487699866295, + 0.013748946599662304, + 0.010655906051397324, + 0.009194827638566494, + 0.007036165334284306, + 0.005193302407860756, + 0.011508990079164505, + 0.0071939886547625065, + 0.025651518255472183, + 0.01690949872136116, + 0.018516357988119125, + 0.016085172072052956, + 0.014024878852069378, + 0.0037536032032221556, + 0.005426712334156036, + 0.00437517324462533, + 0.007828675210475922, + 0.006932858843356371, + 0.006382620427757502, + 0.007022262550890446, + 0.006503376644104719, + 0.005304343067109585, + 0.005454014055430889, + 0.005791439209133387, + 0.007022963371127844, + 0.006053346209228039, + 0.006487272214144468, + 0.006283496040850878, + 0.006407072767615318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014217865653336048 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002492994070053 + ], + "opposite_logprob_tokens": [ + -2.6950249671936035 + ], + "logprob_reward": -0.07002492994070053, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002492994070053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17434869706630707, + 0.34706932306289673, + 0.038991380482912064, + 0.08760648965835571, + 0.09873537719249725, + 0.0544426403939724, + 0.03628171980381012, + 0.09912095963954926, + 0.025139622390270233, + 0.038263820111751556 + ], + "credit_map": [ + 0.01721647009253502, + 0.028543608263134956, + 0.0197039395570755, + 0.02738208696246147, + 0.017905745655298233, + 0.029158340767025948, + 0.01764427125453949, + 0.016794228926301003, + 0.03579180687665939, + 0.06500572711229324, + 0.038738466799259186, + 0.06393065303564072, + 0.0357879176735878, + 0.051231350749731064, + 0.02485942468047142, + 0.031723957508802414, + 0.002837304724380374, + 0.003920661751180887, + 0.006953064817935228, + 0.0039292387664318085, + 0.004102409817278385, + 0.009301182813942432, + 0.003585199126973748, + 0.004362316336482763, + 0.007717113010585308, + 0.008426852524280548, + 0.013422227464616299, + 0.008344137109816074, + 0.010101747699081898, + 0.02136564813554287, + 0.007496115285903215, + 0.010732647962868214, + 0.005018338095396757, + 0.014948037452995777, + 0.012897927314043045, + 0.013598984107375145, + 0.012480850331485271, + 0.013829507865011692, + 0.007262781262397766, + 0.018698934465646744, + 0.0021633037831634283, + 0.004519632086157799, + 0.003940504509955645, + 0.006496725138276815, + 0.003714150283485651, + 0.015878133475780487, + 0.003845857223495841, + 0.013884330168366432, + 0.001974561018869281, + 0.003474715631455183, + 0.0025700984988361597, + 0.006116305012255907, + 0.006022332236170769, + 0.0050815013237297535, + 0.004990569781512022, + 0.006051634903997183, + 0.004082150757312775, + 0.011499537155032158, + 0.00594952842220664, + 0.023555317893624306, + 0.011884305626153946, + 0.015592808835208416, + 0.014559442177414894, + 0.011997862718999386, + 0.0023160334676504135, + 0.002304008463397622, + 0.0025141865480691195, + 0.0031090856064110994, + 0.0034461840987205505, + 0.003395630046725273, + 0.00361420470289886, + 0.004440287593752146, + 0.003671635640785098, + 0.0037543687503784895, + 0.003628319827839732, + 0.00543283112347126, + 0.004308256786316633, + 0.005060513038188219, + 0.008307347074151039, + 0.004100540652871132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.07002492994070053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17176401615142822, + 0.2892864942550659, + 0.08163401484489441, + 0.06304091960191727, + 0.02948661893606186, + 0.07745614647865295, + 0.05288749188184738, + 0.12989184260368347, + 0.050532177090644836, + 0.05402025580406189 + ], + "credit_map": [ + 0.015711067244410515, + 0.0278032086789608, + 0.021556802093982697, + 0.029207706451416016, + 0.0186968594789505, + 0.025959575548768044, + 0.01819637417793274, + 0.01463241782039404, + 0.02875818870961666, + 0.051249802112579346, + 0.028992651030421257, + 0.055696506053209305, + 0.0323556512594223, + 0.044673725962638855, + 0.02244688756763935, + 0.025113074108958244, + 0.005277108401060104, + 0.009042731486260891, + 0.015421027317643166, + 0.007667935453355312, + 0.00691172294318676, + 0.02282305434346199, + 0.005042210686951876, + 0.009448218159377575, + 0.007408422883599997, + 0.005368018988519907, + 0.008974690921604633, + 0.006653033196926117, + 0.007053027395159006, + 0.01448750402778387, + 0.005270196124911308, + 0.007826020009815693, + 0.0012767943553626537, + 0.004056438338011503, + 0.003720517735928297, + 0.005432584322988987, + 0.003744968678802252, + 0.004319673404097557, + 0.0023938422091305256, + 0.004541801754385233, + 0.0031832323875278234, + 0.004853254649788141, + 0.005489908158779144, + 0.009035865776240826, + 0.003567090956494212, + 0.024589423090219498, + 0.004047210793942213, + 0.022690169513225555, + 0.0030498644337058067, + 0.005742175504565239, + 0.00408133352175355, + 0.009418494999408722, + 0.006901576183736324, + 0.008236844092607498, + 0.0072997743263840675, + 0.008157427422702312, + 0.007324379403144121, + 0.013497520238161087, + 0.008996385149657726, + 0.025736723095178604, + 0.013182529248297215, + 0.021596642211079597, + 0.018621688708662987, + 0.02093598246574402, + 0.0033363921102136374, + 0.0035988749004900455, + 0.006212604232132435, + 0.00585022009909153, + 0.004135598428547382, + 0.007728284224867821, + 0.007432600948959589, + 0.012237600050866604, + 0.002477259375154972, + 0.004909397102892399, + 0.004208394791930914, + 0.009621675126254559, + 0.00494746770709753, + 0.007246418856084347, + 0.013959088362753391, + 0.006650553084909916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.011071558110415936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048632293939590454 + ], + "opposite_logprob_tokens": [ + -3.0486323833465576 + ], + "logprob_reward": -0.048632293939590454, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048632293939590454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1182825118303299, + 0.07836944609880447, + 0.2055913358926773, + 0.111955426633358, + 0.03182997182011604, + 0.1694779396057129, + 0.049398332834243774, + 0.11434459686279297, + 0.044307176023721695, + 0.07644321769475937 + ], + "credit_map": [ + 0.01172044686973095, + 0.018073519691824913, + 0.015700705349445343, + 0.017835641279816628, + 0.011060971766710281, + 0.01580696739256382, + 0.012897093780338764, + 0.015187163837254047, + 0.007872258313000202, + 0.009701672941446304, + 0.01185204740613699, + 0.013444025069475174, + 0.0070853084325790405, + 0.010018897242844105, + 0.008017835207283497, + 0.010377400554716587, + 0.013330494053661823, + 0.040530674159526825, + 0.030619751662015915, + 0.03490220010280609, + 0.014541957527399063, + 0.039443086832761765, + 0.01321409735828638, + 0.019009068608283997, + 0.0061438679695129395, + 0.011872919276356697, + 0.015353690832853317, + 0.01751752570271492, + 0.015996884554624557, + 0.021791979670524597, + 0.011407251469790936, + 0.011871309019625187, + 0.0031546070240437984, + 0.0030051060020923615, + 0.004436450079083443, + 0.004194990731775761, + 0.0030058203265070915, + 0.006731831002980471, + 0.0020631083752959967, + 0.005238059442490339, + 0.005716313607990742, + 0.01171583216637373, + 0.009518386796116829, + 0.011930928565561771, + 0.019469942897558212, + 0.054013691842556, + 0.00946767721325159, + 0.047645170241594315, + 0.003077905159443617, + 0.005141471046954393, + 0.003240949474275112, + 0.008320515975356102, + 0.008665954694151878, + 0.007554689422249794, + 0.007338945288211107, + 0.006057903170585632, + 0.00504244165495038, + 0.013300971128046513, + 0.006674150004982948, + 0.02728337049484253, + 0.01402719970792532, + 0.018231570720672607, + 0.015473447740077972, + 0.014311442151665688, + 0.0034983756486326456, + 0.005329668987542391, + 0.004889678210020065, + 0.00601702556014061, + 0.005741320084780455, + 0.005596538539975882, + 0.006856217514723539, + 0.0063783531077206135, + 0.004733134526759386, + 0.010533630847930908, + 0.005647681653499603, + 0.010935721918940544, + 0.010853822343051434, + 0.009784003719687462, + 0.014115620404481888, + 0.009839602746069431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.048632293939590454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward-and-slightly-closed shot takes us deep into a pristine jungle, where a lone explorer stands before a majestic ancient stone gate. He wears a weathered, multi-pocket khaki jacket, its cuffs and hem stained with dirt, a dark gray T-shirt, durable dark overalls, and muddy hiking boots. He appears determined yet slightly exhausted. The gate before him is massive and majestic, carved from a single block of dark rock. Its surface is covered with intricate relief carvings eroded by time. Thick, dark green vines coil around it like giant pythons, almost blending it into the surrounding forest. In the damp, misty air, the explorer, hesitant yet resolute, slowly raises his right hand, clenching it into a fist. Before striking the cold, mossy stone, his fist pauses in mid-air. Dappled sunlight penetrates the dense treetops overhead, casting beams of light and illuminating the dust particles floating in the air. The entire picture is dominated by dark green and earth tones, creating a mysterious, solemn and suspenseful atmosphere, as if a long-buried secret is about to be revealed.", + "seed": 3215, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_2/index_3378_seed_3215_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": -0.056747983768582345, + "vlm_logprob": -0.056747983768582345, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.056747983768582345, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22643251717090607 + ], + "opposite_logprob_tokens": [ + -1.7264325618743896 + ], + "logprob_reward": -0.22643251717090607, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22643251717090607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17948047816753387, + 0.07123897969722748, + 0.16783525049686432, + 0.08733953535556793, + 0.11283177137374878, + 0.06411813199520111, + 0.12174244970083237, + 0.100455641746521, + 0.04009736329317093, + 0.054860394448041916 + ], + "credit_map": [ + 0.013928269036114216, + 0.024388134479522705, + 0.02250359021127224, + 0.0297718346118927, + 0.021495094522833824, + 0.02931065484881401, + 0.021660562604665756, + 0.0164223313331604, + 0.005192066542804241, + 0.0044603049755096436, + 0.0059274244122207165, + 0.01575973443686962, + 0.0102066770195961, + 0.008149644359946251, + 0.008585111238062382, + 0.01295801904052496, + 0.022564172744750977, + 0.01384440716356039, + 0.014711430296301842, + 0.018363965675234795, + 0.036223795264959335, + 0.021497538313269615, + 0.017804525792598724, + 0.022825414314866066, + 0.00948566198348999, + 0.007245814427733421, + 0.008405838161706924, + 0.00765986880287528, + 0.017835743725299835, + 0.012597660534083843, + 0.01189474854618311, + 0.01221420243382454, + 0.011457490734755993, + 0.011659478768706322, + 0.014857858419418335, + 0.010151275433599949, + 0.010742075741291046, + 0.011784179136157036, + 0.022495843470096588, + 0.019683577120304108, + 0.005864517763257027, + 0.007362504489719868, + 0.010006668977439404, + 0.005139917600899935, + 0.010260739363729954, + 0.00502236420288682, + 0.012396353296935558, + 0.00806506909430027, + 0.016033990308642387, + 0.019907305017113686, + 0.01851196587085724, + 0.01441500149667263, + 0.015220179222524166, + 0.009297097101807594, + 0.01482008770108223, + 0.013536821119487286, + 0.009919891133904457, + 0.010316946543753147, + 0.01584881730377674, + 0.012340283952653408, + 0.013310350477695465, + 0.011143459007143974, + 0.01633378118276596, + 0.01124211773276329, + 0.003646757686510682, + 0.0026422441005706787, + 0.006539938505738974, + 0.0059750317595899105, + 0.00580219179391861, + 0.0033726966939866543, + 0.007518020458519459, + 0.004600481130182743, + 0.006151791196316481, + 0.004584873095154762, + 0.00869912002235651, + 0.004654493648558855, + 0.010180117562413216, + 0.005354448687285185, + 0.009457165375351906, + 0.0057783885858953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.22643251717090607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964931100606918 + ], + "opposite_logprob_tokens": [ + -5.130964756011963 + ], + "logprob_reward": -0.005964931100606918, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964931100606918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17021271586418152, + 0.07030147314071655, + 0.13726702332496643, + 0.06007767841219902, + 0.15978194773197174, + 0.058222874999046326, + 0.15187399089336395, + 0.07069876044988632, + 0.030433062463998795, + 0.09113049507141113 + ], + "credit_map": [ + 0.013511206954717636, + 0.020710092037916183, + 0.023578407242894173, + 0.021230651065707207, + 0.02088010497391224, + 0.024557199329137802, + 0.02429652214050293, + 0.02144853211939335, + 0.006702046375721693, + 0.004511076956987381, + 0.008256982080638409, + 0.009779642336070538, + 0.010356244631111622, + 0.011717824265360832, + 0.009667837992310524, + 0.009309816174209118, + 0.01793213188648224, + 0.011859309859573841, + 0.01267478708177805, + 0.01407087966799736, + 0.028707226738333702, + 0.013428863137960434, + 0.019748060032725334, + 0.018845757469534874, + 0.005866037216037512, + 0.0045866225846111774, + 0.0054802848026156425, + 0.004131845664232969, + 0.013825180009007454, + 0.008754365146160126, + 0.010362224653363228, + 0.007071116007864475, + 0.01564170978963375, + 0.011941294185817242, + 0.01760980300605297, + 0.01193246804177761, + 0.03288443386554718, + 0.022775469347834587, + 0.020606668666005135, + 0.026390114799141884, + 0.0034803180024027824, + 0.00814586691558361, + 0.008207926526665688, + 0.0046752868220210075, + 0.008376491256058216, + 0.0075056967325508595, + 0.009158330038189888, + 0.008672963827848434, + 0.014789972454309464, + 0.022443464025855064, + 0.024870101362466812, + 0.012460118159651756, + 0.032231174409389496, + 0.011384347453713417, + 0.018640797585248947, + 0.01505400612950325, + 0.005232127849012613, + 0.008143778890371323, + 0.012624720111489296, + 0.00523257814347744, + 0.010640973225235939, + 0.006394688505679369, + 0.01457199826836586, + 0.007857891730964184, + 0.003110748017206788, + 0.00257564103230834, + 0.004760073963552713, + 0.003974315244704485, + 0.004479629918932915, + 0.0032871884759515524, + 0.0046088676899671555, + 0.0036365974228829145, + 0.0072680991142988205, + 0.012626850977540016, + 0.017189795151352882, + 0.010992751456797123, + 0.011234378442168236, + 0.008539034985005856, + 0.009604373015463352, + 0.013675214722752571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.005964931100606918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635676465928555 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635676465928555, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635676465928555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15334685146808624, + 0.3007796108722687, + 0.18771418929100037, + 0.06562469899654388, + 0.0435795783996582, + 0.10034530609846115, + 0.04905224218964577, + 0.024264561012387276, + 0.04174262285232544, + 0.03355032950639725 + ], + "credit_map": [ + 0.01427715364843607, + 0.01881004124879837, + 0.021913209930062294, + 0.021681951358914375, + 0.019101303070783615, + 0.022347712889313698, + 0.01350993849337101, + 0.02170555107295513, + 0.025959379971027374, + 0.019205065444111824, + 0.03954249247908592, + 0.05180022120475769, + 0.042756587266922, + 0.03169562295079231, + 0.03669973835349083, + 0.053120486438274384, + 0.018959492444992065, + 0.023724915459752083, + 0.014542019926011562, + 0.02589298225939274, + 0.029235251247882843, + 0.03288682922720909, + 0.021729180589318275, + 0.020743532106280327, + 0.007649810519069433, + 0.005288991145789623, + 0.007210935465991497, + 0.006089816335588694, + 0.011444290168583393, + 0.0072502014227211475, + 0.01258005853742361, + 0.008110597729682922, + 0.004730617627501488, + 0.005943304859101772, + 0.006855521351099014, + 0.00554871978238225, + 0.005122049245983362, + 0.00354796857573092, + 0.005445990711450577, + 0.0063854088075459, + 0.007708309683948755, + 0.01367307361215353, + 0.013561716303229332, + 0.008461032062768936, + 0.01701057329773903, + 0.010169139131903648, + 0.018158337101340294, + 0.011603116057813168, + 0.006483763922005892, + 0.008495602756738663, + 0.00793452002108097, + 0.005852557253092527, + 0.0072202119044959545, + 0.004511556588113308, + 0.005800295155495405, + 0.0027537336573004723, + 0.002027420327067375, + 0.002549668774008751, + 0.004146563354879618, + 0.0017429711297154427, + 0.003840084420517087, + 0.0023027067072689533, + 0.0046167876571416855, + 0.003038358176127076, + 0.004357360303401947, + 0.004387913271784782, + 0.007108707446604967, + 0.009821169078350067, + 0.0031758632976561785, + 0.004851046483963728, + 0.00386167224496603, + 0.004178893286734819, + 0.003542785532772541, + 0.003999113105237484, + 0.006236649584025145, + 0.002952957758679986, + 0.004859102889895439, + 0.004141880199313164, + 0.00417027622461319, + 0.0036475672386586666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.008635676465928555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632099531590939 + ], + "opposite_logprob_tokens": [ + -5.379631996154785 + ], + "logprob_reward": -0.004632099531590939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632099531590939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26949137449264526, + 0.19867472350597382, + 0.24224863946437836, + 0.054303862154483795, + 0.029632490128278732, + 0.06957798451185226, + 0.041696663945913315, + 0.031639184802770615, + 0.025370929390192032, + 0.03736414387822151 + ], + "credit_map": [ + 0.016832629218697548, + 0.033096663653850555, + 0.033567629754543304, + 0.03248161822557449, + 0.035169947892427444, + 0.03555504232645035, + 0.04173361882567406, + 0.041054222732782364, + 0.015524483285844326, + 0.013085813261568546, + 0.030357034876942635, + 0.030975501984357834, + 0.021945152431726456, + 0.030880620703101158, + 0.030419498682022095, + 0.025486601516604424, + 0.024735786020755768, + 0.019045796245336533, + 0.023686112836003304, + 0.020430920645594597, + 0.03777945786714554, + 0.04777805879712105, + 0.037150051444768906, + 0.03164243698120117, + 0.0056660957634449005, + 0.004630371928215027, + 0.00574745237827301, + 0.0047071874141693115, + 0.012052236124873161, + 0.0059325252659618855, + 0.008399615995585918, + 0.007168369833379984, + 0.0029656842816621065, + 0.002506023971363902, + 0.004009016323834658, + 0.0035436623729765415, + 0.004630648996680975, + 0.003531625960022211, + 0.003829378867521882, + 0.004616446327418089, + 0.004969101399183273, + 0.00929185375571251, + 0.010948597453534603, + 0.005951758474111557, + 0.009200099855661392, + 0.006841952912509441, + 0.015008870512247086, + 0.007365744095295668, + 0.004936326760798693, + 0.007400211878120899, + 0.006368577014654875, + 0.005174712277948856, + 0.006692111026495695, + 0.0034527985844761133, + 0.00500106206163764, + 0.0026708587538450956, + 0.004254715517163277, + 0.0031488509848713875, + 0.005402449052780867, + 0.005341622978448868, + 0.0035381943453103304, + 0.0038607444148510695, + 0.003556199837476015, + 0.002536405110731721, + 0.0033044111914932728, + 0.0023746138904243708, + 0.004270919132977724, + 0.0043227532878518105, + 0.0026344857178628445, + 0.002706157509237528, + 0.003378462977707386, + 0.002379125915467739, + 0.0036317468620836735, + 0.004420930054038763, + 0.006386683322489262, + 0.005109245423227549, + 0.0056869941763579845, + 0.004529599566012621, + 0.0044670552015304565, + 0.003131889970973134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004632099531590939 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot slow, upward, and slightly closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a lone explorer wearing a weathered khaki jacket, dark gray T-shirt, dark overalls, and muddy hiking boots visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive ancient stone gate carved from a single block of dark rock with intricate relief carvings visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thick dark green vines coiling around the gate visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the explorer slowly raise his right hand, clench it into a fist, and pause it in mid-air before the gate?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807469457387924 + ], + "opposite_logprob_tokens": [ + -3.2880747318267822 + ], + "logprob_reward": -0.03807469457387924, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807469457387924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14521697163581848, + 0.03798557072877884, + 0.08263185620307922, + 0.08036872744560242, + 0.13021224737167358, + 0.33372679352760315, + 0.08087664842605591, + 0.03657807782292366, + 0.04986019805073738, + 0.02254294976592064 + ], + "credit_map": [ + 0.00954309944063425, + 0.00998611655086279, + 0.012457672506570816, + 0.013834141194820404, + 0.03897074982523918, + 0.031526755541563034, + 0.016597187146544456, + 0.012301245704293251, + 0.0035642606671899557, + 0.0021799439564347267, + 0.004222166258841753, + 0.004960858263075352, + 0.004599251318722963, + 0.006295834667980671, + 0.006378640420734882, + 0.00578461354598403, + 0.009324008598923683, + 0.006401513237506151, + 0.009094367735087872, + 0.008023709058761597, + 0.017263786867260933, + 0.011791175231337547, + 0.00928708165884018, + 0.011446213349699974, + 0.009192832745611668, + 0.005421712528914213, + 0.008782219141721725, + 0.006814406253397465, + 0.01397909875959158, + 0.011461541056632996, + 0.014119992032647133, + 0.010596919804811478, + 0.011607445776462555, + 0.010860553942620754, + 0.014161955565214157, + 0.010314260609447956, + 0.023856783285737038, + 0.013146871700882912, + 0.02496931701898575, + 0.02129504457116127, + 0.0273884367197752, + 0.048049163073301315, + 0.043386150151491165, + 0.02848845161497593, + 0.06149795651435852, + 0.032776977866888046, + 0.04884142056107521, + 0.043298203498125076, + 0.014009877108037472, + 0.011290096677839756, + 0.013156810775399208, + 0.011085418984293938, + 0.009155536070466042, + 0.006034181918948889, + 0.010511825792491436, + 0.00563290249556303, + 0.004353034310042858, + 0.004159236326813698, + 0.005769835785031319, + 0.00569227896630764, + 0.005526663735508919, + 0.002717500552535057, + 0.0054084667935967445, + 0.0029510576277971268, + 0.005975661333650351, + 0.004145678598433733, + 0.00881884340196848, + 0.009055882692337036, + 0.004894395358860493, + 0.004893400706350803, + 0.00697214063256979, + 0.005104192066937685, + 0.0016646039439365268, + 0.0028879435267299414, + 0.004145675338804722, + 0.00261349743232131, + 0.002617862308397889, + 0.002147540682926774, + 0.0030062978621572256, + 0.0034595278557389975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03807469457387924 + } + ], + "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/0000122/group_3/index_tensor([1748, 1748])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_tensor([1748, 1748])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..804308e36a55ce52d6be0875c8382459155fe76e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_tensor([1748, 1748])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a static, eye-level perspective within a moving vehicle, the scene unfolds within the harsh, concrete confines of the long tunnel, creating a focused and immersive point-of-view shot. A fiery red sports car, with a low-slung, aerodynamic body and a gleaming, high-gloss finish, initially occupies the very center of the frame. Overhead fluorescent lights pulse rhythmically across its polished surface, accentuating its sleek silhouette. To the right of the screen, a dark green sedan travels in the adjacent lane, slightly ahead of the camera vehicle. Without warning, the red sports car accelerates swiftly and decisively; its dual taillights flash brightly for a moment before surging forward, its powerful body rapidly pulling away down the straight tunnel. As the distance increases, the concrete walls and ceiling blur into gray streaks, but the vibrant red car remains the undeniable focal point, creating a palpable atmosphere of speed, power, and exciting forward momentum.", + "seed": 3605, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_1748_seed_3605_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": -0.0938755703624338, + "vlm_logprob": -0.0938755703624338, + "vlm_yes_no_margin": 3.2249999761581423, + "vlm_token_credit": -0.0938755703624338, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2541750371456146 + ], + "opposite_logprob_tokens": [ + -2.0041749477386475 + ], + "logprob_reward": -0.2541750371456146, + "yes_no_margin_reward": 1.7499998807907104, + "token_credit_reward": -0.2541750371456146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05838989093899727, + 0.018190890550613403, + 0.020860187709331512, + 0.21613864600658417, + 0.10974036902189255, + 0.041594695299863815, + 0.39564892649650574, + 0.04765651747584343, + 0.06598944962024689, + 0.025790398940443993 + ], + "credit_map": [ + 0.008383006788790226, + 0.012226022779941559, + 0.008558059111237526, + 0.005046146456152201, + 0.006823813542723656, + 0.009290390647947788, + 0.004260908346623182, + 0.003801547223702073, + 0.0021220275666564703, + 0.0028894811403006315, + 0.0025544946547597647, + 0.0016067675314843655, + 0.0017808630364015698, + 0.0037615189794451, + 0.001915939268656075, + 0.001559801516123116, + 0.004949322901666164, + 0.004962114617228508, + 0.0017067397711798549, + 0.0014205327024683356, + 0.0016312580555677414, + 0.0035770186223089695, + 0.0011814194731414318, + 0.001431783544830978, + 0.013639461249113083, + 0.09021245688199997, + 0.005377250257879496, + 0.003643106436356902, + 0.00870317779481411, + 0.08063460141420364, + 0.004099756013602018, + 0.009828845970332623, + 0.014041517861187458, + 0.015516318380832672, + 0.02130167931318283, + 0.007529516238719225, + 0.011586621403694153, + 0.019400661811232567, + 0.011188630945980549, + 0.009175430983304977, + 0.005974797066301107, + 0.004361335653811693, + 0.007580663077533245, + 0.0036994782276451588, + 0.004787919577211142, + 0.0074990796856582165, + 0.004471919033676386, + 0.0032195085659623146, + 0.054993271827697754, + 0.16859064996242523, + 0.008594537153840065, + 0.01119330246001482, + 0.015725569799542427, + 0.112946517765522, + 0.010657439939677715, + 0.0129476273432374, + 0.004968744236975908, + 0.007078636437654495, + 0.010661963373422623, + 0.005580944009125233, + 0.007848040200769901, + 0.004503108561038971, + 0.0032468708232045174, + 0.0037682128604501486, + 0.0070299021899700165, + 0.012148908339440823, + 0.008167131803929806, + 0.010918015614151955, + 0.005369191989302635, + 0.01036799605935812, + 0.006536118220537901, + 0.005452191922813654, + 0.00410129688680172, + 0.0023435030598193407, + 0.004439470358192921, + 0.002798401052132249, + 0.0016664544818922877, + 0.004814261570572853, + 0.0023729363456368446, + 0.0032540778629481792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.2541750371456146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002847547410055995 + ], + "opposite_logprob_tokens": [ + -5.877847671508789 + ], + "logprob_reward": -0.002847547410055995, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002847547410055995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06278390437364578, + 0.04805988818407059, + 0.0324566587805748, + 0.43802499771118164, + 0.08256105333566666, + 0.05609296262264252, + 0.11959986388683319, + 0.0499601885676384, + 0.06957777589559555, + 0.04088268801569939 + ], + "credit_map": [ + 0.009979276917874813, + 0.01613522134721279, + 0.006370946764945984, + 0.006417297292500734, + 0.004485513549298048, + 0.010027251206338406, + 0.005282862577587366, + 0.004085536114871502, + 0.005264438688755035, + 0.006837810389697552, + 0.0095288697630167, + 0.006475141271948814, + 0.005347406025975943, + 0.0051754615269601345, + 0.006089015398174524, + 0.0033417444210499525, + 0.009180853143334389, + 0.004061268176883459, + 0.003930121194571257, + 0.0030897348187863827, + 0.0024437461979687214, + 0.004663740284740925, + 0.002936333417892456, + 0.0021508638747036457, + 0.022952260449528694, + 0.1729109287261963, + 0.013580350205302238, + 0.008432816714048386, + 0.022799687460064888, + 0.1638389378786087, + 0.009882917627692223, + 0.023627080023288727, + 0.009701860137283802, + 0.01251518540084362, + 0.010538645088672638, + 0.005206194240599871, + 0.015032668597996235, + 0.014797247014939785, + 0.0059713637456297874, + 0.00879789050668478, + 0.007461595349013805, + 0.00556354271247983, + 0.012383691966533661, + 0.005415224004536867, + 0.0051617431454360485, + 0.009472477249801159, + 0.007016429211944342, + 0.0036182610783725977, + 0.007388892117887735, + 0.04442253336310387, + 0.0062164487317204475, + 0.004144606180489063, + 0.004707932006567717, + 0.04189686104655266, + 0.005221846513450146, + 0.005600748583674431, + 0.004889967385679483, + 0.009455874562263489, + 0.006511054001748562, + 0.005155052989721298, + 0.009982243180274963, + 0.005290105938911438, + 0.0058168028481304646, + 0.0028590841684490442, + 0.007964562624692917, + 0.009213615208864212, + 0.01192636787891388, + 0.015622572042047977, + 0.004657572600990534, + 0.00576182221993804, + 0.009215494617819786, + 0.005215771961957216, + 0.005396642722189426, + 0.0052584754303097725, + 0.005519222468137741, + 0.007629692554473877, + 0.002863542176783085, + 0.005842156708240509, + 0.004017408937215805, + 0.004355546087026596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.002847547410055995 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037527561187744 + ], + "opposite_logprob_tokens": [ + -2.812037467956543 + ], + "logprob_reward": -0.062037527561187744, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037527561187744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18089419603347778, + 0.10354111343622208, + 0.05589780956506729, + 0.20573163032531738, + 0.03836033493280411, + 0.019134128466248512, + 0.30300092697143555, + 0.015093011781573296, + 0.05609595403075218, + 0.022250840440392494 + ], + "credit_map": [ + 0.02683771774172783, + 0.06560074537992477, + 0.01921631395816803, + 0.007340659387409687, + 0.01217536348849535, + 0.037314094603061676, + 0.006196882575750351, + 0.00621240958571434, + 0.014414765872061253, + 0.013810799457132816, + 0.020280463621020317, + 0.007749649230390787, + 0.013435782864689827, + 0.01753348857164383, + 0.004730147309601307, + 0.011586015112698078, + 0.008660209365189075, + 0.013475699350237846, + 0.005932204890996218, + 0.0043302299454808235, + 0.0057195695117115974, + 0.01027421560138464, + 0.0031366567127406597, + 0.004369023721665144, + 0.01177164725959301, + 0.081539586186409, + 0.007035133428871632, + 0.004628849681466818, + 0.01137852855026722, + 0.0748080164194107, + 0.004534913692623377, + 0.010034944862127304, + 0.004846292547881603, + 0.0047074854373931885, + 0.0060674878768622875, + 0.0028512245044112206, + 0.00435936264693737, + 0.007787839975208044, + 0.003744876943528652, + 0.003995764534920454, + 0.002686939900740981, + 0.0026453661266714334, + 0.002305566566064954, + 0.0009760567918419838, + 0.002838507993146777, + 0.003602541983127594, + 0.0015384414000436664, + 0.0025407082866877317, + 0.023088641464710236, + 0.12157458066940308, + 0.007302545942366123, + 0.007494370918720961, + 0.014943176880478859, + 0.10779019445180893, + 0.006659888196736574, + 0.014147557318210602, + 0.0016971621662378311, + 0.0025657485239207745, + 0.0018582289339974523, + 0.0016949664568528533, + 0.0027293399907648563, + 0.0014033779734745622, + 0.0018155449070036411, + 0.0013286422472447157, + 0.006372507195919752, + 0.01006869412958622, + 0.008122149854898453, + 0.010904861614108086, + 0.00420382572337985, + 0.005221759434789419, + 0.007191710639744997, + 0.004010442178696394, + 0.004169462248682976, + 0.0024168805684894323, + 0.003043045988306403, + 0.003265727311372757, + 0.0013788110809400678, + 0.004108302760869265, + 0.0018432700308039784, + 0.0020253395196050406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.062037527561187744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702873349189758 + ], + "opposite_logprob_tokens": [ + -2.127028703689575 + ], + "logprob_reward": -0.12702873349189758, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702873349189758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06871847808361053, + 0.021015258505940437, + 0.03150876611471176, + 0.19986432790756226, + 0.08674922585487366, + 0.029458554461598396, + 0.22122034430503845, + 0.044751863926649094, + 0.20837044715881348, + 0.08834275603294373 + ], + "credit_map": [ + 0.010979779064655304, + 0.020751548931002617, + 0.00666550500318408, + 0.0042405580170452595, + 0.005667225457727909, + 0.013397637754678726, + 0.004221862182021141, + 0.002794364932924509, + 0.0020907192956656218, + 0.0030996380373835564, + 0.0029607827309519053, + 0.0026782227214425802, + 0.0020967540331184864, + 0.0032985759899020195, + 0.0029621729627251625, + 0.001828393549658358, + 0.003821399761363864, + 0.004533146508038044, + 0.004617323167622089, + 0.004074324853718281, + 0.002848210046067834, + 0.004585713613778353, + 0.0041434261947870255, + 0.0028852205723524094, + 0.010337553918361664, + 0.07924523204565048, + 0.0067784651182591915, + 0.003929130267351866, + 0.00839948933571577, + 0.07540974766016006, + 0.005050748120993376, + 0.010713962838053703, + 0.010702401399612427, + 0.010129383765161037, + 0.017554670572280884, + 0.006134710740298033, + 0.008848775178194046, + 0.015371514484286308, + 0.009452172555029392, + 0.008555592969059944, + 0.0037531328853219748, + 0.0035671498626470566, + 0.004673849791288376, + 0.0023980794940143824, + 0.0039315433241426945, + 0.006029762327671051, + 0.002835434628650546, + 0.0022696005180478096, + 0.011608649976551533, + 0.08482207357883453, + 0.007048764266073704, + 0.006187709979712963, + 0.008901271037757397, + 0.08653406798839569, + 0.004690768197178841, + 0.011427042074501514, + 0.005668364930897951, + 0.010321617126464844, + 0.0049931262619793415, + 0.006351799704134464, + 0.0035419887863099575, + 0.005149145610630512, + 0.00597463408485055, + 0.002751188585534692, + 0.021630652248859406, + 0.03446727246046066, + 0.028647061437368393, + 0.040064796805381775, + 0.01687757670879364, + 0.020175082609057426, + 0.029955202713608742, + 0.016552796587347984, + 0.012556592002511024, + 0.010740143246948719, + 0.013286148197948933, + 0.013032002374529839, + 0.006683709565550089, + 0.015134157612919807, + 0.008974683471024036, + 0.007935315370559692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.12702873349189758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328900620341301 + ], + "opposite_logprob_tokens": [ + -3.7732889652252197 + ], + "logprob_reward": -0.02328900620341301, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328900620341301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08853429555892944, + 0.05640384554862976, + 0.08739937841892242, + 0.20912334322929382, + 0.1053394004702568, + 0.034719619899988174, + 0.1767459362745285, + 0.051026325672864914, + 0.13140207529067993, + 0.059305790811777115 + ], + "credit_map": [ + 0.010971841402351856, + 0.032806187868118286, + 0.007668808102607727, + 0.0031372385565191507, + 0.006797927897423506, + 0.019675517454743385, + 0.003542599268257618, + 0.003934176173061132, + 0.007139124907553196, + 0.007941262796521187, + 0.010218051262199879, + 0.005919094663113356, + 0.007182744797319174, + 0.008415302261710167, + 0.00518256239593029, + 0.004405701067298651, + 0.015242503955960274, + 0.018060412257909775, + 0.007544028572738171, + 0.005609961226582527, + 0.011905418708920479, + 0.015224020928144455, + 0.006388120818883181, + 0.007424913812428713, + 0.012740821577608585, + 0.08753788471221924, + 0.00622593192383647, + 0.005187197122722864, + 0.009376909583806992, + 0.07500910758972168, + 0.004825226031243801, + 0.008220268413424492, + 0.013783800415694714, + 0.011867069639265537, + 0.0132807781919837, + 0.006509554106742144, + 0.01669500209391117, + 0.02402782253921032, + 0.008776860311627388, + 0.010398517362773418, + 0.004183451645076275, + 0.003795035881921649, + 0.004765498451888561, + 0.0025612032040953636, + 0.00515751913189888, + 0.007700301241129637, + 0.0034502975177019835, + 0.003106315154582262, + 0.03984237089753151, + 0.07522387057542801, + 0.004526955541223288, + 0.00554411206394434, + 0.00789527129381895, + 0.03209403157234192, + 0.0067115603014826775, + 0.0049077607691287994, + 0.006831311620771885, + 0.009995673783123493, + 0.004159381613135338, + 0.0036770126316696405, + 0.010402402840554714, + 0.009002592414617538, + 0.002696752781048417, + 0.004261201247572899, + 0.013692810200154781, + 0.024086426943540573, + 0.013553047552704811, + 0.020636893808841705, + 0.010525483638048172, + 0.021102124825119972, + 0.015069392509758472, + 0.012735906057059765, + 0.009432092308998108, + 0.005815567448735237, + 0.008043856360018253, + 0.00503557687625289, + 0.0042357794009149075, + 0.014650661498308182, + 0.004027137532830238, + 0.008065118454396725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.02328900620341301 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, eye-level perspective within a moving vehicle, the scene unfolds within the harsh, concrete confines of the long tunnel, creating a focused and immersive point-of-view shot. A fiery red sports car, with a low-slung, aerodynamic body and a gleaming, high-gloss finish, initially occupies the very center of the frame. Overhead fluorescent lights pulse rhythmically across its polished surface, accentuating its sleek silhouette. To the right of the screen, a dark green sedan travels in the adjacent lane, slightly ahead of the camera vehicle. Without warning, the red sports car accelerates swiftly and decisively; its dual taillights flash brightly for a moment before surging forward, its powerful body rapidly pulling away down the straight tunnel. As the distance increases, the concrete walls and ceiling blur into gray streaks, but the vibrant red car remains the undeniable focal point, creating a palpable atmosphere of speed, power, and exciting forward momentum.", + "seed": 3606, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_1748_seed_3606_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": -0.10320570094045252, + "vlm_logprob": -0.10320570094045252, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.10320570094045252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18561191856861115 + ], + "opposite_logprob_tokens": [ + -2.4356119632720947 + ], + "logprob_reward": -0.18561191856861115, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.18561191856861115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0766204372048378, + 0.24141369760036469, + 0.20871533453464508, + 0.069499671459198, + 0.09608288109302521, + 0.0576886422932148, + 0.10605471581220627, + 0.048886969685554504, + 0.047131869941949844, + 0.0479058176279068 + ], + "credit_map": [ + 0.008783752098679543, + 0.01055073831230402, + 0.010806011967360973, + 0.007146123331040144, + 0.008049827069044113, + 0.009397841058671474, + 0.011269398033618927, + 0.010616747662425041, + 0.04253925755620003, + 0.027140310034155846, + 0.03332625329494476, + 0.03278852626681328, + 0.017478300258517265, + 0.026975398883223534, + 0.02811357192695141, + 0.033052071928977966, + 0.056172337383031845, + 0.011533769778907299, + 0.08702006191015244, + 0.013241872191429138, + 0.017326364293694496, + 0.009164337068796158, + 0.0074145011603832245, + 0.006842064205557108, + 0.0075929900631308556, + 0.006073624361306429, + 0.014191395603120327, + 0.010374343954026699, + 0.005985831841826439, + 0.006212806329131126, + 0.011548544280230999, + 0.007520132232457399, + 0.01955614984035492, + 0.007161282934248447, + 0.031112564727663994, + 0.008171909488737583, + 0.007218958809971809, + 0.009483235888183117, + 0.008371567353606224, + 0.005007212515920401, + 0.007309295237064362, + 0.0036245083902031183, + 0.011868560686707497, + 0.006179729476571083, + 0.005497863050550222, + 0.004763307515531778, + 0.008335782214999199, + 0.010109595023095608, + 0.01292745303362608, + 0.011555776000022888, + 0.01427608821541071, + 0.012153367511928082, + 0.014071960933506489, + 0.009767061099410057, + 0.018755031749606133, + 0.012547977268695831, + 0.0049003190360963345, + 0.0038298857398331165, + 0.007012194953858852, + 0.00823410227894783, + 0.004464545752853155, + 0.005826571024954319, + 0.008701053448021412, + 0.005918297916650772, + 0.0035924799740314484, + 0.002756425878033042, + 0.007200656458735466, + 0.012142089195549488, + 0.003498668083921075, + 0.004286277107894421, + 0.007284247316420078, + 0.006371024530380964, + 0.00372312031686306, + 0.005382298491895199, + 0.004065557848662138, + 0.006334536708891392, + 0.003692081430926919, + 0.003674939973279834, + 0.009733373299241066, + 0.011299912817776203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.18561191856861115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001739894854836166 + ], + "opposite_logprob_tokens": [ + -6.376739978790283 + ], + "logprob_reward": -0.001739894854836166, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001739894854836166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06882888078689575, + 0.0207124724984169, + 0.5569276809692383, + 0.022899655625224113, + 0.05257370322942734, + 0.04547874256968498, + 0.026477694511413574, + 0.030614584684371948, + 0.05812321975827217, + 0.11736331880092621 + ], + "credit_map": [ + 0.010292082093656063, + 0.006853668950498104, + 0.010792830027639866, + 0.008641107007861137, + 0.005271168425679207, + 0.008949033915996552, + 0.007999024353921413, + 0.010029965080320835, + 0.0019768676720559597, + 0.0023651958908885717, + 0.00232752226293087, + 0.0026226097252219915, + 0.002121698809787631, + 0.0017522074049338698, + 0.004282141570001841, + 0.0032642297446727753, + 0.15703293681144714, + 0.02925828844308853, + 0.23950356245040894, + 0.03213038668036461, + 0.04754294082522392, + 0.020187800750136375, + 0.016327643766999245, + 0.014944159425795078, + 0.0034400871954858303, + 0.0028131534345448017, + 0.004016713704913855, + 0.0023291956167668104, + 0.0026439945213496685, + 0.002709269290789962, + 0.0020303945057094097, + 0.0029168494511395693, + 0.006056024692952633, + 0.005284655839204788, + 0.009039021097123623, + 0.006982889026403427, + 0.006087669637054205, + 0.007111113518476486, + 0.0071715242229402065, + 0.004840805660933256, + 0.011149821802973747, + 0.00250222347676754, + 0.015444180928170681, + 0.0024425245355814695, + 0.0039058378897607327, + 0.0026051828172057867, + 0.002376060700044036, + 0.0050529115833342075, + 0.0032027990091592073, + 0.002481361385434866, + 0.004387847147881985, + 0.0025011072866618633, + 0.0032608325127512217, + 0.0031289684120565653, + 0.0035712982062250376, + 0.003943482413887978, + 0.003583294339478016, + 0.003411858109757304, + 0.004836536943912506, + 0.0026231957599520683, + 0.0039038255345076323, + 0.003709682496264577, + 0.0031567474361509085, + 0.005389444064348936, + 0.004961959086358547, + 0.0034296424128115177, + 0.011449560523033142, + 0.013930270448327065, + 0.004503041971474886, + 0.004818466026335955, + 0.008817874826490879, + 0.0062124039977788925, + 0.006517907604575157, + 0.007083421107381582, + 0.007606628350913525, + 0.018394947052001953, + 0.005334037821739912, + 0.005253720097243786, + 0.036638595163822174, + 0.030534062534570694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.001739894854836166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156507194042206 + ], + "opposite_logprob_tokens": [ + -1.7015650272369385 + ], + "logprob_reward": -0.20156507194042206, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156507194042206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0824061930179596, + 0.13827873766422272, + 0.20535063743591309, + 0.0469023771584034, + 0.03765365108847618, + 0.1091441661119461, + 0.04182809963822365, + 0.0660826563835144, + 0.1813550889492035, + 0.09099841117858887 + ], + "credit_map": [ + 0.009491668082773685, + 0.009556584991514683, + 0.011339658871293068, + 0.011054744943976402, + 0.006829301826655865, + 0.012360445223748684, + 0.0087409857660532, + 0.013032810762524605, + 0.017241302877664566, + 0.014371183700859547, + 0.017901482060551643, + 0.02004283480346203, + 0.010624472051858902, + 0.015229567885398865, + 0.02115998975932598, + 0.02170790173113346, + 0.05940476432442665, + 0.012760993093252182, + 0.08029337227344513, + 0.011765610426664352, + 0.016798976808786392, + 0.00882840808480978, + 0.008194557391107082, + 0.007303966674953699, + 0.008290916681289673, + 0.00499432347714901, + 0.006559387780725956, + 0.004226448480039835, + 0.006691068410873413, + 0.006320448126643896, + 0.004806445445865393, + 0.005013341084122658, + 0.004488654900342226, + 0.0038232801016420126, + 0.007820275612175465, + 0.003106666961684823, + 0.004375969059765339, + 0.006325538270175457, + 0.004986132960766554, + 0.0027271327562630177, + 0.022867850959300995, + 0.006114494986832142, + 0.03454003483057022, + 0.008972546085715294, + 0.009775146842002869, + 0.011142298579216003, + 0.007156678941100836, + 0.00857512280344963, + 0.004013868048787117, + 0.00571879418566823, + 0.009543451480567455, + 0.003915769048035145, + 0.004920148756355047, + 0.003720971755683422, + 0.004688178189098835, + 0.00530692283064127, + 0.007658163085579872, + 0.005074778106063604, + 0.011040030047297478, + 0.005822622682899237, + 0.008430032059550285, + 0.012896462343633175, + 0.006755461450666189, + 0.008405108936131, + 0.018915941938757896, + 0.01325863879173994, + 0.02647356688976288, + 0.027079615741968155, + 0.0212690532207489, + 0.025061799213290215, + 0.025262806564569473, + 0.02403366193175316, + 0.005583911202847958, + 0.006974379997700453, + 0.005102143157273531, + 0.010958829894661903, + 0.004205190110951662, + 0.004616593476384878, + 0.030328886583447456, + 0.023228483274579048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.20156507194042206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291573196649551 + ], + "opposite_logprob_tokens": [ + -2.2379157543182373 + ], + "logprob_reward": -0.11291573196649551, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291573196649551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019809018820524216, + 0.061687156558036804, + 0.6138721704483032, + 0.02217944525182247, + 0.03546339273452759, + 0.014903674833476543, + 0.016586104407906532, + 0.020901475101709366, + 0.044334299862384796, + 0.15026332437992096 + ], + "credit_map": [ + 0.002825387753546238, + 0.001860669581219554, + 0.00272175669670105, + 0.002049343893304467, + 0.0018711521988734603, + 0.0026586754247546196, + 0.002755505498498678, + 0.003066526958718896, + 0.010089773684740067, + 0.006709292996674776, + 0.008500630035996437, + 0.008728571236133575, + 0.004362244158983231, + 0.006876264698803425, + 0.007956192828714848, + 0.008464179001748562, + 0.17655539512634277, + 0.03223240002989769, + 0.2637348175048828, + 0.034625664353370667, + 0.05261847749352455, + 0.02223948948085308, + 0.01694626919925213, + 0.014919595792889595, + 0.0036115774419158697, + 0.0019523798255249858, + 0.004711918532848358, + 0.0017301049083471298, + 0.0028003123588860035, + 0.002673793351277709, + 0.0018996803555637598, + 0.002799674868583679, + 0.006546199321746826, + 0.003617415437474847, + 0.009567885659635067, + 0.0024641193449497223, + 0.0030817987862974405, + 0.004566655494272709, + 0.003387527074664831, + 0.0022317864932119846, + 0.0024563490878790617, + 0.0007932224543765187, + 0.0037531692069023848, + 0.0015316405333578587, + 0.0016173823969438672, + 0.0014586213510483503, + 0.0012836320092901587, + 0.0020096562802791595, + 0.0014890767633914948, + 0.0012162360362708569, + 0.002689062850549817, + 0.0019937879405915737, + 0.0024210927076637745, + 0.0014975880039855838, + 0.002732010092586279, + 0.0025472494307905436, + 0.002190903527662158, + 0.0019494352163746953, + 0.00485737482085824, + 0.0020323803182691336, + 0.0019586016424000263, + 0.0015890111681073904, + 0.002553921891376376, + 0.003769845701754093, + 0.0026772399432957172, + 0.002554775681346655, + 0.008193605579435825, + 0.008354933001101017, + 0.003787706606090069, + 0.004546596668660641, + 0.007692991755902767, + 0.006526446435600519, + 0.007568043190985918, + 0.008161651901900768, + 0.007965144701302052, + 0.018043583258986473, + 0.005178048741072416, + 0.007876710966229439, + 0.05126738175749779, + 0.04420274496078491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.11291573196649551 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01903144083917141, + 0.009218460880219936, + 0.866420567035675, + 0.012075144797563553, + 0.013311395421624184, + 0.007922659628093243, + 0.008536692708730698, + 0.025426149368286133, + 0.017454318702220917, + 0.020603101700544357 + ], + "credit_map": [ + 0.004646704997867346, + 0.002591898664832115, + 0.00277829822152853, + 0.001813399256207049, + 0.001823905622586608, + 0.0018428657203912735, + 0.0016045228112488985, + 0.0019298456609249115, + 0.0016055903397500515, + 0.0014835626352578402, + 0.0012054693652316928, + 0.0008198777795769274, + 0.0011233361437916756, + 0.001008450984954834, + 0.000869276758749038, + 0.0011028967564925551, + 0.2485964149236679, + 0.04581171274185181, + 0.3727417588233948, + 0.04819246381521225, + 0.07405656576156616, + 0.03123524598777294, + 0.024109287187457085, + 0.02167714759707451, + 0.0024498419370502234, + 0.0015514194965362549, + 0.0019123087404295802, + 0.0009435795363970101, + 0.0016569356666877866, + 0.0011982939904555678, + 0.0012861889554187655, + 0.0010765754850581288, + 0.0013569092843681574, + 0.0011270279064774513, + 0.0018253898015245795, + 0.0021325701382011175, + 0.0015313110779970884, + 0.0016809930093586445, + 0.0019099831115454435, + 0.0017472105100750923, + 0.0011333208531141281, + 0.0006034765974618495, + 0.001737283542752266, + 0.0006445838371291757, + 0.0008901968831196427, + 0.0008950854535214603, + 0.0008660873281769454, + 0.001152624492533505, + 0.000932015129365027, + 0.0009016538970172405, + 0.0020282375626266003, + 0.0006696770433336496, + 0.0010811082320287824, + 0.0007556196069344878, + 0.0007387947989627719, + 0.0014295859728008509, + 0.00344285205937922, + 0.0030123491305857897, + 0.0030900740530341864, + 0.0023512751795351505, + 0.003567658830434084, + 0.0041191536001861095, + 0.002866484923288226, + 0.002976302057504654, + 0.0015155996661633253, + 0.0010352551471441984, + 0.003021492389962077, + 0.002390014473348856, + 0.001344954827800393, + 0.0032190559431910515, + 0.002135074231773615, + 0.0027928717900067568, + 0.0011637466959655285, + 0.0015212609432637691, + 0.0008554660016670823, + 0.002050278475508094, + 0.0015267257113009691, + 0.0015402764547616243, + 0.006425831001251936, + 0.0055195181630551815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.014195887371897697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, eye-level perspective within a moving vehicle, the scene unfolds within the harsh, concrete confines of the long tunnel, creating a focused and immersive point-of-view shot. A fiery red sports car, with a low-slung, aerodynamic body and a gleaming, high-gloss finish, initially occupies the very center of the frame. Overhead fluorescent lights pulse rhythmically across its polished surface, accentuating its sleek silhouette. To the right of the screen, a dark green sedan travels in the adjacent lane, slightly ahead of the camera vehicle. Without warning, the red sports car accelerates swiftly and decisively; its dual taillights flash brightly for a moment before surging forward, its powerful body rapidly pulling away down the straight tunnel. As the distance increases, the concrete walls and ceiling blur into gray streaks, but the vibrant red car remains the undeniable focal point, creating a palpable atmosphere of speed, power, and exciting forward momentum.", + "seed": 3607, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_1748_seed_3607_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": -0.0864281749818474, + "vlm_logprob": -0.0864281749818474, + "vlm_yes_no_margin": 3.524999976158142, + "vlm_token_credit": -0.0864281749818474, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17448315024375916 + ], + "opposite_logprob_tokens": [ + -2.049483060836792 + ], + "logprob_reward": -0.17448315024375916, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.17448315024375916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12957222759723663, + 0.12132083624601364, + 0.09435707330703735, + 0.21596822142601013, + 0.08878684043884277, + 0.06187785044312477, + 0.03775373846292496, + 0.10566862672567368, + 0.10774502158164978, + 0.03694958612322807 + ], + "credit_map": [ + 0.018667273223400116, + 0.019063079729676247, + 0.017480997368693352, + 0.012640196830034256, + 0.012610949575901031, + 0.025649303570389748, + 0.01105048879981041, + 0.012409946881234646, + 0.019478902220726013, + 0.016141468659043312, + 0.018503040075302124, + 0.012450569309294224, + 0.00767475226894021, + 0.02161334455013275, + 0.012398426420986652, + 0.013060341589152813, + 0.014301757328212261, + 0.016091398894786835, + 0.011142369359731674, + 0.00794981699436903, + 0.009045870043337345, + 0.01903422363102436, + 0.007214581593871117, + 0.00957705732434988, + 0.03517332300543785, + 0.028903957456350327, + 0.036069173365831375, + 0.017546871677041054, + 0.015836196020245552, + 0.03994610160589218, + 0.015464533120393753, + 0.02702806331217289, + 0.010193989612162113, + 0.009728718549013138, + 0.011185163632035255, + 0.012150744907557964, + 0.010055352002382278, + 0.008564126677811146, + 0.009403431788086891, + 0.01750531792640686, + 0.01244923286139965, + 0.012564880773425102, + 0.0066298493184149265, + 0.005006639752537012, + 0.009495547041296959, + 0.008209816180169582, + 0.003092919010668993, + 0.004428967367857695, + 0.006190720945596695, + 0.006375879514962435, + 0.0054695154540240765, + 0.0044761705212295055, + 0.004489375744014978, + 0.003240803722292185, + 0.00338547071442008, + 0.004125799983739853, + 0.01070699654519558, + 0.015776606276631355, + 0.011587974615395069, + 0.016793055459856987, + 0.010351350530982018, + 0.010526628233492374, + 0.021134693175554276, + 0.008791324682533741, + 0.01204141415655613, + 0.014390215277671814, + 0.017096489667892456, + 0.016853949055075645, + 0.008027298375964165, + 0.01223695557564497, + 0.018510254099965096, + 0.008588448166847229, + 0.004948630463331938, + 0.004767721053212881, + 0.00600065104663372, + 0.005423549097031355, + 0.0021788249723613262, + 0.0026872276794165373, + 0.006232761312276125, + 0.004710222594439983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.17448315024375916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028401773888617754 + ], + "opposite_logprob_tokens": [ + -5.877840042114258 + ], + "logprob_reward": -0.0028401773888617754, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028401773888617754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2409142404794693, + 0.0700441226363182, + 0.08533255010843277, + 0.10454566776752472, + 0.12103115022182465, + 0.05413620173931122, + 0.09461379796266556, + 0.06618121266365051, + 0.08244694024324417, + 0.08075406402349472 + ], + "credit_map": [ + 0.03897472470998764, + 0.05530857294797897, + 0.026522649452090263, + 0.017803426831960678, + 0.021932806819677353, + 0.0428675077855587, + 0.01723909005522728, + 0.020265495404601097, + 0.008187390863895416, + 0.008909640833735466, + 0.012150832451879978, + 0.008980793878436089, + 0.004451665561646223, + 0.011429507285356522, + 0.008601943030953407, + 0.0073323543183505535, + 0.013150136917829514, + 0.012322921305894852, + 0.012643633410334587, + 0.008762836456298828, + 0.006614492274820805, + 0.01668313518166542, + 0.006092809606343508, + 0.009062587283551693, + 0.01570511981844902, + 0.012520280666649342, + 0.01934872753918171, + 0.010220362804830074, + 0.007429002318531275, + 0.018479755148291588, + 0.008584773167967796, + 0.012257653288543224, + 0.014091841876506805, + 0.014177614822983742, + 0.015932375565171242, + 0.01640239544212818, + 0.011796566657721996, + 0.011107473634183407, + 0.013091125525534153, + 0.024431763216853142, + 0.009249602444469929, + 0.013118904083967209, + 0.005554916337132454, + 0.004916645120829344, + 0.005621943157166243, + 0.008702993392944336, + 0.0031191695015877485, + 0.003852033754810691, + 0.014840850606560707, + 0.012392955832183361, + 0.019278286024928093, + 0.012880652211606503, + 0.009126506745815277, + 0.0072251493111252785, + 0.010421322658658028, + 0.008448079228401184, + 0.008449857123196125, + 0.013153084553778172, + 0.007378367707133293, + 0.009024985134601593, + 0.006467781029641628, + 0.007635453250259161, + 0.008130554109811783, + 0.005941134411841631, + 0.010856748558580875, + 0.014804285950958729, + 0.012614713050425053, + 0.009842406958341599, + 0.006237204186618328, + 0.01023130677640438, + 0.010043282061815262, + 0.007817000150680542, + 0.009786785580217838, + 0.01292202714830637, + 0.009503170847892761, + 0.011038457043468952, + 0.00710277259349823, + 0.007804038003087044, + 0.012803994119167328, + 0.009792820550501347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0028401773888617754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22555877268314362 + ], + "opposite_logprob_tokens": [ + -1.6005587577819824 + ], + "logprob_reward": -0.22555877268314362, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555877268314362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05419757217168808, + 0.02020755037665367, + 0.049166541546583176, + 0.04418839141726494, + 0.6651472449302673, + 0.030883340165019035, + 0.019898945465683937, + 0.03777652978897095, + 0.04321646690368652, + 0.03531735762953758 + ], + "credit_map": [ + 0.007949870079755783, + 0.008341430686414242, + 0.007281255442649126, + 0.005100944545120001, + 0.005095540080219507, + 0.011005314067006111, + 0.0048871710896492004, + 0.00453605130314827, + 0.00274482904933393, + 0.002691990230232477, + 0.0023255706764757633, + 0.0027360005769878626, + 0.001698124222457409, + 0.0035891218576580286, + 0.0023302654735744, + 0.0020916510839015245, + 0.006611804012209177, + 0.006033593788743019, + 0.00544604891911149, + 0.0035521886311471462, + 0.006638833787292242, + 0.012614858336746693, + 0.003544075647369027, + 0.004725146573036909, + 0.007522576954215765, + 0.005598296876996756, + 0.00641894806176424, + 0.004666617140173912, + 0.0035567369777709246, + 0.009327549487352371, + 0.00401676632463932, + 0.003080908441916108, + 0.0795128345489502, + 0.07234422862529755, + 0.08362620323896408, + 0.07568489760160446, + 0.08603107184171677, + 0.0657963678240776, + 0.06727334856987, + 0.13487838208675385, + 0.004720781464129686, + 0.006345754023641348, + 0.003904973156750202, + 0.0017208204371854663, + 0.006933154538273811, + 0.003666510572656989, + 0.0017529265023767948, + 0.0018384242430329323, + 0.0034141812939196825, + 0.0030285771936178207, + 0.0031501103658229113, + 0.002026363741606474, + 0.00276456237770617, + 0.0016961977817118168, + 0.0018784411950036883, + 0.0019405137281864882, + 0.004488839302212, + 0.006925028283149004, + 0.0051551503129303455, + 0.004514412954449654, + 0.003218284109607339, + 0.0034601669758558273, + 0.006417149677872658, + 0.0035975032951682806, + 0.005776972509920597, + 0.005789903458207846, + 0.008764895610511303, + 0.005383692216128111, + 0.003909291233867407, + 0.004184363409876823, + 0.0053271041251719, + 0.004080252721905708, + 0.0052795326337218285, + 0.009229335933923721, + 0.004985356237739325, + 0.0036872567143291235, + 0.0023060881067067385, + 0.002648250898346305, + 0.004185216035693884, + 0.0029963236302137375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.22555877268314362 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183089792728424 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183089792728424, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183089792728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06229669600725174, + 0.04617752879858017, + 0.05297376215457916, + 0.04019971564412117, + 0.4424091577529907, + 0.029777249321341515, + 0.03317392244935036, + 0.12669752538204193, + 0.06996140629053116, + 0.09633301943540573 + ], + "credit_map": [ + 0.0063999188132584095, + 0.011311785317957401, + 0.007405519485473633, + 0.006861722096800804, + 0.006583042908459902, + 0.01106240227818489, + 0.005143089685589075, + 0.007529224269092083, + 0.005710400175303221, + 0.0056767696514725685, + 0.007491411175578833, + 0.005802750587463379, + 0.004472686909139156, + 0.007812138181179762, + 0.004859232809394598, + 0.004352146293967962, + 0.008192953653633595, + 0.00649398984387517, + 0.007559963967651129, + 0.005352326203137636, + 0.0043770014308393, + 0.010114169679582119, + 0.00572185916826129, + 0.005161499138921499, + 0.007665031589567661, + 0.004617673344910145, + 0.009113810956478119, + 0.002948804059997201, + 0.002585452049970627, + 0.005877635441720486, + 0.003218483179807663, + 0.004172830376774073, + 0.05294078588485718, + 0.04774588719010353, + 0.05592181906104088, + 0.0510556697845459, + 0.056366924196481705, + 0.04406873136758804, + 0.04467132315039635, + 0.08963804692029953, + 0.004947328474372625, + 0.005057185422629118, + 0.005510935094207525, + 0.0024477236438542604, + 0.0038615551311522722, + 0.0032194918021559715, + 0.0018768360605463386, + 0.0028561954386532307, + 0.005013241898268461, + 0.004236206412315369, + 0.005571633577346802, + 0.004513904452323914, + 0.0032640050631016493, + 0.003225581720471382, + 0.004602738190442324, + 0.0027466153260320425, + 0.014720429666340351, + 0.021569279953837395, + 0.015915246680378914, + 0.015825677663087845, + 0.015940215438604355, + 0.01640552468597889, + 0.01706635393202305, + 0.009254812262952328, + 0.008507966063916683, + 0.012641272507607937, + 0.010159694589674473, + 0.011396795511245728, + 0.004681585356593132, + 0.00680973194539547, + 0.008653511293232441, + 0.007110842503607273, + 0.012707800604403019, + 0.018426887691020966, + 0.016301224008202553, + 0.014302638359367847, + 0.007210488431155682, + 0.005875155795365572, + 0.013763711787760258, + 0.007745115552097559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.018183089792728424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075684800744057 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075684800744057, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075684800744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07925509661436081, + 0.048073865473270416, + 0.062164634466171265, + 0.04772995412349701, + 0.26331475377082825, + 0.038824815303087234, + 0.046334799379110336, + 0.048950593918561935, + 0.052121344953775406, + 0.31323009729385376 + ], + "credit_map": [ + 0.011091380380094051, + 0.01292201317846775, + 0.013414479792118073, + 0.007207818329334259, + 0.007433351594954729, + 0.013165157288312912, + 0.00698147201910615, + 0.007039427291601896, + 0.006267512682825327, + 0.005571404937654734, + 0.007865099236369133, + 0.006351527292281389, + 0.0038334322161972523, + 0.007070397492498159, + 0.004578533116728067, + 0.006535956170409918, + 0.012315432541072369, + 0.008838898502290249, + 0.009435668587684631, + 0.005567569751292467, + 0.004967754241079092, + 0.01017009187489748, + 0.004457251634448767, + 0.0064119682647287846, + 0.007295956369489431, + 0.006444433704018593, + 0.008443589322268963, + 0.004406759515404701, + 0.003576785558834672, + 0.008379035629332066, + 0.003583748359233141, + 0.005599644500762224, + 0.030699755996465683, + 0.027736827731132507, + 0.03388071432709694, + 0.029762262478470802, + 0.032690491527318954, + 0.02617720700800419, + 0.0272563137114048, + 0.05511118099093437, + 0.0054720062762498856, + 0.00804221536964178, + 0.006597741972655058, + 0.003113955957815051, + 0.005288330372422934, + 0.005131852347403765, + 0.002195225562900305, + 0.0029834886081516743, + 0.009971695952117443, + 0.008397683501243591, + 0.005528723821043968, + 0.003688627388328314, + 0.005579333286732435, + 0.006326225586235523, + 0.0032423194497823715, + 0.0036001899279654026, + 0.00597622012719512, + 0.008394839242100716, + 0.006471221335232258, + 0.006053668446838856, + 0.004861578345298767, + 0.005305481143295765, + 0.006817089393734932, + 0.005070493556559086, + 0.005449495278298855, + 0.007677793502807617, + 0.0071721868589520454, + 0.008560987189412117, + 0.003805877873674035, + 0.004363084211945534, + 0.010628808289766312, + 0.0044631133787333965, + 0.05048082768917084, + 0.0764528289437294, + 0.05567961931228638, + 0.036366865038871765, + 0.020844271406531334, + 0.020913736894726753, + 0.02536884695291519, + 0.027123095467686653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.011075684800744057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a static, eye-level perspective within a moving vehicle, the scene unfolds within the harsh, concrete confines of the long tunnel, creating a focused and immersive point-of-view shot. A fiery red sports car, with a low-slung, aerodynamic body and a gleaming, high-gloss finish, initially occupies the very center of the frame. Overhead fluorescent lights pulse rhythmically across its polished surface, accentuating its sleek silhouette. To the right of the screen, a dark green sedan travels in the adjacent lane, slightly ahead of the camera vehicle. Without warning, the red sports car accelerates swiftly and decisively; its dual taillights flash brightly for a moment before surging forward, its powerful body rapidly pulling away down the straight tunnel. As the distance increases, the concrete walls and ceiling blur into gray streaks, but the vibrant red car remains the undeniable focal point, creating a palpable atmosphere of speed, power, and exciting forward momentum.", + "seed": 3608, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_3/index_1748_seed_3608_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": -0.13039570678956808, + "vlm_logprob": -0.13039570678956808, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.13039570678956808, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1418484002351761 + ], + "opposite_logprob_tokens": [ + -2.39184832572937 + ], + "logprob_reward": -0.1418484002351761, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1418484002351761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0841735452413559, + 0.034879665821790695, + 0.029340682551264763, + 0.023637184873223305, + 0.10518311709165573, + 0.07873383909463882, + 0.04342477023601532, + 0.42331594228744507, + 0.038397178053855896, + 0.13891412317752838 + ], + "credit_map": [ + 0.007868414744734764, + 0.016851257532835007, + 0.010570192709565163, + 0.008636360056698322, + 0.006583999842405319, + 0.012715721502900124, + 0.005350784864276648, + 0.015596804209053516, + 0.0038138553500175476, + 0.003596693743020296, + 0.005577148869633675, + 0.006056664045900106, + 0.004468661732971668, + 0.004288125783205032, + 0.0043037584982812405, + 0.0027747564017772675, + 0.0032616944517940283, + 0.004241776652634144, + 0.0043366169556975365, + 0.0036303966771811247, + 0.0024728765711188316, + 0.003710478078573942, + 0.0042833914048969746, + 0.0034034522250294685, + 0.0035452479496598244, + 0.0031044792849570513, + 0.003725118236616254, + 0.002522464608773589, + 0.0018607859965413809, + 0.0038724259939044714, + 0.0022471616975963116, + 0.0027595001738518476, + 0.013730049133300781, + 0.01516452431678772, + 0.01575261913239956, + 0.007601233199238777, + 0.01097080111503601, + 0.019785260781645775, + 0.007748388685286045, + 0.014430241659283638, + 0.010062447749078274, + 0.007045488338917494, + 0.013103987090289593, + 0.008075379766523838, + 0.009396539069712162, + 0.01152107585221529, + 0.009872985072433949, + 0.009655929170548916, + 0.004976106341928244, + 0.004695176146924496, + 0.005285481456667185, + 0.005237144883722067, + 0.0074739535339176655, + 0.006160701159387827, + 0.004110106732696295, + 0.0054860967211425304, + 0.041941575706005096, + 0.0800078734755516, + 0.03405119851231575, + 0.06493458896875381, + 0.05874130502343178, + 0.04608338326215744, + 0.04616375267505646, + 0.05139228329062462, + 0.0048804255202412605, + 0.0030322845559567213, + 0.006106385495513678, + 0.0023600130807608366, + 0.0035860578063875437, + 0.009861358441412449, + 0.0027232442516833544, + 0.005847410764545202, + 0.016393126919865608, + 0.007640127092599869, + 0.043934166431427, + 0.006252645514905453, + 0.005466227885335684, + 0.010245723649859428, + 0.004677945282310247, + 0.04430415481328964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.1418484002351761 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002217455068603158 + ], + "opposite_logprob_tokens": [ + -6.1272172927856445 + ], + "logprob_reward": -0.002217455068603158, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002217455068603158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4156605899333954, + 0.09905540943145752, + 0.05473588407039642, + 0.028500426560640335, + 0.0928225889801979, + 0.04416217282414436, + 0.06020396202802658, + 0.05064482241868973, + 0.07048410922288895, + 0.08373008668422699 + ], + "credit_map": [ + 0.03880387544631958, + 0.11131294071674347, + 0.0351574532687664, + 0.03981608524918556, + 0.02483079396188259, + 0.05099567770957947, + 0.024984782561659813, + 0.08975895494222641, + 0.011265958659350872, + 0.013657628558576107, + 0.02013316936790943, + 0.008710348047316074, + 0.01377271581441164, + 0.016367696225643158, + 0.006866691168397665, + 0.008281206712126732, + 0.006679529789835215, + 0.007580314762890339, + 0.010565796867012978, + 0.004042612388730049, + 0.007033208385109901, + 0.008270286954939365, + 0.004213821142911911, + 0.0063503095880150795, + 0.0025184231344610453, + 0.007816171273589134, + 0.0027505368925631046, + 0.0031782707665115595, + 0.0019704746082425117, + 0.002653597854077816, + 0.0020196938421577215, + 0.005593257024884224, + 0.008349420502781868, + 0.010376437567174435, + 0.01434410735964775, + 0.019133735448122025, + 0.009155274368822575, + 0.007481405045837164, + 0.013930195942521095, + 0.010052011348307133, + 0.007614739239215851, + 0.004100431688129902, + 0.00587258767336607, + 0.00512028718367219, + 0.0051720584742724895, + 0.006656181998550892, + 0.004555850755423307, + 0.005070033948868513, + 0.009647386148571968, + 0.006214101333171129, + 0.007061780896037817, + 0.009285124018788338, + 0.004781344905495644, + 0.008798835799098015, + 0.007314349990338087, + 0.007101040333509445, + 0.004895811900496483, + 0.008629731833934784, + 0.00484370905905962, + 0.008250289596617222, + 0.006357522215694189, + 0.005142701789736748, + 0.006414846051484346, + 0.0061102076433598995, + 0.006279197055846453, + 0.007328443229198456, + 0.010356004349887371, + 0.004986904561519623, + 0.00927028339356184, + 0.009225109592080116, + 0.004276363644748926, + 0.018761806190013885, + 0.008442431688308716, + 0.006122329737991095, + 0.022225026041269302, + 0.005874296650290489, + 0.004401309881359339, + 0.00842155423015356, + 0.005823161918669939, + 0.02241997979581356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.002217455068603158 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029522329568863 + ], + "opposite_logprob_tokens": [ + -2.350295305252075 + ], + "logprob_reward": -0.10029522329568863, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029522329568863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1566394567489624, + 0.43108299374580383, + 0.04870748147368431, + 0.021731974557042122, + 0.043430786579847336, + 0.04652209207415581, + 0.04103199765086174, + 0.06390577554702759, + 0.033301834017038345, + 0.11364559084177017 + ], + "credit_map": [ + 0.019379960373044014, + 0.033638980239629745, + 0.02266157977283001, + 0.014731819741427898, + 0.010990685783326626, + 0.026163537055253983, + 0.008501063100993633, + 0.020571839064359665, + 0.051618289202451706, + 0.07072889804840088, + 0.07929320633411407, + 0.027172338217496872, + 0.07352627068758011, + 0.06900427490472794, + 0.02405722811818123, + 0.03568249195814133, + 0.006644651293754578, + 0.005166338291019201, + 0.008753974922001362, + 0.003468648297712207, + 0.00525936484336853, + 0.0070596556179225445, + 0.00477279257029295, + 0.007582060527056456, + 0.002803815295919776, + 0.0031725438311696053, + 0.003085237229242921, + 0.002273084130138159, + 0.0017742480849847198, + 0.00338054820895195, + 0.0014779495541006327, + 0.0037645490374416113, + 0.004398786928504705, + 0.005103074014186859, + 0.0065372479148209095, + 0.0049363840371370316, + 0.004173520486801863, + 0.007741597481071949, + 0.00400299858301878, + 0.006537180859595537, + 0.007519163656979799, + 0.0045114969834685326, + 0.00642427196726203, + 0.00584695627912879, + 0.005301840137690306, + 0.008485701866447926, + 0.0030572086106985807, + 0.0053754569962620735, + 0.007369539234787226, + 0.004363425076007843, + 0.0059718117117881775, + 0.0036960747092962265, + 0.0033667839597910643, + 0.008019433356821537, + 0.003613300621509552, + 0.004631626419723034, + 0.006937019526958466, + 0.011909417808055878, + 0.005864185281097889, + 0.009253468364477158, + 0.009078974835574627, + 0.007250672206282616, + 0.006223050411790609, + 0.0073889936320483685, + 0.004119417630136013, + 0.002750075189396739, + 0.004703426733613014, + 0.0020019414369016886, + 0.004728095605969429, + 0.005017702002078295, + 0.002499894006177783, + 0.007481286767870188, + 0.01114292535930872, + 0.006415693089365959, + 0.03613226115703583, + 0.005506850779056549, + 0.004247547592967749, + 0.008587177842855453, + 0.0047029959969222546, + 0.03691013902425766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.10029522329568863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870316445827484 + ], + "opposite_logprob_tokens": [ + -1.1370316743850708 + ], + "logprob_reward": -0.3870316445827484, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870316445827484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15495964884757996, + 0.17816884815692902, + 0.09282493591308594, + 0.0549505390226841, + 0.18999533355236053, + 0.06965169310569763, + 0.043635353446006775, + 0.03976757824420929, + 0.05538587644696236, + 0.12066017836332321 + ], + "credit_map": [ + 0.03243516385555267, + 0.028938375413417816, + 0.030960459262132645, + 0.014882332645356655, + 0.011201820336282253, + 0.017529793083667755, + 0.008746436797082424, + 0.010265273042023182, + 0.021993128582835197, + 0.025679124519228935, + 0.05263744667172432, + 0.011728065088391304, + 0.01923319324851036, + 0.020104246214032173, + 0.0103225102648139, + 0.01647114008665085, + 0.012923181988298893, + 0.014769290573894978, + 0.01429706159979105, + 0.008195185102522373, + 0.013025925494730473, + 0.011517840437591076, + 0.008528714999556541, + 0.009567740373313427, + 0.004260933492332697, + 0.01256042905151844, + 0.006284347735345364, + 0.006186292972415686, + 0.0034895597491413355, + 0.006470540072768927, + 0.0045489054173231125, + 0.011149530299007893, + 0.019776251167058945, + 0.02440684847533703, + 0.026269104331731796, + 0.02402832917869091, + 0.016953855752944946, + 0.03258354216814041, + 0.021512340754270554, + 0.02446506917476654, + 0.009395393542945385, + 0.007704957854002714, + 0.009409037418663502, + 0.012016033753752708, + 0.006709987297654152, + 0.008887894451618195, + 0.008011565543711185, + 0.007516823243349791, + 0.005403885617852211, + 0.0047019231133162975, + 0.010206680744886398, + 0.003452910343185067, + 0.003750317031517625, + 0.008476423099637032, + 0.0035113319754600525, + 0.004131878726184368, + 0.004569535609334707, + 0.004052947741001844, + 0.006129037123173475, + 0.00580156035721302, + 0.004370971582829952, + 0.006125834304839373, + 0.0035933416802436113, + 0.005124347750097513, + 0.007159820292145014, + 0.0061244843527674675, + 0.007503076456487179, + 0.004571432713419199, + 0.0046947915107011795, + 0.009955030865967274, + 0.003540130564942956, + 0.011837108992040157, + 0.010536499321460724, + 0.006792240776121616, + 0.039212342351675034, + 0.006847787648439407, + 0.005129024386405945, + 0.011903359554708004, + 0.005285053048282862, + 0.034953873604536057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.3870316445827484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a static, eye-level view from inside a moving vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a fiery red sports car with a low-slung, aerodynamic body and high-gloss finish visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark green sedan visible in the adjacent lane to the right, slightly ahead of the camera vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the red sports car accelerate swiftly, with its dual taillights flashing brightly before pulling away down the tunnel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the concrete walls and ceiling blur into gray streaks as the red car moves away, while the car remains the focal point?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585810765624046 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585810765624046, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585810765624046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11887769401073456, + 0.06196432560682297, + 0.05502822995185852, + 0.0318012610077858, + 0.07715866714715958, + 0.11501482129096985, + 0.03560828045010567, + 0.16952934861183167, + 0.05223646014928818, + 0.28278088569641113 + ], + "credit_map": [ + 0.0183882936835289, + 0.01962638832628727, + 0.029285283759236336, + 0.010696670971810818, + 0.008247626014053822, + 0.015978140756487846, + 0.006893543992191553, + 0.009761753492057323, + 0.006936448160558939, + 0.010361282154917717, + 0.015695776790380478, + 0.0043317521922290325, + 0.008088166825473309, + 0.006963137071579695, + 0.003567763604223728, + 0.006019999273121357, + 0.00901184231042862, + 0.008690854534506798, + 0.008864694274961948, + 0.004418333061039448, + 0.00577765004709363, + 0.00662419106811285, + 0.005262620747089386, + 0.006378039717674255, + 0.0029833046719431877, + 0.007930678315460682, + 0.004005254711955786, + 0.0031690301839262247, + 0.0023023621179163456, + 0.004839615430682898, + 0.0027937516570091248, + 0.0037772669456899166, + 0.008226005360484123, + 0.009224039502441883, + 0.008590872399508953, + 0.006475447677075863, + 0.007864367216825485, + 0.018437892198562622, + 0.006643061991780996, + 0.011696982197463512, + 0.01699312962591648, + 0.010367224924266338, + 0.021577337756752968, + 0.009612935595214367, + 0.017879698425531387, + 0.017193984240293503, + 0.006426020059734583, + 0.01496448740363121, + 0.005338052753359079, + 0.004417010582983494, + 0.004358967766165733, + 0.0030681604985147715, + 0.0038067682180553675, + 0.007683140225708485, + 0.0028855339623987675, + 0.004050645511597395, + 0.017597537487745285, + 0.031574029475450516, + 0.012651551514863968, + 0.02667723223567009, + 0.02348102256655693, + 0.018117234110832214, + 0.019809087738394737, + 0.019621655344963074, + 0.005596209783107042, + 0.006104509346187115, + 0.007710609119385481, + 0.0037841089069843292, + 0.006727016065269709, + 0.00706904474645853, + 0.003433722071349621, + 0.011811242438852787, + 0.049373798072338104, + 0.021128052845597267, + 0.07672441750764847, + 0.01320877205580473, + 0.016758957877755165, + 0.027884282171726227, + 0.012956428341567516, + 0.06474615633487701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.020585810765624046 + } + ], + "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/0000122/group_4/index_tensor([9609, 9609])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_4/index_tensor([9609, 9609])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e92cbedd4af50303ec8a69a89c053fba4112fc26 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_4/index_tensor([9609, 9609])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle, side-on tracking shot, captured in the style of a striking sports animation, follows a group of track and field athletes sprinting along a bright red plastic track. Under the dazzling stadium floodlights, their sweat-soaked skin reflects off the highlights, and the contours of each muscle tensed by exertion are clearly visible. Suddenly, as the camera pans in sync with the athletes, a leading runner, wearing a bright yellow sleeveless sports vest and sporting short, sleek black hair, bursts forth from the tightly packed front group. With a remarkably decisive gesture, he lowers his center of gravity and swings his arms with a faster tempo and greater amplitude. As he accelerates, the entire frame is filled with intense motion blur and countless white speed lines radiating from behind him. The other runners and the stands in the background instantly transform into flowing, abstract blocks of color. The camera quickly follows and zooms in slightly, focusing on his angular profile. He clenches his teeth, beads of sweat dripping from his temple, and his wide eyes burn with unwavering determination, his pupils fixed firmly on the finish line just beyond his sight. This series of action breakdowns vividly depicts the moment he surpasses his limits and leaves his opponent behind, and the entire scene is filled with a breathtaking tension and adrenaline-pumping excitement.", + "seed": 33465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_4/index_9609_seed_33465_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": -0.23505286406725645, + "vlm_logprob": -0.23505286406725645, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.23505286406725645, + "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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01802024431526661 + ], + "opposite_logprob_tokens": [ + -4.143020153045654 + ], + "logprob_reward": -0.01802024431526661, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01802024431526661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13834208250045776, + 0.2724864184856415, + 0.07556827366352081, + 0.2238328456878662, + 0.07247307896614075, + 0.031786687672138214, + 0.01562531292438507, + 0.028529705479741096, + 0.06454945355653763, + 0.07680613547563553 + ], + "credit_map": [ + 0.01083729974925518, + 0.012652605772018433, + 0.01911875605583191, + 0.021953124552965164, + 0.00995829701423645, + 0.012090965174138546, + 0.02541831135749817, + 0.026312725618481636, + 0.024951132014393806, + 0.031079895794391632, + 0.03736897185444832, + 0.026360465213656425, + 0.02544357068836689, + 0.0335332453250885, + 0.04221739247441292, + 0.05153173208236694, + 0.008134501054883003, + 0.007500536274164915, + 0.009272848255932331, + 0.008786679245531559, + 0.01098354160785675, + 0.006265485193580389, + 0.01544228196144104, + 0.009182398207485676, + 0.013061215169727802, + 0.015317689627408981, + 0.010736105032265186, + 0.015559699386358261, + 0.01965874433517456, + 0.057057615369558334, + 0.02950168214738369, + 0.0629400908946991, + 0.008337447419762611, + 0.012320855632424355, + 0.0064853886142373085, + 0.007402827963232994, + 0.011791565455496311, + 0.009195569902658463, + 0.008189130574464798, + 0.00875029806047678, + 0.003917091526091099, + 0.0027484260499477386, + 0.0038411342538893223, + 0.004859855864197016, + 0.004273105878382921, + 0.004284408874809742, + 0.0030816742219030857, + 0.004780991468578577, + 0.0015439046546816826, + 0.0014475779607892036, + 0.001444554072804749, + 0.00200909236446023, + 0.002028638729825616, + 0.001993637066334486, + 0.0014479871606454253, + 0.003709921846166253, + 0.004360819701105356, + 0.0054482887499034405, + 0.0026306230574846268, + 0.0016202887054532766, + 0.0057403817772865295, + 0.003838097443804145, + 0.002163691446185112, + 0.002727513201534748, + 0.0062206401489675045, + 0.00605051452293992, + 0.005075679160654545, + 0.004186402074992657, + 0.0103611433878541, + 0.017845969647169113, + 0.006012510042637587, + 0.00879659503698349, + 0.02328222431242466, + 0.0052502769976854324, + 0.00529935909435153, + 0.0034145857207477093, + 0.003304774407297373, + 0.006068563554435968, + 0.007416681386530399, + 0.02276967093348503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.01802024431526661 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.1371134519577026 + ], + "opposite_logprob_tokens": [ + -0.387113481760025 + ], + "logprob_reward": -1.1371134519577026, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371134519577026, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1332533359527588, + 0.43866896629333496, + 0.09591217339038849, + 0.12192299962043762, + 0.02457612194120884, + 0.011330539360642433, + 0.014123386703431606, + 0.018697336316108704, + 0.023706762120127678, + 0.11780842393636703 + ], + "credit_map": [ + 0.013139229267835617, + 0.011919519864022732, + 0.01735563576221466, + 0.019649235531687737, + 0.010188237763941288, + 0.014000319875776768, + 0.021481722593307495, + 0.02551943250000477, + 0.03987732529640198, + 0.04455640912055969, + 0.06622330844402313, + 0.04552688077092171, + 0.03780249133706093, + 0.05346280708909035, + 0.07080661505460739, + 0.08041311055421829, + 0.007621983997523785, + 0.006406235042959452, + 0.009638066403567791, + 0.008920108899474144, + 0.009207751601934433, + 0.018960950896143913, + 0.01630893349647522, + 0.018848147243261337, + 0.007206772454082966, + 0.008348428644239902, + 0.006887718103826046, + 0.010607358068227768, + 0.010873583145439625, + 0.028620675206184387, + 0.017456548288464546, + 0.03192191943526268, + 0.003318739589303732, + 0.004270661156624556, + 0.0019793978426605463, + 0.0026969274040311575, + 0.0032775362487882376, + 0.003535112366080284, + 0.0025100125931203365, + 0.002987735206261277, + 0.0015821718843653798, + 0.0009472348028793931, + 0.0015603440115228295, + 0.0015390104381367564, + 0.0012789511820301414, + 0.0013919401681050658, + 0.0014908314915373921, + 0.0015400551492348313, + 0.0013932522851973772, + 0.0013479713816195726, + 0.0011350857093930244, + 0.00135002750903368, + 0.002504923613741994, + 0.0018853021319955587, + 0.0016556890914216638, + 0.002851136028766632, + 0.0024113997351378202, + 0.002151480410248041, + 0.001386774703860283, + 0.001451521529816091, + 0.004927079193294048, + 0.0022710696794092655, + 0.0017304590437561274, + 0.002367553301155567, + 0.0028424591291695833, + 0.0038829315453767776, + 0.002348172478377819, + 0.0018202975625172257, + 0.0037127011455595493, + 0.00400405703112483, + 0.002289483090862632, + 0.0028066597878932953, + 0.02908427268266678, + 0.00855714175850153, + 0.007381741423159838, + 0.010288628749549389, + 0.007555575110018253, + 0.010181128978729248, + 0.017338605597615242, + 0.027421332895755768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -1.1371134519577026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.14268478751182556, + 0.05935239419341087, + 0.05508708953857422, + 0.0989503413438797, + 0.04684659466147423, + 0.037921883165836334, + 0.019120801240205765, + 0.03432023897767067, + 0.059146516025066376, + 0.4465693533420563 + ], + "credit_map": [ + 0.010855484753847122, + 0.012156793847680092, + 0.02318062074482441, + 0.02302299439907074, + 0.009534357115626335, + 0.008763873018324375, + 0.030733363702893257, + 0.02443729154765606, + 0.005557018332183361, + 0.007013959810137749, + 0.0070786974392831326, + 0.0062071792781353, + 0.0055741737596690655, + 0.006772596854716539, + 0.009347084909677505, + 0.011801687069237232, + 0.007658421993255615, + 0.005125293508172035, + 0.00638832151889801, + 0.005592958070337772, + 0.006500269751995802, + 0.004466680344194174, + 0.01136799156665802, + 0.00798715092241764, + 0.0054634506814181805, + 0.006040673237293959, + 0.006648316513746977, + 0.00583989080041647, + 0.008919671177864075, + 0.02359289862215519, + 0.013362452387809753, + 0.02908298559486866, + 0.004445123951882124, + 0.007672117557376623, + 0.00525329727679491, + 0.0062065632082521915, + 0.006379284430295229, + 0.005357543472200632, + 0.005538055207580328, + 0.005994608160108328, + 0.004165120422840118, + 0.0029312947299331427, + 0.00559223722666502, + 0.005195281468331814, + 0.0044895014725625515, + 0.0053200614638626575, + 0.0036914944648742676, + 0.006536893546581268, + 0.002261336427181959, + 0.0020109592005610466, + 0.002582570305094123, + 0.002807391108945012, + 0.002677693497389555, + 0.002491355873644352, + 0.0018997968873009086, + 0.0023896985221654177, + 0.003764088498428464, + 0.0034352766815572977, + 0.0027789019513875246, + 0.002585913985967636, + 0.009544468484818935, + 0.004325411282479763, + 0.002835109131410718, + 0.005051069427281618, + 0.004730208311229944, + 0.008820312097668648, + 0.005712889600545168, + 0.005742840003222227, + 0.01051874179393053, + 0.008810281753540039, + 0.008363014087080956, + 0.0064482251182198524, + 0.16578279435634613, + 0.018749132752418518, + 0.027618195861577988, + 0.009236034005880356, + 0.011276843957602978, + 0.028234869241714478, + 0.022136975079774857, + 0.163534477353096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.003615273628383875 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5216624140739441, + 0.14142782986164093, + 0.0899868980050087, + 0.07035067677497864, + 0.050961386412382126, + 0.02702968381345272, + 0.013843996450304985, + 0.03497808799147606, + 0.018646037206053734, + 0.03111300803720951 + ], + "credit_map": [ + 0.03694663196802139, + 0.023678110912442207, + 0.07953079044818878, + 0.04990370199084282, + 0.028322821483016014, + 0.04173581674695015, + 0.16119152307510376, + 0.10035310685634613, + 0.014500626362860203, + 0.013289234600961208, + 0.021342530846595764, + 0.014487509615719318, + 0.012319367378950119, + 0.01886831782758236, + 0.020102033391594887, + 0.026518210768699646, + 0.007625190075486898, + 0.004998762160539627, + 0.005371381528675556, + 0.005254659336060286, + 0.007321138400584459, + 0.02100764587521553, + 0.014598559588193893, + 0.023809563368558884, + 0.004355384036898613, + 0.0047733779065310955, + 0.005046886391937733, + 0.008091417141258717, + 0.005951362662017345, + 0.01504786778241396, + 0.009751103818416595, + 0.017333272844552994, + 0.005540798418223858, + 0.00627101119607687, + 0.0045094857923686504, + 0.006581067107617855, + 0.009027919732034206, + 0.006314191501587629, + 0.005913993809372187, + 0.006802920717746019, + 0.004358887206763029, + 0.002119677374139428, + 0.0030414003413170576, + 0.0029823624063283205, + 0.004043494816869497, + 0.003910775296390057, + 0.0026344547513872385, + 0.003938630688935518, + 0.0014496390940621495, + 0.0011057244846597314, + 0.0013321149162948132, + 0.0023620035499334335, + 0.002329297596588731, + 0.0015105430502444506, + 0.0012200276833027601, + 0.002534646075218916, + 0.005804416257888079, + 0.007871478796005249, + 0.003105590119957924, + 0.0024816147051751614, + 0.005496592726558447, + 0.00437882449477911, + 0.002000149106606841, + 0.0038394215516746044, + 0.0017397215124219656, + 0.0025709611363708973, + 0.0014343170914798975, + 0.0012767307925969362, + 0.0020924857817590237, + 0.0051253982819616795, + 0.0017344673397019506, + 0.0026719544548541307, + 0.004971323534846306, + 0.002567548304796219, + 0.003169294446706772, + 0.002431761007755995, + 0.0025285305455327034, + 0.005646963138133287, + 0.005231062415987253, + 0.004566526971757412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00673500495031476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07699962705373764, + 0.15206706523895264, + 0.11318166553974152, + 0.09749457985162735, + 0.03759174421429634, + 0.08966678380966187, + 0.07311869412660599, + 0.026355648413300514, + 0.0530468188226223, + 0.2804773151874542 + ], + "credit_map": [ + 0.006988634821027517, + 0.007600638084113598, + 0.010183494538068771, + 0.008196098729968071, + 0.007957163266837597, + 0.006651980336755514, + 0.017437756061553955, + 0.01198386400938034, + 0.013775905594229698, + 0.018995990976691246, + 0.020196583122015, + 0.015676766633987427, + 0.01387980580329895, + 0.019195687025785446, + 0.02142883464694023, + 0.028917476534843445, + 0.01110589224845171, + 0.008883219212293625, + 0.008252217434346676, + 0.007281177211552858, + 0.012063462287187576, + 0.021912183612585068, + 0.0183684304356575, + 0.025315091013908386, + 0.0065904650837183, + 0.00668181199580431, + 0.006816359702497721, + 0.008200653828680515, + 0.009095677174627781, + 0.022079410031437874, + 0.014115830883383751, + 0.023914368823170662, + 0.003618721617385745, + 0.005049462430179119, + 0.002606404945254326, + 0.0048894803039729595, + 0.005715752951800823, + 0.005035299342125654, + 0.004865315277129412, + 0.005811306647956371, + 0.011001943610608578, + 0.00898349191993475, + 0.010223215445876122, + 0.011658797040581703, + 0.014688096009194851, + 0.012401037849485874, + 0.00942106731235981, + 0.011289134621620178, + 0.005722318310290575, + 0.005747171584516764, + 0.008980246260762215, + 0.011165983974933624, + 0.012977145612239838, + 0.008224518969655037, + 0.012024893425405025, + 0.008276418782770634, + 0.004614193458110094, + 0.006206889171153307, + 0.002350050024688244, + 0.002179280389100313, + 0.0032779111061245203, + 0.0029982179403305054, + 0.0016883923672139645, + 0.003040711861103773, + 0.004815276712179184, + 0.005729162599891424, + 0.003173425793647766, + 0.002731176558881998, + 0.006148449145257473, + 0.017365090548992157, + 0.004827036056667566, + 0.008257197216153145, + 0.09321172535419464, + 0.012953824363648891, + 0.01605755090713501, + 0.011672223918139935, + 0.01052815280854702, + 0.02160860225558281, + 0.02358361892402172, + 0.09086163341999054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.009780345484614372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, side-on tracking shot, captured in the style of a striking sports animation, follows a group of track and field athletes sprinting along a bright red plastic track. Under the dazzling stadium floodlights, their sweat-soaked skin reflects off the highlights, and the contours of each muscle tensed by exertion are clearly visible. Suddenly, as the camera pans in sync with the athletes, a leading runner, wearing a bright yellow sleeveless sports vest and sporting short, sleek black hair, bursts forth from the tightly packed front group. With a remarkably decisive gesture, he lowers his center of gravity and swings his arms with a faster tempo and greater amplitude. As he accelerates, the entire frame is filled with intense motion blur and countless white speed lines radiating from behind him. The other runners and the stands in the background instantly transform into flowing, abstract blocks of color. The camera quickly follows and zooms in slightly, focusing on his angular profile. He clenches his teeth, beads of sweat dripping from his temple, and his wide eyes burn with unwavering determination, his pupils fixed firmly on the finish line just beyond his sight. This series of action breakdowns vividly depicts the moment he surpasses his limits and leaves his opponent behind, and the entire scene is filled with a breathtaking tension and adrenaline-pumping excitement.", + "seed": 33466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_4/index_9609_seed_33466_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": -0.462932338565588, + "vlm_logprob": -0.462932338565588, + "vlm_yes_no_margin": 2.799999952316284, + "vlm_token_credit": -0.462932338565588, + "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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03133052960038185 + ], + "opposite_logprob_tokens": [ + -3.5313305854797363 + ], + "logprob_reward": -0.03133052960038185, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03133052960038185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10099027305841446, + 0.031845226883888245, + 0.27192720770835876, + 0.11527454853057861, + 0.08820842206478119, + 0.0629640594124794, + 0.12666930258274078, + 0.06658542156219482, + 0.10568465292453766, + 0.029850836843252182 + ], + "credit_map": [ + 0.012142383493483067, + 0.010485811159014702, + 0.00550253177061677, + 0.006358800921589136, + 0.0055871643126010895, + 0.011560606770217419, + 0.023900602012872696, + 0.025452373549342155, + 0.003068303456529975, + 0.0027161475736647844, + 0.002173482673242688, + 0.0022385476622730494, + 0.002694506663829088, + 0.008353391662240028, + 0.004609726835042238, + 0.0059911212883889675, + 0.05889960005879402, + 0.11977629363536835, + 0.018181130290031433, + 0.012369167059659958, + 0.012883166782557964, + 0.012423647567629814, + 0.021969959139823914, + 0.015424232929944992, + 0.013523824512958527, + 0.019543776288628578, + 0.02196558564901352, + 0.018793396651744843, + 0.007839642465114594, + 0.010114517994225025, + 0.012302032671868801, + 0.01119176670908928, + 0.01034836657345295, + 0.02166489139199257, + 0.026427455246448517, + 0.01055145263671875, + 0.005978017579764128, + 0.005337503273040056, + 0.004942603409290314, + 0.00295813474804163, + 0.007624947931617498, + 0.01172016840428114, + 0.009802297689020634, + 0.006859475746750832, + 0.006802112329751253, + 0.006777482572942972, + 0.006723727099597454, + 0.00665384391322732, + 0.017236685380339622, + 0.013686805963516235, + 0.020959803834557533, + 0.016036411747336388, + 0.011609019711613655, + 0.020285356789827347, + 0.014836353249847889, + 0.012018869630992413, + 0.010281366296112537, + 0.008754469454288483, + 0.00896285567432642, + 0.004817207343876362, + 0.01386096328496933, + 0.008648809045553207, + 0.005089472979307175, + 0.006170275155454874, + 0.01216668076813221, + 0.019432958215475082, + 0.020284578204154968, + 0.008307291194796562, + 0.010895969346165657, + 0.019652754068374634, + 0.005370412487536669, + 0.009574013762176037, + 0.004208149388432503, + 0.004073690623044968, + 0.002119774231687188, + 0.0029503516852855682, + 0.009186425246298313, + 0.0030199335888028145, + 0.002643764717504382, + 0.0016487467801198363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.03133052960038185 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -2.2379608154296875 + ], + "opposite_logprob_tokens": [ + -0.11296088248491287 + ], + "logprob_reward": -2.2379608154296875, + "yes_no_margin_reward": -2.125, + "token_credit_reward": -2.2379608154296875, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14547672867774963, + 0.05365797504782677, + 0.05959411337971687, + 0.08445362746715546, + 0.05460662394762039, + 0.12817977368831635, + 0.13401490449905396, + 0.03138796612620354, + 0.25005534291267395, + 0.05857296288013458 + ], + "credit_map": [ + 0.012073223479092121, + 0.010283542796969414, + 0.008113299496471882, + 0.007641827221959829, + 0.005694519262760878, + 0.014263530261814594, + 0.04350506141781807, + 0.043901726603507996, + 0.0042464942671358585, + 0.003328929888084531, + 0.0033313725143671036, + 0.005177173763513565, + 0.004280630964785814, + 0.011124075390398502, + 0.009659525007009506, + 0.01250977162271738, + 0.006638876162469387, + 0.006531062535941601, + 0.005694359075278044, + 0.00910720881074667, + 0.006131531670689583, + 0.00789385475218296, + 0.009906595572829247, + 0.007690626196563244, + 0.011991161853075027, + 0.02073853462934494, + 0.012776084244251251, + 0.008052859455347061, + 0.0073673236183822155, + 0.011167028918862343, + 0.0067825294099748135, + 0.005578110925853252, + 0.0075576333329081535, + 0.01306308712810278, + 0.009329644031822681, + 0.007334236521273851, + 0.004288765601813793, + 0.0065025584772229195, + 0.003139774315059185, + 0.003390925470739603, + 0.02072281576693058, + 0.0234790351241827, + 0.025753632187843323, + 0.010280294343829155, + 0.009870896115899086, + 0.012094742618501186, + 0.007773876190185547, + 0.018204476684331894, + 0.019201043993234634, + 0.017189856618642807, + 0.02080918848514557, + 0.011488370597362518, + 0.013562939129769802, + 0.019761402159929276, + 0.014678211882710457, + 0.017323896288871765, + 0.0036308499984443188, + 0.004267805255949497, + 0.004756290931254625, + 0.0035978350788354874, + 0.004358666483312845, + 0.003812875133007765, + 0.0034523294307291508, + 0.003511313581839204, + 0.0295166727155447, + 0.048820458352565765, + 0.03348326310515404, + 0.021458197385072708, + 0.035087764263153076, + 0.03463010862469673, + 0.017999473959207535, + 0.0290593970566988, + 0.007760443724691868, + 0.009206087328493595, + 0.005609625484794378, + 0.005222331266850233, + 0.016586370766162872, + 0.007477455772459507, + 0.003488016314804554, + 0.0032226284965872765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -2.2379608154296875 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089242994785309 + ], + "opposite_logprob_tokens": [ + -4.511089324951172 + ], + "logprob_reward": -0.011089242994785309, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089242994785309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1102319210767746, + 0.02739766240119934, + 0.06525643169879913, + 0.13892732560634613, + 0.21176530420780182, + 0.09177910536527634, + 0.12172000110149384, + 0.029199033975601196, + 0.1300347000360489, + 0.0736885592341423 + ], + "credit_map": [ + 0.011609387584030628, + 0.009831559844315052, + 0.007640717551112175, + 0.006992573384195566, + 0.004635289777070284, + 0.012464706785976887, + 0.026074085384607315, + 0.03098359890282154, + 0.0024777387734502554, + 0.0021260143257677555, + 0.0014931675978004932, + 0.002592978300526738, + 0.002158370101824403, + 0.005438352935016155, + 0.004873971454799175, + 0.006237069610506296, + 0.006761654745787382, + 0.0074755242094397545, + 0.006257443688809872, + 0.009681034833192825, + 0.0065012117847800255, + 0.007952218875288963, + 0.011393291875720024, + 0.00923404935747385, + 0.021654661744832993, + 0.04125908389687538, + 0.026708398014307022, + 0.012717925012111664, + 0.008660377003252506, + 0.01135533768683672, + 0.009016860276460648, + 0.0075546749867498875, + 0.062060944736003876, + 0.08002152293920517, + 0.017168430611491203, + 0.011731072328984737, + 0.012556973844766617, + 0.014863469637930393, + 0.005119886249303818, + 0.008243001997470856, + 0.01052266638725996, + 0.014306027442216873, + 0.016709819436073303, + 0.009089727886021137, + 0.013186294585466385, + 0.008854866027832031, + 0.01062674168497324, + 0.008482960052788258, + 0.014360873959958553, + 0.013124276883900166, + 0.021633213385939598, + 0.011641447432339191, + 0.011165736243128777, + 0.01602178066968918, + 0.016456060111522675, + 0.017316613346338272, + 0.003193364944308996, + 0.00304817664436996, + 0.005011960864067078, + 0.003035027300938964, + 0.00487965252250433, + 0.004321293439716101, + 0.0025769262574613094, + 0.0031326322350651026, + 0.015636570751667023, + 0.0173173937946558, + 0.02091578207910061, + 0.014646977186203003, + 0.016566108912229538, + 0.017335347831249237, + 0.010393721982836723, + 0.017222793772816658, + 0.009369480423629284, + 0.010298271663486958, + 0.006051588337868452, + 0.006211248692125082, + 0.02309347875416279, + 0.010119526647031307, + 0.004175086040049791, + 0.004369872622191906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.011089242994785309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082914546132088 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082914546132088, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082914546132088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2743552625179291, + 0.03596444055438042, + 0.0550338551402092, + 0.10259996354579926, + 0.06460979580879211, + 0.09880103170871735, + 0.12174306064844131, + 0.04935811460018158, + 0.13978290557861328, + 0.05775148794054985 + ], + "credit_map": [ + 0.017111193388700485, + 0.017385736107826233, + 0.013577647507190704, + 0.019315529614686966, + 0.010146718472242355, + 0.019945083186030388, + 0.09092723578214645, + 0.08594613522291183, + 0.0031915921717882156, + 0.0025698954705148935, + 0.0018003284931182861, + 0.0037852854002267122, + 0.00296428008005023, + 0.0064474535174667835, + 0.006714874412864447, + 0.008490737527608871, + 0.007642227690666914, + 0.009933196008205414, + 0.004415655508637428, + 0.007130853366106749, + 0.005174970719963312, + 0.00638763140887022, + 0.008322465233504772, + 0.006026856601238251, + 0.013134567067027092, + 0.027172556146979332, + 0.02094150334596634, + 0.009964216500520706, + 0.00727425329387188, + 0.012156208045780659, + 0.0063336496241390705, + 0.005623016972094774, + 0.01134449988603592, + 0.01602398417890072, + 0.009010236710309982, + 0.007328879088163376, + 0.00483004841953516, + 0.007692087907344103, + 0.00385995558463037, + 0.004520110785961151, + 0.017749689519405365, + 0.020218949764966965, + 0.021659282967448235, + 0.007905176840722561, + 0.010156484320759773, + 0.008705665357410908, + 0.00616047065705061, + 0.006245325319468975, + 0.02009483613073826, + 0.014961524866521358, + 0.0170730073004961, + 0.011102001182734966, + 0.012706512585282326, + 0.019218195229768753, + 0.013751952908933163, + 0.012835037894546986, + 0.006045869551599026, + 0.00760033680126071, + 0.008540594018995762, + 0.004744080360978842, + 0.006304455921053886, + 0.007002349942922592, + 0.0045249019749462605, + 0.004595531616359949, + 0.017523685470223427, + 0.01696871407330036, + 0.021801842376589775, + 0.014833571389317513, + 0.020478632301092148, + 0.018191451206803322, + 0.012407276779413223, + 0.017577748745679855, + 0.007239502388983965, + 0.008664751425385475, + 0.0060574873350560665, + 0.005195453763008118, + 0.014881577342748642, + 0.00884456466883421, + 0.0036036663223057985, + 0.0032644865568727255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.016082914546132088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12493883818387985, + 0.021340640261769295, + 0.08273537456989288, + 0.04634673148393631, + 0.17022380232810974, + 0.13252076506614685, + 0.13299153745174408, + 0.04708695039153099, + 0.1867586076259613, + 0.055056799203157425 + ], + "credit_map": [ + 0.012642794288694859, + 0.009129470214247704, + 0.009907256811857224, + 0.008643105626106262, + 0.005222858861088753, + 0.013351053930819035, + 0.03174259513616562, + 0.034299690276384354, + 0.002207776065915823, + 0.0015981151955202222, + 0.0012447155313566327, + 0.002044991124421358, + 0.00158600730355829, + 0.0036668162792921066, + 0.0038944941479712725, + 0.005097721703350544, + 0.01228244137018919, + 0.015623900108039379, + 0.0066962470300495625, + 0.010695790871977806, + 0.007502398453652859, + 0.008293629623949528, + 0.012402239255607128, + 0.009238709695637226, + 0.006445098668336868, + 0.010724332183599472, + 0.007374177221208811, + 0.005661193747073412, + 0.002837246982380748, + 0.0054627759382128716, + 0.004331101197749376, + 0.0035107985604554415, + 0.04992798715829849, + 0.06510934233665466, + 0.014652208425104618, + 0.008812019601464272, + 0.01025015115737915, + 0.0101188188418746, + 0.004584289155900478, + 0.006768963299691677, + 0.024938207119703293, + 0.030399316921830177, + 0.023119406774640083, + 0.010401337407529354, + 0.009640306234359741, + 0.011011751368641853, + 0.007902453653514385, + 0.015107979066669941, + 0.016432208940386772, + 0.015528211370110512, + 0.02471844106912613, + 0.012578041292726994, + 0.011960536241531372, + 0.01644442416727543, + 0.01804334856569767, + 0.017286302521824837, + 0.005388669669628143, + 0.005715983919799328, + 0.006936407182365656, + 0.004617094062268734, + 0.009034676477313042, + 0.006127965170890093, + 0.004217011854052544, + 0.005049135535955429, + 0.025715339928865433, + 0.03503618761897087, + 0.030142562463879585, + 0.013334974646568298, + 0.021933909505605698, + 0.031195497140288353, + 0.011171957477927208, + 0.018228165805339813, + 0.006674087140709162, + 0.007409147918224335, + 0.005653259344398975, + 0.004958407022058964, + 0.013088272884488106, + 0.009828297421336174, + 0.0030542570166289806, + 0.004391063936054707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.01819819025695324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, side-on tracking shot, captured in the style of a striking sports animation, follows a group of track and field athletes sprinting along a bright red plastic track. Under the dazzling stadium floodlights, their sweat-soaked skin reflects off the highlights, and the contours of each muscle tensed by exertion are clearly visible. Suddenly, as the camera pans in sync with the athletes, a leading runner, wearing a bright yellow sleeveless sports vest and sporting short, sleek black hair, bursts forth from the tightly packed front group. With a remarkably decisive gesture, he lowers his center of gravity and swings his arms with a faster tempo and greater amplitude. As he accelerates, the entire frame is filled with intense motion blur and countless white speed lines radiating from behind him. The other runners and the stands in the background instantly transform into flowing, abstract blocks of color. The camera quickly follows and zooms in slightly, focusing on his angular profile. He clenches his teeth, beads of sweat dripping from his temple, and his wide eyes burn with unwavering determination, his pupils fixed firmly on the finish line just beyond his sight. This series of action breakdowns vividly depicts the moment he surpasses his limits and leaves his opponent behind, and the entire scene is filled with a breathtaking tension and adrenaline-pumping excitement.", + "seed": 33467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_4/index_9609_seed_33467_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": -0.19563110433518888, + "vlm_logprob": -0.19563110433518888, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.19563110433518888, + "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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06400232017040253 + ], + "opposite_logprob_tokens": [ + -2.814002275466919 + ], + "logprob_reward": -0.06400232017040253, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06400232017040253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09794168919324875, + 0.05035759508609772, + 0.05077158659696579, + 0.16989628970623016, + 0.01851782016456127, + 0.04119066521525383, + 0.054029129445552826, + 0.044734809547662735, + 0.34208911657333374, + 0.13047130405902863 + ], + "credit_map": [ + 0.01061267964541912, + 0.008755339309573174, + 0.009891219437122345, + 0.007961169816553593, + 0.01276756078004837, + 0.01480933278799057, + 0.016317151486873627, + 0.016827236860990524, + 0.0058904592879116535, + 0.005394126288592815, + 0.006349267438054085, + 0.005107914563268423, + 0.006426636595278978, + 0.006392049603164196, + 0.007620283402502537, + 0.007176858838647604, + 0.004512255545705557, + 0.004481425043195486, + 0.0062834639102220535, + 0.004996321629732847, + 0.004786546807736158, + 0.008895521983504295, + 0.0057807741686701775, + 0.01103527843952179, + 0.017331499606370926, + 0.02504824660718441, + 0.018636424094438553, + 0.021924300119280815, + 0.023694375529885292, + 0.016694406047463417, + 0.025077713653445244, + 0.021489322185516357, + 0.0025764477904886007, + 0.0024752661120146513, + 0.0024880908895283937, + 0.0012756850337609649, + 0.0021211181301623583, + 0.003654185915365815, + 0.0023501708637923002, + 0.0015768558951094747, + 0.007532869931310415, + 0.00603122916072607, + 0.006445298902690411, + 0.00495483772829175, + 0.003865351201966405, + 0.004214110318571329, + 0.004597983788698912, + 0.003548982320353389, + 0.005991965997964144, + 0.005799951031804085, + 0.0072311717085540295, + 0.006827253382652998, + 0.009878075681626797, + 0.005673102103173733, + 0.0063359844498336315, + 0.006291626486927271, + 0.009235798381268978, + 0.00749686174094677, + 0.0070644281804561615, + 0.0031026690267026424, + 0.004372339230030775, + 0.00477979052811861, + 0.00372930895537138, + 0.004953613970428705, + 0.11040664464235306, + 0.06924349069595337, + 0.02072341926395893, + 0.019354550167918205, + 0.031132543459534645, + 0.04488926753401756, + 0.023272976279258728, + 0.023066220805048943, + 0.010803448036313057, + 0.018200213089585304, + 0.014514843001961708, + 0.020005468279123306, + 0.010236863978207111, + 0.015877651050686836, + 0.02613644301891327, + 0.014696372672915459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.06400232017040253 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983632922172546 + ], + "opposite_logprob_tokens": [ + -0.5233632922172546 + ], + "logprob_reward": -0.8983632922172546, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983632922172546, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08955052495002747, + 0.03020480088889599, + 0.04857739806175232, + 0.05128103494644165, + 0.029377443715929985, + 0.03694319725036621, + 0.07439029961824417, + 0.23380334675312042, + 0.3182629942893982, + 0.08760897815227509 + ], + "credit_map": [ + 0.007586311548948288, + 0.006455502472817898, + 0.008938612416386604, + 0.0078573664650321, + 0.013065903447568417, + 0.012596085667610168, + 0.018504343926906586, + 0.014546399004757404, + 0.0027669742703437805, + 0.0023405635729432106, + 0.0028813460376113653, + 0.004107591696083546, + 0.004198000766336918, + 0.004208835307508707, + 0.005240302998572588, + 0.00446118600666523, + 0.004426993429660797, + 0.003164427587762475, + 0.004416057374328375, + 0.004191319923847914, + 0.0047806487418711185, + 0.009281294420361519, + 0.005612161010503769, + 0.012704497203230858, + 0.008077665232121944, + 0.005843097344040871, + 0.008319852873682976, + 0.006598478648811579, + 0.005279892589896917, + 0.005635347682982683, + 0.005151782184839249, + 0.0063749197870492935, + 0.0050331405363976955, + 0.005659053102135658, + 0.0036195404827594757, + 0.0018734504701569676, + 0.002548712771385908, + 0.003990596625953913, + 0.004287583753466606, + 0.0023653642274439335, + 0.004789401777088642, + 0.004946828354150057, + 0.0064700087532401085, + 0.004077321384102106, + 0.005257055163383484, + 0.0038350156974047422, + 0.0042236242443323135, + 0.003343940479680896, + 0.012325333431363106, + 0.01343119889497757, + 0.011257986538112164, + 0.004047672729939222, + 0.007917914539575577, + 0.008500663563609123, + 0.004487421363592148, + 0.012422117404639721, + 0.052706435322761536, + 0.04751821607351303, + 0.03776809200644493, + 0.014868644066154957, + 0.01901349611580372, + 0.019520169124007225, + 0.014772675931453705, + 0.027635615319013596, + 0.10287173092365265, + 0.0635664090514183, + 0.018937386572360992, + 0.017061466351151466, + 0.027072783559560776, + 0.04200806841254234, + 0.02192601002752781, + 0.02481912449002266, + 0.008461393415927887, + 0.013621016405522823, + 0.008742987178266048, + 0.009919507429003716, + 0.009373864158987999, + 0.01359848864376545, + 0.011794901452958584, + 0.01209682784974575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.8983632922172546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07267215847969055, + 0.038212355226278305, + 0.04345962405204773, + 0.07431179285049438, + 0.027854297310113907, + 0.06629674136638641, + 0.03924084082245827, + 0.1319996416568756, + 0.30816078186035156, + 0.19779179990291595 + ], + "credit_map": [ + 0.008109376765787601, + 0.0079898526892066, + 0.007432861253619194, + 0.006503621116280556, + 0.011653082445263863, + 0.01169927790760994, + 0.009077431634068489, + 0.010206645354628563, + 0.0031383519526571035, + 0.003002539975568652, + 0.003863525576889515, + 0.005133950617164373, + 0.006782045587897301, + 0.005183026660233736, + 0.005838322918862104, + 0.0052705914713442326, + 0.004713238682597876, + 0.00610383739694953, + 0.004527416080236435, + 0.004884718917310238, + 0.005680430680513382, + 0.004202198702841997, + 0.0054375543259084225, + 0.007910228334367275, + 0.010537582449615002, + 0.01129093300551176, + 0.006234948057681322, + 0.011911118403077126, + 0.006329549942165613, + 0.009619854390621185, + 0.007309185341000557, + 0.011078624054789543, + 0.0039235432632267475, + 0.004357742145657539, + 0.002788358833640814, + 0.0022866353392601013, + 0.003822342026978731, + 0.002573442878201604, + 0.005214038770645857, + 0.0028881931211799383, + 0.008509541861712933, + 0.009329359978437424, + 0.007916111499071121, + 0.009245169349014759, + 0.009940199553966522, + 0.008595865219831467, + 0.0067682513035833836, + 0.005992238875478506, + 0.005222302861511707, + 0.005193097982555628, + 0.005954679101705551, + 0.003387744538486004, + 0.0059708599001169205, + 0.004355516284704208, + 0.0033973383251577616, + 0.005759302992373705, + 0.03252485394477844, + 0.02186933532357216, + 0.021432217210531235, + 0.008157453499734402, + 0.013503213413059711, + 0.011818957515060902, + 0.010335488244891167, + 0.012358109466731548, + 0.06289801001548767, + 0.0548940971493721, + 0.013443579897284508, + 0.01695689558982849, + 0.024017196148633957, + 0.03600219264626503, + 0.03480241820216179, + 0.06514637172222137, + 0.02100216969847679, + 0.026485290378332138, + 0.032066699117422104, + 0.0229997206479311, + 0.01918608881533146, + 0.027581987902522087, + 0.0259652491658926, + 0.022504596039652824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028178293723613024 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028178293723613024, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028178293723613024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044927798211574554, + 0.02824336290359497, + 0.032970208674669266, + 0.054258350282907486, + 0.03290865570306778, + 0.05441444367170334, + 0.13075311481952667, + 0.08555500954389572, + 0.05113404989242554, + 0.4848349988460541 + ], + "credit_map": [ + 0.004822651389986277, + 0.0032410086132586002, + 0.003929716069251299, + 0.0034857499413192272, + 0.008685339242219925, + 0.006081432569772005, + 0.006771397776901722, + 0.007910500280559063, + 0.003384784096851945, + 0.0024141878820955753, + 0.002697763964533806, + 0.003339448943734169, + 0.004297812003642321, + 0.0032810880802571774, + 0.004621428437530994, + 0.0042068492621183395, + 0.0027977884747087955, + 0.003078936832025647, + 0.0038535338826477528, + 0.0031194626353681087, + 0.00443820096552372, + 0.0040182918310165405, + 0.004097762051969767, + 0.0075662326999008656, + 0.006075178738683462, + 0.008049781434237957, + 0.006943207699805498, + 0.007885841652750969, + 0.005173286888748407, + 0.006235517095774412, + 0.006490766070783138, + 0.007404768373817205, + 0.005441301967948675, + 0.004886439070105553, + 0.005064734257757664, + 0.002327857306227088, + 0.0028720456175506115, + 0.006027041003108025, + 0.0034778770059347153, + 0.002811357844620943, + 0.006248513236641884, + 0.009320943616330624, + 0.005798917263746262, + 0.007350891828536987, + 0.007757102604955435, + 0.006408891640603542, + 0.005886743310838938, + 0.005642435979098082, + 0.01923314668238163, + 0.022344574332237244, + 0.024460066109895706, + 0.009047569707036018, + 0.016904475167393684, + 0.01141700241714716, + 0.011359580792486668, + 0.015986699610948563, + 0.016605554148554802, + 0.019917314872145653, + 0.012219827622175217, + 0.0054214042611420155, + 0.008223545737564564, + 0.007573904003947973, + 0.007953012362122536, + 0.007640453986823559, + 0.01681489869952202, + 0.011973120272159576, + 0.0028264166321605444, + 0.0019905243534594774, + 0.004389946814626455, + 0.003711891360580921, + 0.003004414029419422, + 0.006422838196158409, + 0.04038450866937637, + 0.06660453975200653, + 0.0488644503057003, + 0.0756460502743721, + 0.03734825178980827, + 0.05727921053767204, + 0.09733163565397263, + 0.06137635558843613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.0028178293723613024 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05485962703824043, + 0.027597161009907722, + 0.032944608479738235, + 0.05293409898877144, + 0.014100983738899231, + 0.05715841054916382, + 0.0580623596906662, + 0.02543279528617859, + 0.34719735383987427, + 0.32971256971359253 + ], + "credit_map": [ + 0.005654301960021257, + 0.004083257168531418, + 0.004498570691794157, + 0.0049968683160841465, + 0.008617687970399857, + 0.005583853926509619, + 0.013105935417115688, + 0.008319155313074589, + 0.0023887110874056816, + 0.003228492569178343, + 0.0025719180703163147, + 0.0037514371797442436, + 0.003994880244135857, + 0.0042589460499584675, + 0.003918853122740984, + 0.0034839233849197626, + 0.002996510360389948, + 0.002631428185850382, + 0.0037755384109914303, + 0.0027250912971794605, + 0.002970481291413307, + 0.005200990475714207, + 0.003997633699327707, + 0.008646936155855656, + 0.0057867467403411865, + 0.007841426879167557, + 0.006774954032152891, + 0.008239291608333588, + 0.0049943337216973305, + 0.007105814293026924, + 0.006127719301730394, + 0.006063816137611866, + 0.002351907314732671, + 0.0023040499072521925, + 0.0019161747768521309, + 0.000877206155564636, + 0.00135202263481915, + 0.0022826821077615023, + 0.0016764502506703138, + 0.001340491813607514, + 0.0059800418093800545, + 0.009613808244466782, + 0.007877003401517868, + 0.008040644228458405, + 0.008309882134199142, + 0.005926709156483412, + 0.005783708766102791, + 0.005626617465168238, + 0.0069928960874676704, + 0.00928651262074709, + 0.0074881333857774734, + 0.0047280690632760525, + 0.009157393127679825, + 0.005738851614296436, + 0.004217028152197599, + 0.0104534812271595, + 0.003605068428441882, + 0.004209698177874088, + 0.003070427104830742, + 0.0025298800319433212, + 0.0031390124931931496, + 0.0027268542908132076, + 0.0034767696633934975, + 0.0026750850956887007, + 0.11906274408102036, + 0.07986616343259811, + 0.01938403956592083, + 0.01325197797268629, + 0.026313606649637222, + 0.025622935965657234, + 0.020118089392781258, + 0.04357784241437912, + 0.027667241171002388, + 0.04448460787534714, + 0.035332091152668, + 0.0505792573094368, + 0.0254568662494421, + 0.0398816242814064, + 0.06439641863107681, + 0.041914504021406174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.00978058110922575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle, side-on tracking shot, captured in the style of a striking sports animation, follows a group of track and field athletes sprinting along a bright red plastic track. Under the dazzling stadium floodlights, their sweat-soaked skin reflects off the highlights, and the contours of each muscle tensed by exertion are clearly visible. Suddenly, as the camera pans in sync with the athletes, a leading runner, wearing a bright yellow sleeveless sports vest and sporting short, sleek black hair, bursts forth from the tightly packed front group. With a remarkably decisive gesture, he lowers his center of gravity and swings his arms with a faster tempo and greater amplitude. As he accelerates, the entire frame is filled with intense motion blur and countless white speed lines radiating from behind him. The other runners and the stands in the background instantly transform into flowing, abstract blocks of color. The camera quickly follows and zooms in slightly, focusing on his angular profile. He clenches his teeth, beads of sweat dripping from his temple, and his wide eyes burn with unwavering determination, his pupils fixed firmly on the finish line just beyond his sight. This series of action breakdowns vividly depicts the moment he surpasses his limits and leaves his opponent behind, and the entire scene is filled with a breathtaking tension and adrenaline-pumping excitement.", + "seed": 33468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_4/index_9609_seed_33468_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": -0.29149733325466515, + "vlm_logprob": -0.29149733325466515, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.29149733325466515, + "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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03132949024438858 + ], + "opposite_logprob_tokens": [ + -3.531329393386841 + ], + "logprob_reward": -0.03132949024438858, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03132949024438858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16090913116931915, + 0.08681600540876389, + 0.17069634795188904, + 0.047913700342178345, + 0.0848667100071907, + 0.03860746696591377, + 0.029331112280488014, + 0.23551137745380402, + 0.0799785628914833, + 0.06536965817213058 + ], + "credit_map": [ + 0.01907925121486187, + 0.02198750711977482, + 0.029169220477342606, + 0.01902288943529129, + 0.015164951793849468, + 0.017562847584486008, + 0.017052436247467995, + 0.021870017051696777, + 0.009107890538871288, + 0.011869997717440128, + 0.005865104496479034, + 0.014303768053650856, + 0.013124172575771809, + 0.01819148100912571, + 0.008265258744359016, + 0.006088330410420895, + 0.03505561873316765, + 0.016940196976065636, + 0.00995966512709856, + 0.012605362571775913, + 0.021428139880299568, + 0.05487265810370445, + 0.010030421428382397, + 0.009804286994040012, + 0.007075081113725901, + 0.009627911262214184, + 0.005170394200831652, + 0.004669638816267252, + 0.004695861600339413, + 0.005465722177177668, + 0.0068986257538199425, + 0.00431046774610877, + 0.014883804135024548, + 0.02977316454052925, + 0.004171338863670826, + 0.0036482983268797398, + 0.014414975419640541, + 0.0055833784863352776, + 0.008044635877013206, + 0.004347109701484442, + 0.007021159399300814, + 0.007288639433681965, + 0.0035557951778173447, + 0.002861005486920476, + 0.006243946496397257, + 0.004341879393905401, + 0.00440471526235342, + 0.0028903260827064514, + 0.005126101430505514, + 0.004022582434117794, + 0.0021870313212275505, + 0.0038811336271464825, + 0.0025706372689455748, + 0.00468134880065918, + 0.0039188312366604805, + 0.0029434484895318747, + 0.06482981890439987, + 0.09613543748855591, + 0.010695784352719784, + 0.009519302286207676, + 0.018874328583478928, + 0.009951934218406677, + 0.015789609402418137, + 0.009715168736875057, + 0.013918042182922363, + 0.03337261825799942, + 0.0037783998996019363, + 0.003471899777650833, + 0.008864554576575756, + 0.006060055457055569, + 0.0050882743671536446, + 0.005424719303846359, + 0.01181931421160698, + 0.006009113509207964, + 0.009564605541527271, + 0.005576198920607567, + 0.00996342021971941, + 0.008517805486917496, + 0.008513606153428555, + 0.00540559459477663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.03132949024438858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.4063441753387451 + ], + "opposite_logprob_tokens": [ + -0.2813442051410675 + ], + "logprob_reward": -1.4063441753387451, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063441753387451, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09899522364139557, + 0.1143798679113388, + 0.09084241092205048, + 0.05788488686084747, + 0.06351909786462784, + 0.07211936265230179, + 0.02531714178621769, + 0.2367449849843979, + 0.09012646973133087, + 0.15007056295871735 + ], + "credit_map": [ + 0.011561725288629532, + 0.016016654670238495, + 0.011928613297641277, + 0.018449347466230392, + 0.01103804912418127, + 0.010173682123422623, + 0.010216240771114826, + 0.00961090438067913, + 0.011851712130010128, + 0.017206912860274315, + 0.009150638245046139, + 0.014045076444745064, + 0.0193887110799551, + 0.02448018454015255, + 0.0102093406021595, + 0.008047282695770264, + 0.020857125520706177, + 0.010040583088994026, + 0.00502043729647994, + 0.00577550521120429, + 0.010482999496161938, + 0.029235780239105225, + 0.00525272823870182, + 0.0041772485710680485, + 0.008746030740439892, + 0.010647278279066086, + 0.0067786043509840965, + 0.004740497563034296, + 0.0059288437478244305, + 0.0066521610133349895, + 0.00741328252479434, + 0.006978179793804884, + 0.013693829998373985, + 0.015636583790183067, + 0.006875375751405954, + 0.0036253060679882765, + 0.007629071827977896, + 0.006087432149797678, + 0.005741256754845381, + 0.004230234771966934, + 0.0142095647752285, + 0.012700800783932209, + 0.00650868471711874, + 0.005730397533625364, + 0.009944343008100986, + 0.008780230768024921, + 0.0068791271187365055, + 0.007366209290921688, + 0.004981754347681999, + 0.004160101525485516, + 0.001840938930399716, + 0.004783697426319122, + 0.0018199929036200047, + 0.002380057703703642, + 0.0027211331762373447, + 0.002629463095217943, + 0.06490643322467804, + 0.09246267378330231, + 0.008900781162083149, + 0.00773649662733078, + 0.017974000424146652, + 0.021417368203401566, + 0.011778876185417175, + 0.011568332090973854, + 0.008763669058680534, + 0.030634548515081406, + 0.004459941294044256, + 0.006162729579955339, + 0.010129183530807495, + 0.008430084213614464, + 0.006061334628611803, + 0.015484970062971115, + 0.0320914164185524, + 0.01153480913490057, + 0.010668433271348476, + 0.014104967936873436, + 0.03165249153971672, + 0.016258813440799713, + 0.022209078073501587, + 0.011550545692443848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -1.4063441753387451 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463957479223609 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.00463957479223609, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463957479223609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08656568825244904, + 0.07293535768985748, + 0.135635644197464, + 0.050011370331048965, + 0.16266097128391266, + 0.05213487893342972, + 0.03557445853948593, + 0.24058528244495392, + 0.03356923907995224, + 0.13032712042331696 + ], + "credit_map": [ + 0.011425236240029335, + 0.009088345803320408, + 0.012944952584803104, + 0.013108470477163792, + 0.010219749994575977, + 0.0076727597042918205, + 0.010589237324893475, + 0.011516930535435677, + 0.007299399469047785, + 0.010591992177069187, + 0.00551572535187006, + 0.008691549301147461, + 0.012512274086475372, + 0.014176401309669018, + 0.008972644805908203, + 0.005175367463380098, + 0.025618581101298332, + 0.011671827174723148, + 0.008699240162968636, + 0.01102167647331953, + 0.017634468153119087, + 0.043937038630247116, + 0.010020259767770767, + 0.007032538298517466, + 0.008132852613925934, + 0.012413710355758667, + 0.003843768499791622, + 0.003257520031183958, + 0.005622450727969408, + 0.005619630217552185, + 0.00514683872461319, + 0.005974599625915289, + 0.03549065813422203, + 0.061283405870199203, + 0.005337720271199942, + 0.005508402828127146, + 0.029315749183297157, + 0.008416957221925259, + 0.010768643580377102, + 0.0065394435077905655, + 0.010025528259575367, + 0.010919421911239624, + 0.00457201711833477, + 0.0033006942830979824, + 0.007945528253912926, + 0.005400723312050104, + 0.005566972307860851, + 0.0044039953500032425, + 0.00552033819258213, + 0.004896835424005985, + 0.0038342324551194906, + 0.005682965740561485, + 0.0031539045739918947, + 0.004684783052653074, + 0.00424240343272686, + 0.00355899496935308, + 0.06609819084405899, + 0.09156624227762222, + 0.014228954911231995, + 0.009124111384153366, + 0.01826152391731739, + 0.015188244171440601, + 0.014760546386241913, + 0.011357485316693783, + 0.002974971430376172, + 0.007967155426740646, + 0.002568231662735343, + 0.002782362513244152, + 0.00359989982098341, + 0.004022594541311264, + 0.003656471148133278, + 0.005997552536427975, + 0.019834714010357857, + 0.012598356232047081, + 0.011530493386089802, + 0.010838027112185955, + 0.02986166439950466, + 0.020263103768229485, + 0.01556405983865261, + 0.009836703538894653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.00463957479223609 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095619544386864 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095619544386864, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095619544386864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07355895638465881, + 0.09715736657381058, + 0.10136206448078156, + 0.04007996618747711, + 0.10982146114110947, + 0.044555533677339554, + 0.05277954041957855, + 0.12673445045948029, + 0.036639146506786346, + 0.31731149554252625 + ], + "credit_map": [ + 0.009127938188612461, + 0.006612722761929035, + 0.008526015095412731, + 0.011393730528652668, + 0.01236545667052269, + 0.006598814390599728, + 0.009529003873467445, + 0.009405277669429779, + 0.010088934563100338, + 0.013234536163508892, + 0.0069154598750174046, + 0.009864035062491894, + 0.016038542613387108, + 0.022560788318514824, + 0.011056007817387581, + 0.007399062160402536, + 0.017690706998109818, + 0.012123160064220428, + 0.009652076289057732, + 0.007825717329978943, + 0.010470053181052208, + 0.025290247052907944, + 0.009724182076752186, + 0.008585922420024872, + 0.007476412691175938, + 0.011820885352790356, + 0.003283002180978656, + 0.0026224965695291758, + 0.004892309661954641, + 0.004141252022236586, + 0.0033083390444517136, + 0.0025352698285132647, + 0.021537741646170616, + 0.03553325682878494, + 0.005067093297839165, + 0.006570690777152777, + 0.01563984528183937, + 0.007110271602869034, + 0.011194426566362381, + 0.00716814212501049, + 0.011029957793653011, + 0.010176788084208965, + 0.0030996815767139196, + 0.002747555263340473, + 0.00618318235501647, + 0.0040099178440868855, + 0.003936417866498232, + 0.003372031496837735, + 0.010687943547964096, + 0.012882464565336704, + 0.004210219252854586, + 0.006157732103019953, + 0.00430294731631875, + 0.005226414650678635, + 0.004754664376378059, + 0.004557155072689056, + 0.03140714764595032, + 0.020978696644306183, + 0.03264978155493736, + 0.008125213906168938, + 0.00783473439514637, + 0.010412770323455334, + 0.008221848867833614, + 0.007104259915649891, + 0.004328386392444372, + 0.012517089024186134, + 0.002289649099111557, + 0.0024625728838145733, + 0.0037651227321475744, + 0.0036097364500164986, + 0.004184582736343145, + 0.003482006723061204, + 0.07428036630153656, + 0.02141302078962326, + 0.03863843157887459, + 0.026582596823573112, + 0.07858643680810928, + 0.02100272662937641, + 0.03685309737920761, + 0.019954828545451164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.004095619544386864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 track and field athletes visible sprinting along a bright red plastic track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene lit by dazzling stadium floodlights showing sweat-soaked skin and tensed muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a leading runner wearing a bright yellow sleeveless sports vest and short black hair burst from the front group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frame filled with intense motion blur and white speed lines radiating from behind the leading runner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the runner with a low-angle side-on tracking shot that includes a slight zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0539851151406765, + 0.07375869154930115, + 0.167280375957489, + 0.04285990446805954, + 0.12498337030410767, + 0.06355734169483185, + 0.07681453227996826, + 0.20916861295700073, + 0.04184463992714882, + 0.14574743807315826 + ], + "credit_map": [ + 0.006008588243275881, + 0.0056703719310462475, + 0.005808320827782154, + 0.009722880087792873, + 0.0065869856625795364, + 0.006206256803125143, + 0.007269572000950575, + 0.006712139584124088, + 0.007660557050257921, + 0.01021235715597868, + 0.007530310191214085, + 0.008826090022921562, + 0.012472596950829029, + 0.011944939382374287, + 0.007558554410934448, + 0.007553288713097572, + 0.03046061471104622, + 0.017095470800995827, + 0.010978271253407001, + 0.014618870802223682, + 0.020944520831108093, + 0.04599643871188164, + 0.016923341900110245, + 0.01026284322142601, + 0.007603418547660112, + 0.011879701167345047, + 0.0034702843986451626, + 0.002883798209950328, + 0.004652282223105431, + 0.0043640765361487865, + 0.004187811631709337, + 0.0038185317534953356, + 0.02792634814977646, + 0.04412872716784477, + 0.005197706166654825, + 0.005124840419739485, + 0.01950993947684765, + 0.00711619621142745, + 0.0086442856118083, + 0.0073353201150894165, + 0.013344463892281055, + 0.015156514942646027, + 0.004465010017156601, + 0.004411946516484022, + 0.008717592805624008, + 0.006559431087225676, + 0.0066118245013058186, + 0.004290561191737652, + 0.012324794195592403, + 0.009168102405965328, + 0.006662367843091488, + 0.009956888854503632, + 0.006826369091868401, + 0.011166750453412533, + 0.01051880232989788, + 0.01019045989960432, + 0.027320554479956627, + 0.07140175998210907, + 0.052507560700178146, + 0.009525535628199577, + 0.011483793146908283, + 0.01707926020026207, + 0.011167030781507492, + 0.00868312083184719, + 0.006251814775168896, + 0.011004343628883362, + 0.003956153057515621, + 0.0035465210676193237, + 0.0035608187317848206, + 0.004468952771276236, + 0.004991678986698389, + 0.004064359236508608, + 0.027327466756105423, + 0.012810948304831982, + 0.010733678936958313, + 0.009484214708209038, + 0.04225432500243187, + 0.020764794200658798, + 0.013572043739259243, + 0.00879995059221983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.01107780635356903 + } + ], + "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/0000122/group_5/index_tensor([1538, 1538])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_5/index_tensor([1538, 1538])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8458a0ae1e4d0339d4a2cd0f993e8151ab648317 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_5/index_tensor([1538, 1538])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Captured in stunning 8K Ultra HD with a macro lens, this extreme close-up remains perfectly still, focusing on the human eye with razor-sharp clarity. The iris is a complex galaxy of deep hazel, flecked with subtle amber, its intricate fibrous pattern rendered in surreal detail, while the slightly dilated pupil holds a soft spotlight, suggesting a dimly lit, intimate space. The peripheral retina is faintly tinged with blood-red, a delicate network of red veins betraying a deep, unspoken sorrow. Slowly, a crystal tear inexorably pours out, gathering along the lower waterline until surface tension creates a shimmering, convex pool of liquid, magnifying the thin, black eyelashes beneath. For a breathtaking moment, the tear maintains its form, trembling at the edge of the lower eyelid. Then, with a slow, inescapable force of gravity, it spills over the precipice, a perfect droplet tracing a shimmering, silent path along the skin. The entire poignant micro-drama unfolds amidst an atmosphere of profound intimacy and heartbreaking melancholy.", + "seed": 22717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_5/index_1538_seed_22717_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": -0.18427972886711358, + "vlm_logprob": -0.18427972886711358, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.18427972886711358, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006466061808168888 + ], + "opposite_logprob_tokens": [ + -5.881465911865234 + ], + "logprob_reward": -0.006466061808168888, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006466061808168888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37440839409828186, + 0.06626302003860474, + 0.031754542142152786, + 0.034387826919555664, + 0.10135219991207123, + 0.03555898740887642, + 0.07151758670806885, + 0.17840096354484558, + 0.07449347525835037, + 0.031863026320934296 + ], + "credit_map": [ + 0.02298569306731224, + 0.009196294471621513, + 0.010360677726566792, + 0.015122288838028908, + 0.02618538960814476, + 0.133286252617836, + 0.017874054610729218, + 0.1393977403640747, + 0.007269084919244051, + 0.005370818544179201, + 0.011753307655453682, + 0.005694221705198288, + 0.010507674887776375, + 0.007685174699872732, + 0.00858067162334919, + 0.009402064606547356, + 0.005944699514657259, + 0.0021117369178682566, + 0.004067814443260431, + 0.003147607669234276, + 0.005029013846069574, + 0.0032584993168711662, + 0.004204767290502787, + 0.003990404307842255, + 0.00468796119093895, + 0.0021622017957270145, + 0.0050186472944915295, + 0.0037155048921704292, + 0.005132874008268118, + 0.0038141959812492132, + 0.005197472870349884, + 0.0046589672565460205, + 0.01271379180252552, + 0.005910526495426893, + 0.012979459017515182, + 0.010071231052279472, + 0.01343451626598835, + 0.009750551544129848, + 0.0204623956233263, + 0.016029732301831245, + 0.0031106537207961082, + 0.0028833618853241205, + 0.0054984851740300655, + 0.00416311202570796, + 0.004995068535208702, + 0.0033770145382732153, + 0.007529212161898613, + 0.004002079833298922, + 0.0056668356992304325, + 0.006660915911197662, + 0.011729315854609013, + 0.007733289618045092, + 0.01228860579431057, + 0.009146537631750107, + 0.008558697067201138, + 0.009733393788337708, + 0.038518115878105164, + 0.01258933823555708, + 0.032808493822813034, + 0.013452158309519291, + 0.01908116415143013, + 0.020372984930872917, + 0.015383555553853512, + 0.02619515173137188, + 0.01714908331632614, + 0.005567790009081364, + 0.01314136665314436, + 0.005085819400846958, + 0.007881861180067062, + 0.008928976953029633, + 0.006679868791252375, + 0.010058706626296043, + 0.005170103628188372, + 0.0028379797004163265, + 0.004951672162860632, + 0.003391142701730132, + 0.005192517768591642, + 0.003570987842977047, + 0.003139507956802845, + 0.003609113860875368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.006466061808168888 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005704939365387 + ], + "opposite_logprob_tokens": [ + -2.6950571537017822 + ], + "logprob_reward": -0.07005704939365387, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005704939365387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06122134253382683, + 0.10340963304042816, + 0.16069509088993073, + 0.0666244849562645, + 0.11194583773612976, + 0.045370228588581085, + 0.05245901271700859, + 0.06952174752950668, + 0.04729011654853821, + 0.28146257996559143 + ], + "credit_map": [ + 0.005640322808176279, + 0.005790072027593851, + 0.007100329268723726, + 0.00429764948785305, + 0.013773689046502113, + 0.009385737590491772, + 0.008754719980061054, + 0.006478826515376568, + 0.01531736459583044, + 0.008056163787841797, + 0.008704237639904022, + 0.005758179817348719, + 0.026898235082626343, + 0.013202947564423084, + 0.008869744837284088, + 0.016602754592895508, + 0.015056990087032318, + 0.014162040315568447, + 0.016881821677088737, + 0.008848891593515873, + 0.04012702405452728, + 0.021359365433454514, + 0.017289219424128532, + 0.026969734579324722, + 0.003923698328435421, + 0.0049288892187178135, + 0.0057987794280052185, + 0.006474652327597141, + 0.016106897965073586, + 0.008603768423199654, + 0.012394757010042667, + 0.008393037132918835, + 0.011405537836253643, + 0.007256580516695976, + 0.013529003597795963, + 0.011691100895404816, + 0.01652620919048786, + 0.010369518771767616, + 0.025482183322310448, + 0.015685714781284332, + 0.004281563684344292, + 0.0037119006738066673, + 0.006457597017288208, + 0.004621125292032957, + 0.005750548560172319, + 0.004764290526509285, + 0.008385955356061459, + 0.007397247478365898, + 0.007563683204352856, + 0.004718967247754335, + 0.009034091606736183, + 0.0037874348927289248, + 0.008055041544139385, + 0.0066070472821593285, + 0.0049745202995836735, + 0.007718226406723261, + 0.013774308376014233, + 0.005685330834239721, + 0.012096366845071316, + 0.0048261964693665504, + 0.00951379630714655, + 0.008207142353057861, + 0.005773223005235195, + 0.009645381942391396, + 0.009602755308151245, + 0.0036376221105456352, + 0.006194644141942263, + 0.0033360866364091635, + 0.008448222652077675, + 0.006296427454799414, + 0.004648290574550629, + 0.005126067437231541, + 0.054275207221508026, + 0.030569899827241898, + 0.04025459662079811, + 0.021227341145277023, + 0.04354904219508171, + 0.034751471132040024, + 0.025191759690642357, + 0.03164324909448624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.07005704939365387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307379201054573 + ], + "opposite_logprob_tokens": [ + -3.168073892593384 + ], + "logprob_reward": -0.04307379201054573, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307379201054573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19977083802223206, + 0.04562204331159592, + 0.06098750978708267, + 0.04929901659488678, + 0.4107341468334198, + 0.06001460552215576, + 0.043747954070568085, + 0.049555301666259766, + 0.04500095546245575, + 0.03526762127876282 + ], + "credit_map": [ + 0.014214804396033287, + 0.005170382093638182, + 0.008617174811661243, + 0.007633749395608902, + 0.017318466678261757, + 0.06470780819654465, + 0.009244152344763279, + 0.07286430150270462, + 0.006511163432151079, + 0.003892559092491865, + 0.0070618330501019955, + 0.004945404827594757, + 0.005377204157412052, + 0.005841354373842478, + 0.005296727642416954, + 0.006695798132568598, + 0.012660784646868706, + 0.0041788979433476925, + 0.00952856708317995, + 0.004413285758346319, + 0.007524267770349979, + 0.009002656675875187, + 0.005587202962487936, + 0.008091842755675316, + 0.0060169147327542305, + 0.003644718788564205, + 0.007010894827544689, + 0.0050202831625938416, + 0.006414505187422037, + 0.00642964243888855, + 0.008397375233471394, + 0.006364678032696247, + 0.05087459832429886, + 0.025203948840498924, + 0.0559673011302948, + 0.03853316232562065, + 0.0555931031703949, + 0.03679664060473442, + 0.09191899001598358, + 0.05584639310836792, + 0.006942980457097292, + 0.0036240126937627792, + 0.011632845737040043, + 0.006981920450925827, + 0.006516674999147654, + 0.006490214262157679, + 0.010367452166974545, + 0.007458509411662817, + 0.007843751460313797, + 0.0028782901354134083, + 0.006747109349817038, + 0.004335098899900913, + 0.005302634555846453, + 0.004435828421264887, + 0.003918656148016453, + 0.008286586031317711, + 0.00875517725944519, + 0.003907287493348122, + 0.008203120902180672, + 0.0030576298013329506, + 0.007617649622261524, + 0.005673523526638746, + 0.0048615108244121075, + 0.007479404099285603, + 0.007807999383658171, + 0.0035509951412677765, + 0.007840132340788841, + 0.004524812567979097, + 0.0072321114130318165, + 0.004777989350259304, + 0.0036874031648039818, + 0.005579512100666761, + 0.007207653485238552, + 0.003302759025245905, + 0.0060465638525784016, + 0.0029255140107125044, + 0.004090231377631426, + 0.004170338623225689, + 0.0031339379493147135, + 0.004390623886138201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.04307379201054573 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761857032775879 + ], + "opposite_logprob_tokens": [ + -0.8261857032775879 + ], + "logprob_reward": -0.5761857032775879, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761857032775879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3208909034729004, + 0.07246540486812592, + 0.0469030998647213, + 0.04284614697098732, + 0.31749942898750305, + 0.04405447095632553, + 0.05886272341012955, + 0.03553816303610802, + 0.03917965292930603, + 0.02176004648208618 + ], + "credit_map": [ + 0.01943575218319893, + 0.008230616338551044, + 0.013634596951305866, + 0.013544763438403606, + 0.026155265048146248, + 0.10309898853302002, + 0.014343847520649433, + 0.12244703620672226, + 0.009667587466537952, + 0.004689733497798443, + 0.011361059732735157, + 0.008888880722224712, + 0.008771086111664772, + 0.007207182701677084, + 0.010793582536280155, + 0.011086286045610905, + 0.008808623068034649, + 0.0032158379908651114, + 0.007159092463552952, + 0.0035429089330136776, + 0.00619022361934185, + 0.006836963817477226, + 0.004583035595715046, + 0.0065664141438901424, + 0.003276518313214183, + 0.0028580406215041876, + 0.004310559015721083, + 0.005264220759272575, + 0.005055124871432781, + 0.003919281531125307, + 0.012209568172693253, + 0.005952827632427216, + 0.04510166868567467, + 0.018330998718738556, + 0.0424068346619606, + 0.032622288912534714, + 0.04031234607100487, + 0.027801502496004105, + 0.06104717403650284, + 0.04987657815217972, + 0.004047862254083157, + 0.0030752508901059628, + 0.008676462806761265, + 0.004999805707484484, + 0.005135402083396912, + 0.004651505500078201, + 0.00680769607424736, + 0.006660480983555317, + 0.008937537670135498, + 0.0042527769692242146, + 0.009330020286142826, + 0.004217399284243584, + 0.007684292271733284, + 0.006054176017642021, + 0.006345976144075394, + 0.01204054243862629, + 0.0027736604679375887, + 0.0032988539896905422, + 0.004100256599485874, + 0.003879275405779481, + 0.007897567935287952, + 0.003853141563013196, + 0.0037278197705745697, + 0.006007586605846882, + 0.00625037495046854, + 0.003606734797358513, + 0.005756318103522062, + 0.0037168373819440603, + 0.00701133394613862, + 0.004329524002969265, + 0.0035386516246944666, + 0.004969875328242779, + 0.0042829751037061214, + 0.0019800951704382896, + 0.0035902229137718678, + 0.0018088718643411994, + 0.002825418720021844, + 0.002473200438544154, + 0.0023918449878692627, + 0.0024074148386716843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.5761857032775879 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561603784561157 + ], + "opposite_logprob_tokens": [ + -1.6006159782409668 + ], + "logprob_reward": -0.22561603784561157, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561603784561157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2185746282339096, + 0.08151347190141678, + 0.04518953338265419, + 0.0660523846745491, + 0.30295655131340027, + 0.1031082272529602, + 0.050530560314655304, + 0.043295279145240784, + 0.042910341173410416, + 0.04586896300315857 + ], + "credit_map": [ + 0.015176071785390377, + 0.007104970049113035, + 0.013190493918955326, + 0.010859785601496696, + 0.0185098759829998, + 0.06416420638561249, + 0.010545581579208374, + 0.07902362942695618, + 0.0101019237190485, + 0.005608802195638418, + 0.013570536859333515, + 0.010105838067829609, + 0.01040017232298851, + 0.008158382959663868, + 0.012316761538386345, + 0.01125105656683445, + 0.006717597134411335, + 0.002895032986998558, + 0.006793182343244553, + 0.003854932961985469, + 0.006748263258486986, + 0.005787650588899851, + 0.005730506964027882, + 0.006662365049123764, + 0.008023206144571304, + 0.00347780529409647, + 0.008650660514831543, + 0.006453753914684057, + 0.00772512424737215, + 0.006475960835814476, + 0.015859367325901985, + 0.009386507794260979, + 0.052251510322093964, + 0.018783750012516975, + 0.04939467832446098, + 0.029612086713314056, + 0.035751160234212875, + 0.02688857913017273, + 0.04749661684036255, + 0.042778171598911285, + 0.007837392389774323, + 0.0063532693311572075, + 0.019562220200896263, + 0.01083797961473465, + 0.01245284266769886, + 0.01129013393074274, + 0.019633987918496132, + 0.015140398405492306, + 0.007475605234503746, + 0.0033637387678027153, + 0.009190455079078674, + 0.00395204545930028, + 0.006570178549736738, + 0.004938879981637001, + 0.005299243610352278, + 0.009740411303937435, + 0.004520490765571594, + 0.004223352298140526, + 0.005021894350647926, + 0.004644128959625959, + 0.008645455352962017, + 0.004404740873724222, + 0.004091709386557341, + 0.007743506226688623, + 0.005593523848801851, + 0.00428045867010951, + 0.00562948826700449, + 0.004234046675264835, + 0.00886036828160286, + 0.004783371929079294, + 0.0037765749730169773, + 0.005752507597208023, + 0.00914777535945177, + 0.004896616097539663, + 0.007137068547308445, + 0.0037122999783605337, + 0.006078196223825216, + 0.005616747308522463, + 0.003921125549823046, + 0.0053591374307870865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.22561603784561157 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in stunning 8K Ultra HD with a macro lens, this extreme close-up remains perfectly still, focusing on the human eye with razor-sharp clarity. The iris is a complex galaxy of deep hazel, flecked with subtle amber, its intricate fibrous pattern rendered in surreal detail, while the slightly dilated pupil holds a soft spotlight, suggesting a dimly lit, intimate space. The peripheral retina is faintly tinged with blood-red, a delicate network of red veins betraying a deep, unspoken sorrow. Slowly, a crystal tear inexorably pours out, gathering along the lower waterline until surface tension creates a shimmering, convex pool of liquid, magnifying the thin, black eyelashes beneath. For a breathtaking moment, the tear maintains its form, trembling at the edge of the lower eyelid. Then, with a slow, inescapable force of gravity, it spills over the precipice, a perfect droplet tracing a shimmering, silent path along the skin. The entire poignant micro-drama unfolds amidst an atmosphere of profound intimacy and heartbreaking melancholy.", + "seed": 22718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_5/index_1538_seed_22718_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": -0.154817795753479, + "vlm_logprob": -0.154817795753479, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.154817795753479, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01578443869948387 + ], + "opposite_logprob_tokens": [ + -4.39078426361084 + ], + "logprob_reward": -0.01578443869948387, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01578443869948387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04299991577863693, + 0.10211902856826782, + 0.022626176476478577, + 0.04148763790726662, + 0.04201579838991165, + 0.2309393733739853, + 0.08850660920143127, + 0.3915133774280548, + 0.01830120198428631, + 0.019490890204906464 + ], + "credit_map": [ + 0.004764176905155182, + 0.003896527225151658, + 0.005975963082164526, + 0.003466584486886859, + 0.00708155520260334, + 0.0054190391674637794, + 0.007543972693383694, + 0.004852093290537596, + 0.01855308748781681, + 0.0076788803562521935, + 0.012368015013635159, + 0.009123937226831913, + 0.010987985879182816, + 0.01862940564751625, + 0.009351355023682117, + 0.015426358208060265, + 0.002888001734390855, + 0.0011355449678376317, + 0.0032328960951417685, + 0.0013080830685794353, + 0.0021397829987108707, + 0.006950949784368277, + 0.0020085712894797325, + 0.0029623452574014664, + 0.00553404726088047, + 0.002961578778922558, + 0.0055312588810920715, + 0.004472148604691029, + 0.004134658258408308, + 0.008127721026539803, + 0.0047332895919680595, + 0.005992932245135307, + 0.005372549872845411, + 0.0022918053437024355, + 0.006654185708612204, + 0.002440862590447068, + 0.005357710178941488, + 0.011697898618876934, + 0.00403596181422472, + 0.004164820071309805, + 0.013130168430507183, + 0.004953500349074602, + 0.08106305450201035, + 0.00793799664825201, + 0.007194720208644867, + 0.09247035533189774, + 0.009729765355587006, + 0.014459787867963314, + 0.00572229316458106, + 0.0024416078813374043, + 0.012625295668840408, + 0.0031048699747771025, + 0.004887117072939873, + 0.037479955703020096, + 0.0046386150643229485, + 0.01760684885084629, + 0.020231779664754868, + 0.010399133898317814, + 0.04020852595567703, + 0.011991938576102257, + 0.01797880232334137, + 0.17317801713943481, + 0.013233842328190804, + 0.10429133474826813, + 0.0032642686273902655, + 0.0015516035491600633, + 0.0031757911201566458, + 0.0019410401582717896, + 0.0017777439206838608, + 0.0018930814694613218, + 0.002241607988253236, + 0.002456064336001873, + 0.0015847192844375968, + 0.0009509279043413699, + 0.005422037094831467, + 0.0016218399396166205, + 0.0023716711439192295, + 0.004625245928764343, + 0.0014542602002620697, + 0.001460186205804348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.01578443869948387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373636677861214 + ], + "opposite_logprob_tokens": [ + -3.40873646736145 + ], + "logprob_reward": -0.03373636677861214, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373636677861214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25801607966423035, + 0.10534011572599411, + 0.025735516101121902, + 0.12831561267375946, + 0.02758912183344364, + 0.18413202464580536, + 0.04294984042644501, + 0.1565234214067459, + 0.008624077774584293, + 0.0627741813659668 + ], + "credit_map": [ + 0.03155626729130745, + 0.02682153694331646, + 0.020736457780003548, + 0.028765857219696045, + 0.04040423408150673, + 0.0315634049475193, + 0.045503880828619, + 0.03266442194581032, + 0.01874811016023159, + 0.009496144019067287, + 0.011737663298845291, + 0.010647487826645374, + 0.012720688246190548, + 0.012254574336111546, + 0.011781524866819382, + 0.01795392669737339, + 0.003116535721346736, + 0.002219256479293108, + 0.00290410197339952, + 0.0022807519417256117, + 0.005659783724695444, + 0.004064110107719898, + 0.0027288133278489113, + 0.0027621639892458916, + 0.013305027969181538, + 0.009415901266038418, + 0.016137726604938507, + 0.01081419363617897, + 0.014846419915556908, + 0.03389928862452507, + 0.010586640797555447, + 0.01931041106581688, + 0.002911705756559968, + 0.0024108700454235077, + 0.0028705468866974115, + 0.0019006117945536971, + 0.007808510214090347, + 0.003978710621595383, + 0.0036181623581796885, + 0.0020900038070976734, + 0.010511582717299461, + 0.004107411485165358, + 0.06350177526473999, + 0.008630364201962948, + 0.006485730409622192, + 0.06660206615924835, + 0.012126828543841839, + 0.012166270054876804, + 0.0036871235352009535, + 0.0020483280532062054, + 0.00312807597219944, + 0.0025108668487519026, + 0.006016843020915985, + 0.01219442579895258, + 0.003943631891161203, + 0.009420545771718025, + 0.00946967862546444, + 0.0051065064035356045, + 0.016667982563376427, + 0.006535678170621395, + 0.008350752294063568, + 0.06478564441204071, + 0.006692120339721441, + 0.03891507163643837, + 0.0012088948860764503, + 0.000627115077804774, + 0.0015954719856381416, + 0.0007743429741822183, + 0.001246757572516799, + 0.0008226899662986398, + 0.0010340713197365403, + 0.0013147341087460518, + 0.003200683742761612, + 0.001848897198215127, + 0.0219956673681736, + 0.0036739578936249018, + 0.0026651921216398478, + 0.02260756306350231, + 0.0028917689342051744, + 0.0038904526736587286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.03373636677861214 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866340756416321 + ], + "opposite_logprob_tokens": [ + -3.048663377761841 + ], + "logprob_reward": -0.04866340756416321, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866340756416321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21434666216373444, + 0.10462255030870438, + 0.029974913224577904, + 0.02835681103169918, + 0.03456903621554375, + 0.4211635887622833, + 0.02641429752111435, + 0.09201565384864807, + 0.010191245004534721, + 0.038345158100128174 + ], + "credit_map": [ + 0.02542281523346901, + 0.02055811509490013, + 0.01870199479162693, + 0.020903848111629486, + 0.03655640035867691, + 0.030156617984175682, + 0.04001948609948158, + 0.022027399390935898, + 0.020193427801132202, + 0.007931618951261044, + 0.013448009267449379, + 0.008880369365215302, + 0.010783875361084938, + 0.01810537837445736, + 0.009401999413967133, + 0.015877874568104744, + 0.002883803565055132, + 0.0017946206498891115, + 0.003940201364457607, + 0.0029078389052301645, + 0.004630704410374165, + 0.007027690298855305, + 0.00308997998945415, + 0.003700073342770338, + 0.00504198158159852, + 0.001972937723621726, + 0.004279376473277807, + 0.002302883891388774, + 0.0036421252880245447, + 0.006005639210343361, + 0.0025296309031546116, + 0.0025822357274591923, + 0.0032842704094946384, + 0.0025783968158066273, + 0.004111244343221188, + 0.0016608197474852204, + 0.007699169684201479, + 0.007669564802199602, + 0.003732128767296672, + 0.0038334436248987913, + 0.02153942920267582, + 0.00837551336735487, + 0.15190424025058746, + 0.017161015421152115, + 0.012695075944066048, + 0.1623411625623703, + 0.021456278860569, + 0.025690877810120583, + 0.0025773111265152693, + 0.00083385338075459, + 0.0027481908909976482, + 0.0010884287767112255, + 0.0009591164998710155, + 0.009816745296120644, + 0.0020663621835410595, + 0.006324291229248047, + 0.005664278287440538, + 0.003173748729750514, + 0.009098357520997524, + 0.003938424866646528, + 0.005283767823129892, + 0.037395257502794266, + 0.004119564313441515, + 0.023342259228229523, + 0.0017774314619600773, + 0.0006309404270723462, + 0.002551374025642872, + 0.0009631162974983454, + 0.0009474916150793433, + 0.0013218640815466642, + 0.0008696737349964678, + 0.001129354233853519, + 0.003040160285308957, + 0.002427099272608757, + 0.009756120853126049, + 0.003065407508984208, + 0.006541260983794928, + 0.007769464515149593, + 0.0032969245221465826, + 0.0024487164337188005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.04866340756416321 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743114411830902 + ], + "opposite_logprob_tokens": [ + -0.9743114709854126 + ], + "logprob_reward": -0.4743114411830902, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743114411830902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2538517415523529, + 0.15487781167030334, + 0.06828318536281586, + 0.04587337002158165, + 0.0372917577624321, + 0.2788000702857971, + 0.0286631491035223, + 0.052174851298332214, + 0.020761968567967415, + 0.059422045946121216 + ], + "credit_map": [ + 0.02693355642259121, + 0.024342888966202736, + 0.02318480797111988, + 0.025306355208158493, + 0.04570648446679115, + 0.034274034202098846, + 0.05119869112968445, + 0.022904926910996437, + 0.02321239747107029, + 0.011828971095383167, + 0.01593983732163906, + 0.014133820310235023, + 0.02649819850921631, + 0.02285853587090969, + 0.017451783642172813, + 0.022954272106289864, + 0.005841279402375221, + 0.005178489256650209, + 0.00478270323947072, + 0.005917844828218222, + 0.022494615986943245, + 0.007783553563058376, + 0.010410688817501068, + 0.005874013062566519, + 0.006193230859935284, + 0.00376112200319767, + 0.005023949313908815, + 0.004444882273674011, + 0.009154705330729485, + 0.006480344105511904, + 0.005586098879575729, + 0.0052290367893874645, + 0.0063662477768957615, + 0.0029873063322156668, + 0.004840643610805273, + 0.0035566750448197126, + 0.003762026783078909, + 0.003560456680133939, + 0.004290332086384296, + 0.007928070612251759, + 0.02044999785721302, + 0.010199809446930885, + 0.07475775480270386, + 0.01694573275744915, + 0.022316977381706238, + 0.07042834162712097, + 0.04072962701320648, + 0.022971834987401962, + 0.002677182899788022, + 0.0011408182326704264, + 0.003246405627578497, + 0.0012709154980257154, + 0.0012912319507449865, + 0.010928517207503319, + 0.0018159181345254183, + 0.006292160600423813, + 0.0037642924580723047, + 0.0026756867300719023, + 0.004539883229881525, + 0.002756329020485282, + 0.004118857439607382, + 0.017785709351301193, + 0.0035221767611801624, + 0.013011916540563107, + 0.0027668809052556753, + 0.001263984595425427, + 0.005555643234401941, + 0.0021370623726397753, + 0.0018104601185768843, + 0.001921693212352693, + 0.0023236675187945366, + 0.0029825770761817694, + 0.004980600904673338, + 0.0032854382880032063, + 0.01715189963579178, + 0.0039697266183793545, + 0.006141059100627899, + 0.015394375659525394, + 0.004092615097761154, + 0.00440633250400424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.4743114411830902 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015933245420456 + ], + "opposite_logprob_tokens": [ + -1.701593279838562 + ], + "logprob_reward": -0.2015933245420456, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015933245420456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08548344671726227, + 0.1303885281085968, + 0.05303175374865532, + 0.053832877427339554, + 0.039188627153635025, + 0.46490779519081116, + 0.0415758416056633, + 0.08486642688512802, + 0.018606023862957954, + 0.028118640184402466 + ], + "credit_map": [ + 0.014006244018673897, + 0.008350520394742489, + 0.009662413969635963, + 0.010362761095166206, + 0.008975001983344555, + 0.011101716198027134, + 0.010303416289389133, + 0.012721378356218338, + 0.02045738883316517, + 0.00951642356812954, + 0.01455837581306696, + 0.011788325384259224, + 0.01915868930518627, + 0.02191745862364769, + 0.013501663692295551, + 0.019490212202072144, + 0.005041159223765135, + 0.004025905858725309, + 0.004268697928637266, + 0.0047978367656469345, + 0.01642194576561451, + 0.005720006301999092, + 0.007924199104309082, + 0.004832008387893438, + 0.009077789261937141, + 0.003913513384759426, + 0.0064318012446165085, + 0.006825692486017942, + 0.006482085213065147, + 0.007999946363270283, + 0.0053102001547813416, + 0.0077918535098433495, + 0.004867044277489185, + 0.0031037013977766037, + 0.004390818532556295, + 0.004013766534626484, + 0.0062007480300962925, + 0.005190111696720123, + 0.005106049589812756, + 0.0063163889572024345, + 0.025773711502552032, + 0.010089388117194176, + 0.16308830678462982, + 0.021562011912465096, + 0.015847116708755493, + 0.16740958392620087, + 0.030952779576182365, + 0.030184928327798843, + 0.002677495125681162, + 0.0012953280238434672, + 0.004649242386221886, + 0.001275738817639649, + 0.0017473406624048948, + 0.01807573437690735, + 0.0016591588500887156, + 0.010195807553827763, + 0.004267245065420866, + 0.002289379946887493, + 0.007793772034347057, + 0.0025683268904685974, + 0.003858541138470173, + 0.03763703256845474, + 0.0028802286833524704, + 0.023571902886033058, + 0.002548057585954666, + 0.0011399632785469294, + 0.005133673548698425, + 0.001858361647464335, + 0.0015265215188264847, + 0.0029303831979632378, + 0.0017610351787880063, + 0.0017080288380384445, + 0.0033284786622971296, + 0.0027489005587995052, + 0.004828665871173143, + 0.002920702565461397, + 0.004721620120108128, + 0.002808045595884323, + 0.003036390757188201, + 0.003725838614627719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.2015933245420456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in stunning 8K Ultra HD with a macro lens, this extreme close-up remains perfectly still, focusing on the human eye with razor-sharp clarity. The iris is a complex galaxy of deep hazel, flecked with subtle amber, its intricate fibrous pattern rendered in surreal detail, while the slightly dilated pupil holds a soft spotlight, suggesting a dimly lit, intimate space. The peripheral retina is faintly tinged with blood-red, a delicate network of red veins betraying a deep, unspoken sorrow. Slowly, a crystal tear inexorably pours out, gathering along the lower waterline until surface tension creates a shimmering, convex pool of liquid, magnifying the thin, black eyelashes beneath. For a breathtaking moment, the tear maintains its form, trembling at the edge of the lower eyelid. Then, with a slow, inescapable force of gravity, it spills over the precipice, a perfect droplet tracing a shimmering, silent path along the skin. The entire poignant micro-drama unfolds amidst an atmosphere of profound intimacy and heartbreaking melancholy.", + "seed": 22719, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_5/index_1538_seed_22719_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": -0.08088615778833627, + "vlm_logprob": -0.08088615778833627, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.08088615778833627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0049993982538580894 + ], + "opposite_logprob_tokens": [ + -6.004999160766602 + ], + "logprob_reward": -0.0049993982538580894, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0049993982538580894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03774689882993698, + 0.10211681574583054, + 0.15245585143566132, + 0.10420544445514679, + 0.29138317704200745, + 0.1867879033088684, + 0.03529050573706627, + 0.016094738617539406, + 0.017323285341262817, + 0.05659537389874458 + ], + "credit_map": [ + 0.005273306742310524, + 0.0029576155357062817, + 0.004474993795156479, + 0.004411306232213974, + 0.003756376216188073, + 0.003404422430321574, + 0.004786804784089327, + 0.008682074956595898, + 0.013238376006484032, + 0.009438457898795605, + 0.010513581335544586, + 0.007235164754092693, + 0.017566461116075516, + 0.010256895795464516, + 0.01605120301246643, + 0.01781667396426201, + 0.022056905552744865, + 0.014414583332836628, + 0.018579818308353424, + 0.014253808185458183, + 0.024343904107809067, + 0.016972878947854042, + 0.01748523861169815, + 0.02434871345758438, + 0.009948636405169964, + 0.010960609652101994, + 0.010033379308879375, + 0.01444848533719778, + 0.014741401188075542, + 0.008914717473089695, + 0.01713176630437374, + 0.018026452511548996, + 0.029175981879234314, + 0.028401901945471764, + 0.029912052676081657, + 0.04059535637497902, + 0.03841477632522583, + 0.023188767954707146, + 0.06678485125303268, + 0.03490949794650078, + 0.018403422087430954, + 0.017784133553504944, + 0.024930015206336975, + 0.043411046266555786, + 0.02528383396565914, + 0.015432714484632015, + 0.024399902671575546, + 0.017142847180366516, + 0.0046541946940124035, + 0.003739954438060522, + 0.0038550642784684896, + 0.002264919923618436, + 0.005887460429221392, + 0.003644954413175583, + 0.0035270326770842075, + 0.007716925349086523, + 0.002848248463124037, + 0.0010814913548529148, + 0.0021384262945502996, + 0.0014015069464221597, + 0.0016402831533923745, + 0.0017676908755674958, + 0.001281236414797604, + 0.003935855347663164, + 0.0024399789981544018, + 0.001686674659140408, + 0.003024828853085637, + 0.0017147889593616128, + 0.002257408807054162, + 0.0016055015148594975, + 0.0015007230686023831, + 0.003093381179496646, + 0.004635270219296217, + 0.0018854057416319847, + 0.017449725419282913, + 0.004305203445255756, + 0.0035431706346571445, + 0.00351777789182961, + 0.0028491499833762646, + 0.018409669399261475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.0049993982538580894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011098556220531464 + ], + "opposite_logprob_tokens": [ + -4.511098384857178 + ], + "logprob_reward": -0.011098556220531464, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011098556220531464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08017025887966156, + 0.2511206567287445, + 0.0657632052898407, + 0.050354257225990295, + 0.3061869144439697, + 0.14342795312404633, + 0.022228799760341644, + 0.035236526280641556, + 0.010713942348957062, + 0.034797441214323044 + ], + "credit_map": [ + 0.011593693867325783, + 0.00452269334346056, + 0.009365764446556568, + 0.007179730571806431, + 0.007480745669454336, + 0.007803747430443764, + 0.012603692710399628, + 0.019620200619101524, + 0.03351452946662903, + 0.026852605864405632, + 0.02445746771991253, + 0.019147250801324844, + 0.04689262434840202, + 0.02584940753877163, + 0.03274860233068466, + 0.04165817052125931, + 0.008679705671966076, + 0.006604354362934828, + 0.00751351285725832, + 0.005628627724945545, + 0.011889558285474777, + 0.007283872924745083, + 0.007791096810251474, + 0.010372483171522617, + 0.006813725456595421, + 0.0038322270847857, + 0.0063652354292571545, + 0.007687009871006012, + 0.005063703283667564, + 0.005133529659360647, + 0.007507480680942535, + 0.007951348088681698, + 0.03011639416217804, + 0.03143666684627533, + 0.03132758289575577, + 0.042503345757722855, + 0.041173454374074936, + 0.023948639631271362, + 0.06821625679731369, + 0.03746459633111954, + 0.014352299273014069, + 0.012969008646905422, + 0.019355520606040955, + 0.03443823382258415, + 0.018685176968574524, + 0.011459054425358772, + 0.01827528513967991, + 0.013893385417759418, + 0.0023957861121743917, + 0.0023875501938164234, + 0.002864275360479951, + 0.0021337487269192934, + 0.0042260754853487015, + 0.002173786051571369, + 0.0032018942292779684, + 0.002845684764906764, + 0.004460766911506653, + 0.0022324593737721443, + 0.004515206906944513, + 0.0038103654514998198, + 0.005649595521390438, + 0.004088648594915867, + 0.003926938399672508, + 0.006552549544721842, + 0.0012161460472270846, + 0.001372158294543624, + 0.001055252505466342, + 0.0010493198642507195, + 0.002485663164407015, + 0.0014305353397503495, + 0.0010242167627438903, + 0.0010806508362293243, + 0.002637539990246296, + 0.0019948347471654415, + 0.007764388341456652, + 0.004853253718465567, + 0.00274518970400095, + 0.00271061179228127, + 0.0036495598033070564, + 0.008442066609859467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.011098556220531464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294946178793907 + ], + "opposite_logprob_tokens": [ + -3.7732949256896973 + ], + "logprob_reward": -0.023294946178793907, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294946178793907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07702212035655975, + 0.050094690173864365, + 0.10709163546562195, + 0.05309537425637245, + 0.3323690891265869, + 0.24063058197498322, + 0.029447486624121666, + 0.036238525062799454, + 0.021840775385499, + 0.05216967687010765 + ], + "credit_map": [ + 0.011370418593287468, + 0.0033698175102472305, + 0.009281198494136333, + 0.007282160688191652, + 0.00525654898956418, + 0.006511777639389038, + 0.012301674112677574, + 0.021648533642292023, + 0.0054417927749454975, + 0.005332845728844404, + 0.005240521859377623, + 0.003657574998214841, + 0.00989010464400053, + 0.005465558730065823, + 0.007479224819689989, + 0.007587067782878876, + 0.014004146680235863, + 0.011833691969513893, + 0.011279481463134289, + 0.008759983815252781, + 0.01992255635559559, + 0.011075466871261597, + 0.013495937921106815, + 0.016720373183488846, + 0.008237329311668873, + 0.0032626602333039045, + 0.0075041623786091805, + 0.00744945602491498, + 0.004704708699136972, + 0.005359691102057695, + 0.007091345731168985, + 0.009486015886068344, + 0.033242519944906235, + 0.03436798229813576, + 0.033765532076358795, + 0.04358697310090065, + 0.045320261269807816, + 0.026484251022338867, + 0.07440876215696335, + 0.04119282215833664, + 0.022974330931901932, + 0.020846538245677948, + 0.034937143325805664, + 0.062031883746385574, + 0.027333587408065796, + 0.017598118633031845, + 0.03107132762670517, + 0.023837652057409286, + 0.0041853985749185085, + 0.0024786207359284163, + 0.003144026268273592, + 0.0029778589960187674, + 0.003290527733042836, + 0.0025020665489137173, + 0.006087157875299454, + 0.004781828261911869, + 0.006074896082282066, + 0.0031992215663194656, + 0.004563724156469107, + 0.0023277848958969116, + 0.004904295317828655, + 0.003378876019269228, + 0.003455708734691143, + 0.008334019221365452, + 0.002460223389789462, + 0.001184556051157415, + 0.006203882861882448, + 0.0016243158606812358, + 0.0017982174176722765, + 0.0015392309287562966, + 0.0013593259500339627, + 0.005671022925525904, + 0.004402810242027044, + 0.002237469656392932, + 0.015009012073278427, + 0.004324339330196381, + 0.003448327537626028, + 0.0031250757165253162, + 0.00377813633531332, + 0.01584450714290142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.023294946178793907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520991563796997 + ], + "opposite_logprob_tokens": [ + -1.5020991563796997 + ], + "logprob_reward": -0.2520991563796997, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520991563796997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05082494765520096, + 0.046902526170015335, + 0.15612775087356567, + 0.04074011743068695, + 0.3566856384277344, + 0.30688557028770447, + 0.010995890013873577, + 0.015820078551769257, + 0.007119785062968731, + 0.007897681556642056 + ], + "credit_map": [ + 0.008039158768951893, + 0.0025131250731647015, + 0.006475992035120726, + 0.0049471487291157246, + 0.003965458367019892, + 0.004312876611948013, + 0.006493366323411465, + 0.014077824540436268, + 0.005272481590509415, + 0.004919836297631264, + 0.004839620087295771, + 0.0036429432220757008, + 0.008841006085276604, + 0.004892818629741669, + 0.007426278665661812, + 0.0070675392635166645, + 0.021720603108406067, + 0.01774558238685131, + 0.016560960561037064, + 0.012223764322698116, + 0.029196172952651978, + 0.01628214120864868, + 0.018411926925182343, + 0.023986587300896645, + 0.006478975061327219, + 0.0035848375409841537, + 0.0044153183698654175, + 0.0029809139668941498, + 0.005358064081519842, + 0.0039899093098938465, + 0.004350749310106039, + 0.009581349790096283, + 0.036504656076431274, + 0.03445997089147568, + 0.03582992032170296, + 0.047446783632040024, + 0.04549187049269676, + 0.02796121872961521, + 0.08472508937120438, + 0.044266123324632645, + 0.028721189126372337, + 0.023972751572728157, + 0.04840109869837761, + 0.08746834099292755, + 0.026781504973769188, + 0.01999148353934288, + 0.03806859999895096, + 0.03348059952259064, + 0.0018552284454926848, + 0.0007708558696322143, + 0.0013440417824313045, + 0.0007804006454534829, + 0.0011763240909203887, + 0.0009473716490902007, + 0.0014343536458909512, + 0.00268731452524662, + 0.001908803591504693, + 0.0012843076838180423, + 0.0020207236520946026, + 0.0019002672052010894, + 0.002325737150385976, + 0.0020004138350486755, + 0.002184616168960929, + 0.0021952101960778236, + 0.0011002838145941496, + 0.0007336623384617269, + 0.0009900163859128952, + 0.0009341294644400477, + 0.0009040636941790581, + 0.0007405286305584013, + 0.0007158323423936963, + 0.0010012678103521466, + 0.0012899318244308233, + 0.0006024730973877013, + 0.0014039414236322045, + 0.000891908654011786, + 0.0008077751263044775, + 0.0008232010877691209, + 0.0009600675548426807, + 0.0011183818569406867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.2520991563796997 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293873190879822 + ], + "opposite_logprob_tokens": [ + -2.237938642501831 + ], + "logprob_reward": -0.11293873190879822, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293873190879822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02820153906941414, + 0.052666276693344116, + 0.1125781387090683, + 0.04837442561984062, + 0.3848397731781006, + 0.2662886679172516, + 0.007200810592621565, + 0.019938701763749123, + 0.042399514466524124, + 0.03751214221119881 + ], + "credit_map": [ + 0.0044104536063969135, + 0.0015922993188723922, + 0.0035587535239756107, + 0.0027308445423841476, + 0.00208218302577734, + 0.002398873446509242, + 0.003890755819156766, + 0.007537379395216703, + 0.0060862088575959206, + 0.0058060321025550365, + 0.005299732554703951, + 0.004126341547816992, + 0.010154739022254944, + 0.005544022656977177, + 0.007824869826436043, + 0.007824335247278214, + 0.015662390738725662, + 0.012401402927935123, + 0.012381537817418575, + 0.00898843351751566, + 0.02027870900928974, + 0.011518153361976147, + 0.013890137895941734, + 0.017457377165555954, + 0.008863819763064384, + 0.0026574551593512297, + 0.00664010364562273, + 0.005612869281321764, + 0.004519690293818712, + 0.0050315228290855885, + 0.00530202966183424, + 0.009746934287250042, + 0.037003617733716965, + 0.0379071980714798, + 0.04051033779978752, + 0.0578136183321476, + 0.04978470504283905, + 0.029071634635329247, + 0.0864938348531723, + 0.04625484347343445, + 0.025269126519560814, + 0.02209251932799816, + 0.040370915085077286, + 0.07305800169706345, + 0.026448985561728477, + 0.018118901178240776, + 0.03324791043996811, + 0.027682308107614517, + 0.0012788260355591774, + 0.0003340865659993142, + 0.0009456967818550766, + 0.0005203249747864902, + 0.0006044254987500608, + 0.0008301431662403047, + 0.0007559548248536885, + 0.0019313521916046739, + 0.003450086573138833, + 0.0013428847305476665, + 0.0022558753844350576, + 0.002034715609624982, + 0.0022291315253823996, + 0.002598615363240242, + 0.0021193225402384996, + 0.003908070269972086, + 0.003022765275090933, + 0.0013484357623383403, + 0.013907095417380333, + 0.0022310339845716953, + 0.0026473463512957096, + 0.00246808142401278, + 0.0020188114140182734, + 0.014755941927433014, + 0.00300526712089777, + 0.0011935896473005414, + 0.011495580896735191, + 0.0029494562186300755, + 0.002266679424792528, + 0.0022462892811745405, + 0.002288215095177293, + 0.012067064642906189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.11293873190879822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in stunning 8K Ultra HD with a macro lens, this extreme close-up remains perfectly still, focusing on the human eye with razor-sharp clarity. The iris is a complex galaxy of deep hazel, flecked with subtle amber, its intricate fibrous pattern rendered in surreal detail, while the slightly dilated pupil holds a soft spotlight, suggesting a dimly lit, intimate space. The peripheral retina is faintly tinged with blood-red, a delicate network of red veins betraying a deep, unspoken sorrow. Slowly, a crystal tear inexorably pours out, gathering along the lower waterline until surface tension creates a shimmering, convex pool of liquid, magnifying the thin, black eyelashes beneath. For a breathtaking moment, the tear maintains its form, trembling at the edge of the lower eyelid. Then, with a slow, inescapable force of gravity, it spills over the precipice, a perfect droplet tracing a shimmering, silent path along the skin. The entire poignant micro-drama unfolds amidst an atmosphere of profound intimacy and heartbreaking melancholy.", + "seed": 22720, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_5/index_1538_seed_22720_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": -0.19624498598277568, + "vlm_logprob": -0.19624498598277568, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.19624498598277568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02798420377075672 + ], + "opposite_logprob_tokens": [ + -3.777984142303467 + ], + "logprob_reward": -0.02798420377075672, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02798420377075672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18179239332675934, + 0.19692234694957733, + 0.14548398554325104, + 0.11932604014873505, + 0.08467309921979904, + 0.040532760322093964, + 0.048173002898693085, + 0.05234384536743164, + 0.03711700811982155, + 0.09363555908203125 + ], + "credit_map": [ + 0.011636877432465553, + 0.012872496619820595, + 0.025599678978323936, + 0.013237068429589272, + 0.02948739565908909, + 0.02740384079515934, + 0.041038159281015396, + 0.020516877993941307, + 0.010030283592641354, + 0.010393600910902023, + 0.01717524416744709, + 0.011409544385969639, + 0.023029353469610214, + 0.05248107388615608, + 0.05357625335454941, + 0.01882699877023697, + 0.01046743430197239, + 0.010800711810588837, + 0.011721905320882797, + 0.011644536629319191, + 0.019530141726136208, + 0.040982991456985474, + 0.02008734457194805, + 0.02024892531335354, + 0.013712283223867416, + 0.013271909207105637, + 0.01094826776534319, + 0.01618637703359127, + 0.01588675193488598, + 0.016736099496483803, + 0.015614239498972893, + 0.01697009988129139, + 0.008722197264432907, + 0.010279505513608456, + 0.007812520489096642, + 0.01225278526544571, + 0.011150937527418137, + 0.011215281672775745, + 0.012120179831981659, + 0.011119691655039787, + 0.0036403394769877195, + 0.004921110346913338, + 0.0035742195323109627, + 0.004837729502469301, + 0.005425384268164635, + 0.006039765197783709, + 0.0062278821133077145, + 0.005866331979632378, + 0.0048409318551421165, + 0.003735961392521858, + 0.003291072091087699, + 0.0045712790451943874, + 0.007539384998381138, + 0.004781676456332207, + 0.01423290278762579, + 0.0051797968335449696, + 0.00586652522906661, + 0.006206234451383352, + 0.004620070569217205, + 0.006863162852823734, + 0.007036158815026283, + 0.007096064742654562, + 0.00851464457809925, + 0.006140982266515493, + 0.0029724533669650555, + 0.0026326740626245737, + 0.003123729256913066, + 0.0032647314947098494, + 0.004889371804893017, + 0.0031638769432902336, + 0.011356968432664871, + 0.005713200196623802, + 0.006434101611375809, + 0.007607995066791773, + 0.008109419606626034, + 0.011131919920444489, + 0.012183016166090965, + 0.012526690028607845, + 0.021918198093771935, + 0.013724219985306263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02798420377075672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311500862240791 + ], + "opposite_logprob_tokens": [ + -3.1681149005889893 + ], + "logprob_reward": -0.04311500862240791, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311500862240791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2672666311264038, + 0.06647472083568573, + 0.10723019391298294, + 0.09519772231578827, + 0.03696977719664574, + 0.031958695501089096, + 0.04809344559907913, + 0.13500137627124786, + 0.025085965171456337, + 0.18672142922878265 + ], + "credit_map": [ + 0.015965543687343597, + 0.011234612204134464, + 0.019518697634339333, + 0.01728919893503189, + 0.028617778792977333, + 0.07862226665019989, + 0.07371403276920319, + 0.022304518148303032, + 0.005855354480445385, + 0.0058126854710280895, + 0.007532892748713493, + 0.004988165572285652, + 0.00878059957176447, + 0.017630523070693016, + 0.008289722725749016, + 0.007584788370877504, + 0.006213338579982519, + 0.007929591462016106, + 0.007088412065058947, + 0.008270575664937496, + 0.019102534279227257, + 0.013808031566441059, + 0.03353350982069969, + 0.01128421537578106, + 0.010682492516934872, + 0.00958380475640297, + 0.008401221595704556, + 0.010509178042411804, + 0.013590776361525059, + 0.01095205545425415, + 0.018853049725294113, + 0.012625149451196194, + 0.004978623241186142, + 0.00428896164521575, + 0.0032535779755562544, + 0.0038434260059148073, + 0.005096511449664831, + 0.004248284734785557, + 0.00740324892103672, + 0.003857147181406617, + 0.0030872602947056293, + 0.003674654522910714, + 0.003178805811330676, + 0.0036451120395213366, + 0.0047137741930782795, + 0.003773011500015855, + 0.00613181060180068, + 0.0037542679347097874, + 0.005398280918598175, + 0.006239103619009256, + 0.004113906063139439, + 0.00586478179320693, + 0.006782422307878733, + 0.0061961328610777855, + 0.008559209294617176, + 0.004939616192132235, + 0.011425141245126724, + 0.021268054842948914, + 0.010874123312532902, + 0.018636230379343033, + 0.023791242390871048, + 0.019825685769319534, + 0.018030425533652306, + 0.011150479316711426, + 0.0021564362104982138, + 0.0029099758248776197, + 0.0028563777450472116, + 0.0030880901031196117, + 0.004705633502453566, + 0.002827615709975362, + 0.0036167872603982687, + 0.0029250523075461388, + 0.013783164322376251, + 0.02788466215133667, + 0.01961398310959339, + 0.02254880592226982, + 0.030355723574757576, + 0.02768329717218876, + 0.025149716064333916, + 0.019702086225152016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.04311500862240791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867146909236908 + ], + "opposite_logprob_tokens": [ + -3.0486714839935303 + ], + "logprob_reward": -0.04867146909236908, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867146909236908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.695953369140625, + 0.05484115704894066, + 0.06344688683748245, + 0.03899311646819115, + 0.023447180166840553, + 0.016430802643299103, + 0.04114687815308571, + 0.018223993480205536, + 0.009232391603291035, + 0.038284335285425186 + ], + "credit_map": [ + 0.027293860912322998, + 0.016451576724648476, + 0.033434879034757614, + 0.020389514043927193, + 0.04724700003862381, + 0.2967621684074402, + 0.05403701961040497, + 0.2003372609615326, + 0.0027863739524036646, + 0.0027778479270637035, + 0.00399425346404314, + 0.003433494595810771, + 0.008631233125925064, + 0.008640887215733528, + 0.018742632120847702, + 0.0058344281278550625, + 0.005020770709961653, + 0.005233710166066885, + 0.004940674640238285, + 0.007315151393413544, + 0.010142606683075428, + 0.007121891714632511, + 0.015308010391891003, + 0.008364064618945122, + 0.004346118774265051, + 0.004605975933372974, + 0.0036176054272800684, + 0.005131735000759363, + 0.004998824559152126, + 0.0051109143532812595, + 0.0053854272700846195, + 0.005796509794890881, + 0.0020091389305889606, + 0.003135089296847582, + 0.0019161792006343603, + 0.0028354129754006863, + 0.003299717092886567, + 0.003026676131412387, + 0.0038674292154610157, + 0.0033575345296412706, + 0.0026699858717620373, + 0.0018811837071552873, + 0.0018158600432798266, + 0.0017240672605112195, + 0.0018786948639899492, + 0.0022163190878927708, + 0.002401631558313966, + 0.001843057805672288, + 0.0028550911229103804, + 0.004776547197252512, + 0.003217030316591263, + 0.004854799713939428, + 0.006000495981425047, + 0.006052777636796236, + 0.008122012950479984, + 0.005268115550279617, + 0.0015534493140876293, + 0.002113548805937171, + 0.0014472436159849167, + 0.002738091628998518, + 0.0025971787981688976, + 0.002328727627173066, + 0.003557260148227215, + 0.0018884893506765366, + 0.0011073306668549776, + 0.0008653724216856062, + 0.0011393605964258313, + 0.000950292160268873, + 0.0014285623328760266, + 0.0011064548743888736, + 0.0013583905529230833, + 0.0012766266008839011, + 0.003004328813403845, + 0.003636231878772378, + 0.004175187554210424, + 0.005799715872853994, + 0.004821350332349539, + 0.005302635487169027, + 0.0069519891403615475, + 0.004592892248183489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04867146909236908 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743316173553467 + ], + "opposite_logprob_tokens": [ + -0.9743316173553467 + ], + "logprob_reward": -0.4743316173553467, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743316173553467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3308509886264801, + 0.08716247230768204, + 0.20387035608291626, + 0.05710811913013458, + 0.031082337722182274, + 0.01782444305717945, + 0.059200748801231384, + 0.029590431600809097, + 0.05675003305077553, + 0.1265600472688675 + ], + "credit_map": [ + 0.015318046323955059, + 0.011003674007952213, + 0.027976052835583687, + 0.014597378671169281, + 0.024385564029216766, + 0.12653212249279022, + 0.06374157220125198, + 0.04729658365249634, + 0.004332191776484251, + 0.004578833468258381, + 0.007588991895318031, + 0.006069382186979055, + 0.011984909884631634, + 0.017793195322155952, + 0.025700010359287262, + 0.009114956483244896, + 0.009739582426846027, + 0.011661007069051266, + 0.011821523308753967, + 0.014973988756537437, + 0.03314417228102684, + 0.03771518915891647, + 0.070663221180439, + 0.014151668176054955, + 0.006202593445777893, + 0.0054649533703923225, + 0.0048544821329414845, + 0.006857744418084621, + 0.008438489399850368, + 0.006852208636701107, + 0.0109010124579072, + 0.00753663619980216, + 0.0038509198930114508, + 0.003489747643470764, + 0.002986891195178032, + 0.004032305907458067, + 0.0042380234226584435, + 0.00394990062341094, + 0.0049655865877866745, + 0.0035689626820385456, + 0.0017061616526916623, + 0.0022101907525211573, + 0.0015834961086511612, + 0.0019326910842210054, + 0.002369052730500698, + 0.002351232338696718, + 0.0028916203882545233, + 0.0027799976523965597, + 0.004245679825544357, + 0.00723203644156456, + 0.0049915132112801075, + 0.006759622599929571, + 0.008764592930674553, + 0.008781544864177704, + 0.01149237621575594, + 0.006933380849659443, + 0.002959411358460784, + 0.002542938804253936, + 0.0030439042020589113, + 0.004576107021421194, + 0.0034592540469020605, + 0.004555096384137869, + 0.0044777835719287395, + 0.003975935745984316, + 0.003530307440087199, + 0.004350967239588499, + 0.004547063261270523, + 0.0050789243541657925, + 0.008044226095080376, + 0.008237403817474842, + 0.014122148044407368, + 0.00883899163454771, + 0.014342720620334148, + 0.0157387126237154, + 0.014869226142764091, + 0.014123586006462574, + 0.015002254396677017, + 0.01654816046357155, + 0.016949309036135674, + 0.018986085429787636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.4743316173553467 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 human eye with razor-sharp clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the iris hazel with flecks of amber and a visible fibrous pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the pupil slightly dilated and does the peripheral retina show faint red veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a crystal tear form a convex pool on the lower waterline before spilling over the eyelid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tear trace a shimmering path along the skin after spilling over the eyelid?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38712263107299805 + ], + "opposite_logprob_tokens": [ + -1.137122631072998 + ], + "logprob_reward": -0.38712263107299805, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38712263107299805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25866439938545227, + 0.0844544768333435, + 0.11188954859972, + 0.09289836883544922, + 0.05208431929349899, + 0.030749941244721413, + 0.11474625766277313, + 0.03809530660510063, + 0.07436120510101318, + 0.1420561969280243 + ], + "credit_map": [ + 0.014368869364261627, + 0.012666555121541023, + 0.023537691682577133, + 0.015940535813570023, + 0.035123053938150406, + 0.042639050632715225, + 0.08403356373310089, + 0.030355071648955345, + 0.0038617076352238655, + 0.0040308465249836445, + 0.007086918689310551, + 0.0056947385892271996, + 0.012561298906803131, + 0.014899040572345257, + 0.02767729014158249, + 0.008642633445560932, + 0.006509661208838224, + 0.006822241470217705, + 0.006620202679187059, + 0.007802035193890333, + 0.018039770424365997, + 0.01747872866690159, + 0.037647172808647156, + 0.010969733819365501, + 0.009120362810790539, + 0.010816024616360664, + 0.008135616779327393, + 0.012338971719145775, + 0.0126603739336133, + 0.011581002734601498, + 0.014538806863129139, + 0.013707200065255165, + 0.0037014037370681763, + 0.007284143939614296, + 0.0039013661444187164, + 0.005543699953705072, + 0.007668858859688044, + 0.006658289581537247, + 0.009213192388415337, + 0.008113360963761806, + 0.003459915053099394, + 0.003799094120040536, + 0.0029836934991180897, + 0.003374251304194331, + 0.0033879491966217756, + 0.004245102871209383, + 0.004329868592321873, + 0.005170065443962812, + 0.007851069793105125, + 0.01483149640262127, + 0.009200473316013813, + 0.013839060440659523, + 0.016071267426013947, + 0.017957206815481186, + 0.020592207089066505, + 0.01440347358584404, + 0.0037571664433926344, + 0.003634001361206174, + 0.0038230346981436014, + 0.005989623721688986, + 0.004572450648993254, + 0.006381387822329998, + 0.004660327918827534, + 0.0052773114293813705, + 0.004566288087517023, + 0.005744414869695902, + 0.005555214826017618, + 0.00717574916779995, + 0.0106640150770545, + 0.011539878323674202, + 0.018199210986495018, + 0.010916430503129959, + 0.014795220457017422, + 0.01879412680864334, + 0.016343139111995697, + 0.015135480090975761, + 0.018039433285593987, + 0.019499756395816803, + 0.01850714720785618, + 0.02094189077615738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.38712263107299805 + } + ], + "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/0000122/group_6/index_tensor([2814, 2814])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_tensor([2814, 2814])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..840cf5d2e769d435d9b8ed683c26406b3b70274f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_tensor([2814, 2814])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, shot handheld, creates a constant, slightly shaky feel, creating an unstable and immersive visual experience. The camera focuses on a vibrant red maple leaf suspended in mid-air, its color so saturated it seems to ignite with autumn's final vitality. Warm afternoon sunlight penetrates the leaf from behind, revealing its thin, wispy structure and network of crimson veins, its edges slightly curled from dryness. The background is completely blurred, a soft bokeh of golden and orange specks, suggestive of a forest in late autumn. An invisible breeze sweeps across the leaf, causing it to tremble slightly. Then, starting at its sharp corner, it doesn't shatter, but rather gracefully and gradually disintegrates, its material form transforming into energy. The bright red fiber structure first becomes transparent, and then turns into the first shining point of light. This process quickly spreads to the entire leaf. In an instant, the entire maple leaf decomposes into countless bright, warm, firefly-like golden and red points of light. They form a short, spiraling trajectory in the air, slowly drifting and flickering as the wind blows, and finally disappearing in the bright halo. The whole scene is full of dreamy, poetic and fleeting beauty.", + "seed": 13592, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_2814_seed_13592_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": -0.1065949976677075, + "vlm_logprob": -0.1065949976677075, + "vlm_yes_no_margin": 5.3, + "vlm_token_credit": -0.1065949976677075, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0047973766922950745 + ], + "opposite_logprob_tokens": [ + -6.754797458648682 + ], + "logprob_reward": -0.0047973766922950745, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0047973766922950745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037130750715732574, + 0.020038126036524773, + 0.012316151522099972, + 0.04065721854567528, + 0.025201881304383278, + 0.05822790786623955, + 0.1850220263004303, + 0.03364158049225807, + 0.03611494600772858, + 0.5516494512557983 + ], + "credit_map": [ + 0.0061325980350375175, + 0.005122174974530935, + 0.007935358211398125, + 0.00531239015981555, + 0.0026473787147551775, + 0.0025885803624987602, + 0.0030105612240731716, + 0.004381710663437843, + 0.0036919633857905865, + 0.005541727412492037, + 0.0021570553071796894, + 0.002065124223008752, + 0.0015522363828495145, + 0.001673733931966126, + 0.0016211827751249075, + 0.0017351022688671947, + 0.0015704709803685546, + 0.0011945964070037007, + 0.0011607279302552342, + 0.0011681128526106477, + 0.0011237720027565956, + 0.0025215523783117533, + 0.0011588077759370208, + 0.0024181113112717867, + 0.004795175511389971, + 0.004166486207395792, + 0.00427890382707119, + 0.0035887183621525764, + 0.006438414100557566, + 0.004619022365659475, + 0.007151131983846426, + 0.005619366187602282, + 0.0038171273190528154, + 0.0022126457188278437, + 0.0043424940668046474, + 0.0027802239637821913, + 0.002333875745534897, + 0.005023910664021969, + 0.0017978667747229338, + 0.0028937372844666243, + 0.0036203661002218723, + 0.005335754249244928, + 0.013787011615931988, + 0.004530385136604309, + 0.002190720522776246, + 0.004356491845101118, + 0.0020794502925127745, + 0.02232772670686245, + 0.008765552192926407, + 0.014990177005529404, + 0.04984177649021149, + 0.015143735334277153, + 0.005703667178750038, + 0.009678482078015804, + 0.00903285387903452, + 0.07186578214168549, + 0.0028195332270115614, + 0.0032605116721242666, + 0.004622284788638353, + 0.004950003698468208, + 0.003982958849519491, + 0.004281958099454641, + 0.004930832423269749, + 0.004793497733771801, + 0.0036936441902071238, + 0.004366845823824406, + 0.005053553730249405, + 0.0037957923486828804, + 0.004016576800495386, + 0.005506837274879217, + 0.004190612118691206, + 0.005491083953529596, + 0.10646679997444153, + 0.0374356284737587, + 0.06401639431715012, + 0.0641995444893837, + 0.03506028279662132, + 0.04852647706866264, + 0.034867335110902786, + 0.16107697784900665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.0047973766922950745 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233348608016968 + ], + "opposite_logprob_tokens": [ + -0.8983348608016968 + ], + "logprob_reward": -0.5233348608016968, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233348608016968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05961930751800537, + 0.026565080508589745, + 0.013175630941987038, + 0.014579366892576218, + 0.027307815849781036, + 0.029749831184744835, + 0.14258386194705963, + 0.054496631026268005, + 0.05242322012782097, + 0.5794993042945862 + ], + "credit_map": [ + 0.011003557592630386, + 0.0111863287165761, + 0.006089814938604832, + 0.0069357785396277905, + 0.0049939341843128204, + 0.005127852316945791, + 0.006169501692056656, + 0.008112541399896145, + 0.0033750247675925493, + 0.005571916699409485, + 0.004731860943138599, + 0.003386127995327115, + 0.0018612584099173546, + 0.001975367544218898, + 0.0025695757940411568, + 0.003093946725130081, + 0.0018714925972744823, + 0.0018002528231590986, + 0.0017522048437967896, + 0.0014297559391707182, + 0.0009840034181252122, + 0.002329274546355009, + 0.0009173983708024025, + 0.002091248519718647, + 0.0025132885202765465, + 0.003096130909398198, + 0.001997033366933465, + 0.001074643456377089, + 0.0013907203683629632, + 0.0015189999248832464, + 0.0012782460544258356, + 0.0017103040590882301, + 0.0039001936092972755, + 0.0029514373745769262, + 0.006091886665672064, + 0.002823071787133813, + 0.0021402332931756973, + 0.003022353397682309, + 0.002725923666730523, + 0.003652716986835003, + 0.0026256521232426167, + 0.006900870706886053, + 0.0033250225242227316, + 0.003779999678954482, + 0.002851288765668869, + 0.0035239579156041145, + 0.0025630532763898373, + 0.00417998805642128, + 0.0072416942566633224, + 0.01335956808179617, + 0.0348801463842392, + 0.01129972655326128, + 0.0054633719846606255, + 0.010477060452103615, + 0.009004143998026848, + 0.05085814744234085, + 0.004296167753636837, + 0.005369876511394978, + 0.006029090844094753, + 0.006561899092048407, + 0.009568626992404461, + 0.006177846807986498, + 0.008497430942952633, + 0.007995693944394588, + 0.004488580860197544, + 0.005294693633913994, + 0.00766195310279727, + 0.009048155508935452, + 0.006246128119528294, + 0.007508374750614166, + 0.00655202753841877, + 0.005623307079076767, + 0.10982783138751984, + 0.039588622748851776, + 0.0443473681807518, + 0.08446881920099258, + 0.027835695073008537, + 0.04918324202299118, + 0.032638221979141235, + 0.1916094869375229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.5233348608016968 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011870964663103223 + ], + "opposite_logprob_tokens": [ + -6.751187324523926 + ], + "logprob_reward": -0.0011870964663103223, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011870964663103223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09496905654668808, + 0.0774526298046112, + 0.02256888709962368, + 0.020278997719287872, + 0.029414236545562744, + 0.08271085470914841, + 0.1971641182899475, + 0.05022895336151123, + 0.020220980048179626, + 0.4049912691116333 + ], + "credit_map": [ + 0.007998157292604446, + 0.010075630620121956, + 0.010908277705311775, + 0.00904279574751854, + 0.010529705323278904, + 0.01830374449491501, + 0.016038190573453903, + 0.0120725492015481, + 0.011390602216124535, + 0.012445075437426567, + 0.011681157164275646, + 0.010665270499885082, + 0.006231662351638079, + 0.009004494175314903, + 0.007374968379735947, + 0.00865939725190401, + 0.0025199749507009983, + 0.006636152509599924, + 0.0024855465162545443, + 0.0019241715781390667, + 0.0020618478301912546, + 0.0030183475464582443, + 0.001264619524590671, + 0.002658227691426873, + 0.002673053415492177, + 0.004794224631041288, + 0.0028465017676353455, + 0.0018795982468873262, + 0.002424227073788643, + 0.0016473433934152126, + 0.0018567085498943925, + 0.002157340757548809, + 0.005174939054995775, + 0.0037639224901795387, + 0.004372449591755867, + 0.0032879316713660955, + 0.0022653862833976746, + 0.0053433021530508995, + 0.0017783137736842036, + 0.003427991410717368, + 0.006046187598258257, + 0.00951181910932064, + 0.015122518874704838, + 0.007702195551246405, + 0.0048054419457912445, + 0.008772927336394787, + 0.003958235029131174, + 0.026791520416736603, + 0.009572337381541729, + 0.01623697206377983, + 0.04930368810892105, + 0.016140304505825043, + 0.008263900876045227, + 0.011348235420882702, + 0.013773911632597446, + 0.07252476364374161, + 0.003584078745916486, + 0.005487461108714342, + 0.007637769915163517, + 0.00609537772834301, + 0.004885409958660603, + 0.008170767687261105, + 0.005391027312725782, + 0.00897705927491188, + 0.0016401982866227627, + 0.0016092644073069096, + 0.0041762650944292545, + 0.002106515923514962, + 0.0032901684753596783, + 0.00213219178840518, + 0.0022613389883190393, + 0.003005038248375058, + 0.06051021069288254, + 0.024550290778279305, + 0.03299705311655998, + 0.09626308083534241, + 0.01328327227383852, + 0.029173308983445168, + 0.019017061218619347, + 0.12919697165489197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0011870964663103223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017160940915346146 + ], + "opposite_logprob_tokens": [ + -6.376716136932373 + ], + "logprob_reward": -0.0017160940915346146, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017160940915346146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08488669991493225, + 0.03363191336393356, + 0.028527861461043358, + 0.019160335883498192, + 0.014756528660655022, + 0.0833854079246521, + 0.3122228980064392, + 0.17136980593204498, + 0.034031178802251816, + 0.21802732348442078 + ], + "credit_map": [ + 0.012878318317234516, + 0.008235065266489983, + 0.01849230006337166, + 0.01127657201141119, + 0.005393669940531254, + 0.010525734163820744, + 0.00846373662352562, + 0.009621302597224712, + 0.005722592584788799, + 0.006313348654657602, + 0.003974044229835272, + 0.003819826291874051, + 0.0024235451128333807, + 0.002302810549736023, + 0.004339674487709999, + 0.004736071918159723, + 0.004508297890424728, + 0.004738317336887121, + 0.003410405246540904, + 0.002778369467705488, + 0.002034141682088375, + 0.004622326698154211, + 0.0027184709906578064, + 0.003717533079907298, + 0.0023537450470030308, + 0.0024710304569453, + 0.00217529502697289, + 0.0015711550367996097, + 0.002971542766317725, + 0.0021072449162602425, + 0.003099823836237192, + 0.0024104996118694544, + 0.001774907112121582, + 0.0018478012643754482, + 0.0020808824338018894, + 0.001643214258365333, + 0.001848773448728025, + 0.002239587251096964, + 0.0014582989970222116, + 0.0018630631966516376, + 0.005703454837203026, + 0.014783352613449097, + 0.01370515301823616, + 0.008318360894918442, + 0.005980255082249641, + 0.007292726077139378, + 0.0070647988468408585, + 0.02053731493651867, + 0.015239573083817959, + 0.02362077310681343, + 0.08735834062099457, + 0.023349691182374954, + 0.007624435238540173, + 0.016486939042806625, + 0.010697017423808575, + 0.1278461515903473, + 0.014251595363020897, + 0.01563303731381893, + 0.01828373409807682, + 0.020613716915249825, + 0.027339108288288116, + 0.02092406339943409, + 0.02753959409892559, + 0.02678496763110161, + 0.003896012669429183, + 0.00586222019046545, + 0.004227830097079277, + 0.003217389341443777, + 0.0034898812882602215, + 0.005199620965868235, + 0.0030074373353272676, + 0.005130787380039692, + 0.042341068387031555, + 0.01606687530875206, + 0.019553406164050102, + 0.020338233560323715, + 0.013133839704096317, + 0.02033282443881035, + 0.013007580302655697, + 0.07325349003076553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0017160940915346146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019395602867007256 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.0019395602867007256, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019395602867007256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040437277406454086, + 0.013517019338905811, + 0.024259056895971298, + 0.052791252732276917, + 0.012715628370642662, + 0.0673580914735794, + 0.09105304628610611, + 0.023224281147122383, + 0.04711439833045006, + 0.6275299787521362 + ], + "credit_map": [ + 0.004692905582487583, + 0.004182937555015087, + 0.007258151192218065, + 0.005320204421877861, + 0.00345396320335567, + 0.007162698078900576, + 0.0037497172597795725, + 0.004616699647158384, + 0.0018485005712136626, + 0.0015699415234848857, + 0.0020805022213608027, + 0.0017188071506097913, + 0.0010761698940768838, + 0.0012785729486495256, + 0.0018487117486074567, + 0.0020958136301487684, + 0.0038582554552704096, + 0.006563777104020119, + 0.003128394950181246, + 0.0022377183195203543, + 0.0017000846564769745, + 0.002323912689462304, + 0.0013791363453492522, + 0.0030677770264446735, + 0.006845239549875259, + 0.015538294799625874, + 0.005534091033041477, + 0.004408480599522591, + 0.005881075281649828, + 0.004729235544800758, + 0.0031801906879991293, + 0.006674646865576506, + 0.0016420567408204079, + 0.0012216612230986357, + 0.0019676322117447853, + 0.001274228561669588, + 0.0013280988205224276, + 0.002312441822141409, + 0.001083654584363103, + 0.001885853474959731, + 0.004520792048424482, + 0.00617124093696475, + 0.01421236339956522, + 0.00491131329908967, + 0.003925570752471685, + 0.007703356444835663, + 0.0038901849184185266, + 0.02202327363193035, + 0.004598930943757296, + 0.007385853677988052, + 0.024616437032818794, + 0.00725421542301774, + 0.0032302283216267824, + 0.005258386954665184, + 0.003876541508361697, + 0.03483245149254799, + 0.002327223774045706, + 0.003343333024531603, + 0.0027508484199643135, + 0.003157759318128228, + 0.002652857219800353, + 0.0034655164927244186, + 0.002661836566403508, + 0.00286490679718554, + 0.004415066447108984, + 0.007504571229219437, + 0.005012269131839275, + 0.006881740875542164, + 0.005309634376317263, + 0.006233702413737774, + 0.00524117611348629, + 0.006516238208860159, + 0.10125287622213364, + 0.0390603169798851, + 0.04395964741706848, + 0.09909338504076004, + 0.02831721119582653, + 0.05575314909219742, + 0.033789828419685364, + 0.2263035625219345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.0019395602867007256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot handheld, creates a constant, slightly shaky feel, creating an unstable and immersive visual experience. The camera focuses on a vibrant red maple leaf suspended in mid-air, its color so saturated it seems to ignite with autumn's final vitality. Warm afternoon sunlight penetrates the leaf from behind, revealing its thin, wispy structure and network of crimson veins, its edges slightly curled from dryness. The background is completely blurred, a soft bokeh of golden and orange specks, suggestive of a forest in late autumn. An invisible breeze sweeps across the leaf, causing it to tremble slightly. Then, starting at its sharp corner, it doesn't shatter, but rather gracefully and gradually disintegrates, its material form transforming into energy. The bright red fiber structure first becomes transparent, and then turns into the first shining point of light. This process quickly spreads to the entire leaf. In an instant, the entire maple leaf decomposes into countless bright, warm, firefly-like golden and red points of light. They form a short, spiraling trajectory in the air, slowly drifting and flickering as the wind blows, and finally disappearing in the bright halo. The whole scene is full of dreamy, poetic and fleeting beauty.", + "seed": 13593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_2814_seed_13593_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": -0.13027533276472242, + "vlm_logprob": -0.13027533276472242, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.13027533276472242, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006918048951774836 + ], + "opposite_logprob_tokens": [ + -5.881917953491211 + ], + "logprob_reward": -0.006918048951774836, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006918048951774836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10178316384553909, + 0.09218553453683853, + 0.032940126955509186, + 0.044246286153793335, + 0.03175421804189682, + 0.08741309493780136, + 0.4700169861316681, + 0.05918487161397934, + 0.05861803889274597, + 0.021857701241970062 + ], + "credit_map": [ + 0.011688358150422573, + 0.00820097140967846, + 0.009564056061208248, + 0.012002987787127495, + 0.009869111701846123, + 0.01892564631998539, + 0.014204071834683418, + 0.01732795685529709, + 0.008145112544298172, + 0.008615044876933098, + 0.013105720281600952, + 0.012587142176926136, + 0.010089104995131493, + 0.013233019039034843, + 0.01563495397567749, + 0.010775433853268623, + 0.004676505923271179, + 0.0030651171691715717, + 0.005931271705776453, + 0.003816665383055806, + 0.00340005150064826, + 0.003154675243422389, + 0.00540295522660017, + 0.0034928827080875635, + 0.007514097262173891, + 0.0042526922188699245, + 0.005807272158563137, + 0.005526157096028328, + 0.005374854430556297, + 0.005326044280081987, + 0.005603021010756493, + 0.0048421453684568405, + 0.0032487225253134966, + 0.004875061102211475, + 0.0034027057699859142, + 0.004505849909037352, + 0.004003241658210754, + 0.004363300744444132, + 0.0037072475533932447, + 0.00364808551967144, + 0.008894577622413635, + 0.011818313971161842, + 0.00870964489877224, + 0.009206674993038177, + 0.014246497303247452, + 0.009704980067908764, + 0.012317617423832417, + 0.012514784932136536, + 0.02712281234562397, + 0.0345718115568161, + 0.1374800056219101, + 0.20018188655376434, + 0.012067976407706738, + 0.02387097105383873, + 0.023523878306150436, + 0.011197610758244991, + 0.0056530204601585865, + 0.006869623903185129, + 0.009471979923546314, + 0.012247160077095032, + 0.004274821374565363, + 0.0037309755571186543, + 0.011460020206868649, + 0.005477267783135176, + 0.004820340313017368, + 0.006377881858497858, + 0.00517053110525012, + 0.006400388199836016, + 0.004989455919712782, + 0.010807140730321407, + 0.01075194776058197, + 0.009300349280238152, + 0.0021370933391153812, + 0.0017989337211474776, + 0.002810887061059475, + 0.00350879761390388, + 0.002721362514421344, + 0.002858845517039299, + 0.002854939317330718, + 0.0031668427400290966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.006918048951774836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328172087669373 + ], + "opposite_logprob_tokens": [ + -0.7578172087669373 + ], + "logprob_reward": -0.6328172087669373, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328172087669373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13597406446933746, + 0.29113680124282837, + 0.08094228059053421, + 0.06462492793798447, + 0.07051068544387817, + 0.05660990998148918, + 0.2252279371023178, + 0.029067648574709892, + 0.030722273513674736, + 0.015183448791503906 + ], + "credit_map": [ + 0.01910870335996151, + 0.010868944227695465, + 0.021150272339582443, + 0.013646152801811695, + 0.009414979256689548, + 0.021781355142593384, + 0.017384635284543037, + 0.02261902391910553, + 0.03685544803738594, + 0.024647783488035202, + 0.03298518434166908, + 0.03305675461888313, + 0.02883855625987053, + 0.04736261069774628, + 0.035736482590436935, + 0.051653988659381866, + 0.010117992758750916, + 0.009506270289421082, + 0.008914116770029068, + 0.010568144731223583, + 0.008682089857757092, + 0.008126497268676758, + 0.01326073706150055, + 0.011766433715820312, + 0.007625846657902002, + 0.006863555405288935, + 0.00803704559803009, + 0.008991869166493416, + 0.009044735692441463, + 0.008919673040509224, + 0.010030196979641914, + 0.005112008191645145, + 0.012338732369244099, + 0.007485928479582071, + 0.007892647758126259, + 0.010500643402338028, + 0.006319883745163679, + 0.01034841313958168, + 0.005931188352406025, + 0.009693251922726631, + 0.006930762901902199, + 0.010091128759086132, + 0.005456114653497934, + 0.007588891312479973, + 0.007861888967454433, + 0.006448117550462484, + 0.005903318058699369, + 0.006329688709229231, + 0.015772493556141853, + 0.015251871198415756, + 0.06385526806116104, + 0.0835016742348671, + 0.013639179989695549, + 0.010295525193214417, + 0.013598021119832993, + 0.009313913993537426, + 0.002436887938529253, + 0.003078531939536333, + 0.0038780332542955875, + 0.006959461607038975, + 0.0022261356934905052, + 0.0017951307818293571, + 0.006377774756401777, + 0.0023156937677413225, + 0.0020991708151996136, + 0.003153387689962983, + 0.004039301071316004, + 0.00403214804828167, + 0.002975102048367262, + 0.004341527353972197, + 0.005310804583132267, + 0.004770832136273384, + 0.0013068130938336253, + 0.0015265309484675527, + 0.0016405052738264203, + 0.0022458164021372795, + 0.0022416722495108843, + 0.002103956649079919, + 0.0018070816295221448, + 0.002311073010787368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.6328172087669373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002496223896741867 + ], + "opposite_logprob_tokens": [ + -6.002496242523193 + ], + "logprob_reward": -0.002496223896741867, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002496223896741867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15128140151500702, + 0.11725225299596786, + 0.04228953272104263, + 0.033475324511528015, + 0.0236176960170269, + 0.06682916730642319, + 0.4268668591976166, + 0.04075393080711365, + 0.06340311467647552, + 0.03423076122999191 + ], + "credit_map": [ + 0.01844760589301586, + 0.013473779894411564, + 0.01412192638963461, + 0.019205518066883087, + 0.012759228236973286, + 0.02594776079058647, + 0.01813482493162155, + 0.02919073775410652, + 0.013640557415783405, + 0.007561480160802603, + 0.019765064120292664, + 0.015005608089268208, + 0.010613995604217052, + 0.01600668951869011, + 0.019356619566679, + 0.015302247367799282, + 0.006948385387659073, + 0.005444781854748726, + 0.007619969546794891, + 0.004936465062201023, + 0.002739028772339225, + 0.003870293265208602, + 0.005351454950869083, + 0.005379151552915573, + 0.0036010616458952427, + 0.005400131922215223, + 0.003122161142528057, + 0.005139822605997324, + 0.003947193268686533, + 0.004199131391942501, + 0.004892494063824415, + 0.0031733305659145117, + 0.0026128024328500032, + 0.0026165114250034094, + 0.0025015610735863447, + 0.003196221310645342, + 0.002895795041695237, + 0.0027073232922703028, + 0.0027035835664719343, + 0.004383898340165615, + 0.012408359907567501, + 0.012236641719937325, + 0.007112020626664162, + 0.007601127959787846, + 0.009477818384766579, + 0.005134114995598793, + 0.007456410676240921, + 0.005402674432843924, + 0.024642273783683777, + 0.030688321217894554, + 0.1275395005941391, + 0.1795428991317749, + 0.013774246908724308, + 0.019045913591980934, + 0.019581831991672516, + 0.012051884084939957, + 0.0037774094380438328, + 0.004706073086708784, + 0.0059696584939956665, + 0.0084149818867445, + 0.0035322606563568115, + 0.0030202947091311216, + 0.0076058548875153065, + 0.0037273967172950506, + 0.005205123219639063, + 0.004771105479449034, + 0.006298489402979612, + 0.007198100909590721, + 0.004423931706696749, + 0.010283795185387135, + 0.015154853463172913, + 0.010067717172205448, + 0.002538671949878335, + 0.003299799980595708, + 0.0035659694112837315, + 0.004170290660113096, + 0.0059463209472596645, + 0.004513842985033989, + 0.005506955552846193, + 0.00468891067430377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.002496223896741867 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763415452092886 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.00763415452092886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763415452092886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031688984483480453, + 0.19825123250484467, + 0.039225414395332336, + 0.019832706078886986, + 0.015984848141670227, + 0.029777780175209045, + 0.5922369360923767, + 0.02863926626741886, + 0.03784254193305969, + 0.006520209368318319 + ], + "credit_map": [ + 0.004052063450217247, + 0.003893357701599598, + 0.002896925900131464, + 0.0039909109473228455, + 0.002654668875038624, + 0.004274645820260048, + 0.004552439320832491, + 0.005373971536755562, + 0.025296224281191826, + 0.015936383977532387, + 0.02471908926963806, + 0.02191605418920517, + 0.01738203503191471, + 0.033549852669239044, + 0.024750202894210815, + 0.03470136970281601, + 0.003958441782742739, + 0.002985263941809535, + 0.003698192071169615, + 0.0064331237226724625, + 0.00398374255746603, + 0.0031203373800963163, + 0.011800635606050491, + 0.003245679894462228, + 0.0034346729516983032, + 0.001650257152505219, + 0.0031616168562322855, + 0.0025005715433508158, + 0.002114291535690427, + 0.002925596199929714, + 0.0023438921198248863, + 0.0017018073704093695, + 0.0018368430901318789, + 0.0017303983913734555, + 0.00141476234421134, + 0.0026636479888111353, + 0.0027819410897791386, + 0.0019037250895053148, + 0.0017390353605151176, + 0.0019144945545122027, + 0.005433017387986183, + 0.0038211543578654528, + 0.004036266822367907, + 0.004005094990134239, + 0.004713115282356739, + 0.002511270111426711, + 0.0031217995565384626, + 0.00213606096804142, + 0.031627386808395386, + 0.0451066792011261, + 0.17461952567100525, + 0.2516435980796814, + 0.01559471059590578, + 0.02925868146121502, + 0.027741845697164536, + 0.016644548624753952, + 0.002734504174441099, + 0.003457401180639863, + 0.0029722938779741526, + 0.003959116991609335, + 0.00293133407831192, + 0.0025084938388317823, + 0.007081770338118076, + 0.002994351089000702, + 0.0032741734758019447, + 0.002749586245045066, + 0.003589065745472908, + 0.004177661146968603, + 0.0026669756043702364, + 0.0060841538943350315, + 0.008859951980412006, + 0.006440974306315184, + 0.0005459975218400359, + 0.0006718610529787838, + 0.0007260056445375085, + 0.0007707543554715812, + 0.0009887871565297246, + 0.001031725201755762, + 0.000757251342292875, + 0.001027827151119709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.00763415452092886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001511027687229216 + ], + "opposite_logprob_tokens": [ + -6.501511096954346 + ], + "logprob_reward": -0.001511027687229216, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001511027687229216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09132494777441025, + 0.04822590574622154, + 0.07853452861309052, + 0.04527734965085983, + 0.03566407039761543, + 0.08890905231237411, + 0.46409371495246887, + 0.05695135146379471, + 0.05472864583134651, + 0.03629036992788315 + ], + "credit_map": [ + 0.013409832492470741, + 0.007409109268337488, + 0.012220568023622036, + 0.00912128109484911, + 0.006618541199713945, + 0.016826746985316277, + 0.0092534851282835, + 0.01646539382636547, + 0.005335342604666948, + 0.005057221744209528, + 0.005296862218528986, + 0.005475152283906937, + 0.005838378332555294, + 0.007548835128545761, + 0.006316853687167168, + 0.007357263006269932, + 0.011011962778866291, + 0.0072100223042070866, + 0.01013968512415886, + 0.011124975979328156, + 0.006822755094617605, + 0.006462282966822386, + 0.01728898473083973, + 0.00847385823726654, + 0.008298509754240513, + 0.009224431589245796, + 0.005809590220451355, + 0.005072521977126598, + 0.00326160853728652, + 0.0049644229002296925, + 0.0029448720160871744, + 0.005701398476958275, + 0.003918521571904421, + 0.003986464813351631, + 0.004163528326898813, + 0.007168329320847988, + 0.003977518063038588, + 0.003625238314270973, + 0.003907098434865475, + 0.004917372949421406, + 0.017039915546774864, + 0.013677033595740795, + 0.01226425264030695, + 0.012386362999677658, + 0.010755936615169048, + 0.007603241130709648, + 0.009537829086184502, + 0.0056444802321493626, + 0.02738526463508606, + 0.03400564193725586, + 0.1377539485692978, + 0.1954936683177948, + 0.012896341271698475, + 0.023398365825414658, + 0.020182864740490913, + 0.01297767087817192, + 0.0045400941744446754, + 0.006225156597793102, + 0.00644845562055707, + 0.011052029207348824, + 0.006040174048393965, + 0.00402422621846199, + 0.013347391970455647, + 0.005273828748613596, + 0.0037613892927765846, + 0.003588864579796791, + 0.005497165489941835, + 0.006656779907643795, + 0.004190504085272551, + 0.009210744872689247, + 0.012305333279073238, + 0.009517861530184746, + 0.0028343703597784042, + 0.004730176646262407, + 0.003673403989523649, + 0.0053880405612289906, + 0.005444630980491638, + 0.003855513408780098, + 0.005294549744576216, + 0.005069688893854618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.001511027687229216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot handheld, creates a constant, slightly shaky feel, creating an unstable and immersive visual experience. The camera focuses on a vibrant red maple leaf suspended in mid-air, its color so saturated it seems to ignite with autumn's final vitality. Warm afternoon sunlight penetrates the leaf from behind, revealing its thin, wispy structure and network of crimson veins, its edges slightly curled from dryness. The background is completely blurred, a soft bokeh of golden and orange specks, suggestive of a forest in late autumn. An invisible breeze sweeps across the leaf, causing it to tremble slightly. Then, starting at its sharp corner, it doesn't shatter, but rather gracefully and gradually disintegrates, its material form transforming into energy. The bright red fiber structure first becomes transparent, and then turns into the first shining point of light. This process quickly spreads to the entire leaf. In an instant, the entire maple leaf decomposes into countless bright, warm, firefly-like golden and red points of light. They form a short, spiraling trajectory in the air, slowly drifting and flickering as the wind blows, and finally disappearing in the bright halo. The whole scene is full of dreamy, poetic and fleeting beauty.", + "seed": 13594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_2814_seed_13594_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": -0.04929047445766628, + "vlm_logprob": -0.04929047445766628, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.04929047445766628, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0065897065214812756 + ], + "opposite_logprob_tokens": [ + -6.006589889526367 + ], + "logprob_reward": -0.0065897065214812756, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0065897065214812756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16572339832782745, + 0.06271623820066452, + 0.05119200050830841, + 0.06266667693853378, + 0.08663398027420044, + 0.04046949744224548, + 0.1577189415693283, + 0.1104273870587349, + 0.15339012444019318, + 0.10906179249286652 + ], + "credit_map": [ + 0.017291150987148285, + 0.016716014593839645, + 0.02537045255303383, + 0.015404683537781239, + 0.020254191011190414, + 0.019160693511366844, + 0.029633162543177605, + 0.02189304307103157, + 0.0055287969298660755, + 0.0035894045140594244, + 0.017750917002558708, + 0.006440252531319857, + 0.0035722616594284773, + 0.0049613346345722675, + 0.005569214001297951, + 0.01530405692756176, + 0.006433357018977404, + 0.00642760843038559, + 0.007002794183790684, + 0.00595669262111187, + 0.005133618135005236, + 0.004764471668750048, + 0.006519480608403683, + 0.008953975513577461, + 0.00800886657088995, + 0.005583399441093206, + 0.01329033449292183, + 0.006570691242814064, + 0.0049241385422647, + 0.006767601706087589, + 0.007437419146299362, + 0.010084230452775955, + 0.01356367114931345, + 0.008751440793275833, + 0.010555353946983814, + 0.005282167345285416, + 0.009110158309340477, + 0.009154646657407284, + 0.006752825807780027, + 0.023463714867830276, + 0.004316715523600578, + 0.003994146827608347, + 0.006861921865493059, + 0.0037033495027571917, + 0.0045824917033314705, + 0.005175347905606031, + 0.005289336666464806, + 0.0065461862832307816, + 0.009927763603627682, + 0.005431624595075846, + 0.014261377975344658, + 0.007675780449062586, + 0.007592995185405016, + 0.05471440404653549, + 0.006455488968640566, + 0.05165950953960419, + 0.009351653978228569, + 0.009444408118724823, + 0.00907329935580492, + 0.008817105554044247, + 0.007129475474357605, + 0.02867332100868225, + 0.009376419708132744, + 0.028561707586050034, + 0.02541978284716606, + 0.0181247740983963, + 0.020326096564531326, + 0.016443263739347458, + 0.017674297094345093, + 0.014046561904251575, + 0.015746913850307465, + 0.025608427822589874, + 0.009766059927642345, + 0.011338312178850174, + 0.010463575832545757, + 0.007152417674660683, + 0.012389939278364182, + 0.027582138776779175, + 0.00802113302052021, + 0.022348206490278244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0065897065214812756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557570040225983 + ], + "opposite_logprob_tokens": [ + -1.6005756855010986 + ], + "logprob_reward": -0.22557570040225983, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557570040225983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09034927934408188, + 0.0890391618013382, + 0.08300448209047318, + 0.13009533286094666, + 0.031318046152591705, + 0.05390310660004616, + 0.27711448073387146, + 0.024683436378836632, + 0.13516294956207275, + 0.08532971143722534 + ], + "credit_map": [ + 0.009335952810943127, + 0.012278519570827484, + 0.01258259266614914, + 0.00762039003893733, + 0.00964871421456337, + 0.00835720356553793, + 0.012589684687554836, + 0.01793622598052025, + 0.008177874609827995, + 0.008305362425744534, + 0.020785316824913025, + 0.00914132222533226, + 0.007783147972077131, + 0.007953023537993431, + 0.006606730632483959, + 0.020286383107304573, + 0.00882283877581358, + 0.006159820128232241, + 0.019385667517781258, + 0.006333346478641033, + 0.006510511506348848, + 0.007692921441048384, + 0.007592284586280584, + 0.020507099106907845, + 0.01205362193286419, + 0.007068672217428684, + 0.033652883023023605, + 0.009095845744013786, + 0.01010433305054903, + 0.012088089250028133, + 0.010057088918983936, + 0.035974808037281036, + 0.003679178422316909, + 0.004253033548593521, + 0.003800576087087393, + 0.0034458499867469072, + 0.0032375946175307035, + 0.003762464039027691, + 0.004410817753523588, + 0.0047285291366279125, + 0.008336766622960567, + 0.004407197702676058, + 0.00883456040173769, + 0.004275302868336439, + 0.005107691045850515, + 0.00512883672490716, + 0.004626058042049408, + 0.013186691328883171, + 0.01667029969394207, + 0.004677870310842991, + 0.03153633326292038, + 0.007370810955762863, + 0.007768271490931511, + 0.10891972482204437, + 0.0069277603179216385, + 0.09324339777231216, + 0.0030929988715797663, + 0.0029987129382789135, + 0.0030610926914960146, + 0.0037072072736918926, + 0.0023795089218765497, + 0.00330833881162107, + 0.0030387421138584614, + 0.0030968347564339638, + 0.02337392419576645, + 0.014032427221536636, + 0.014338773675262928, + 0.00948489923030138, + 0.012553093954920769, + 0.020928002893924713, + 0.010350795462727547, + 0.03010104037821293, + 0.008020604960620403, + 0.012880812399089336, + 0.01306716538965702, + 0.011199839413166046, + 0.008618254214525223, + 0.012957832776010036, + 0.006218843162059784, + 0.012366361916065216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.22557570040225983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028246049769222736 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028246049769222736, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028246049769222736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13501255214214325, + 0.1322123259305954, + 0.10079320520162582, + 0.07591605931520462, + 0.05660707503557205, + 0.0379992239177227, + 0.25300952792167664, + 0.052609097212553024, + 0.07569346576929092, + 0.08014751225709915 + ], + "credit_map": [ + 0.020390858873724937, + 0.019568316638469696, + 0.022058580070734024, + 0.010182221420109272, + 0.014551266096532345, + 0.01865660771727562, + 0.014748726040124893, + 0.01485596876591444, + 0.011644931510090828, + 0.009262756444513798, + 0.03772849217057228, + 0.011890221387147903, + 0.007334862370043993, + 0.010479643940925598, + 0.007217708043754101, + 0.03665371984243393, + 0.011891167610883713, + 0.0043834238313138485, + 0.031049808487296104, + 0.006096648517996073, + 0.003257758216932416, + 0.007105047348886728, + 0.004506334196776152, + 0.03250300511717796, + 0.009361416101455688, + 0.008115255273878574, + 0.018680915236473083, + 0.005758014041930437, + 0.00620093010365963, + 0.005718534346669912, + 0.005326963495463133, + 0.0167540255934, + 0.005480749066919088, + 0.0037702855188399553, + 0.0192020982503891, + 0.003408790333196521, + 0.002955838805064559, + 0.010989039205014706, + 0.00225636875256896, + 0.008543905802071095, + 0.003690623678267002, + 0.0034569683484733105, + 0.009610362350940704, + 0.002943887608125806, + 0.0029890520963817835, + 0.0027717871125787497, + 0.0027677493635565042, + 0.009768789634108543, + 0.015105611644685268, + 0.006152331829071045, + 0.0670943334698677, + 0.009375626221299171, + 0.007605802267789841, + 0.10272818803787231, + 0.006269641220569611, + 0.03867797926068306, + 0.00879631470888853, + 0.0035383361391723156, + 0.005261057987809181, + 0.003075728891417384, + 0.0028177921194583178, + 0.010604515671730042, + 0.003863215446472168, + 0.014652135781943798, + 0.01412332896143198, + 0.0076315863989293575, + 0.00898787472397089, + 0.0063048661686480045, + 0.007022487465292215, + 0.00789126381278038, + 0.009512562304735184, + 0.014219493605196476, + 0.009455635212361813, + 0.012858211994171143, + 0.01165842916816473, + 0.011314371600747108, + 0.007712970022112131, + 0.00975396204739809, + 0.007075663655996323, + 0.010318266227841377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0028246049769222736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646076545119286 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008646076545119286, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646076545119286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05139467492699623, + 0.09518460184335709, + 0.07518307864665985, + 0.05028347671031952, + 0.021387111395597458, + 0.024866633117198944, + 0.24198181927204132, + 0.03523940593004227, + 0.2992500960826874, + 0.10522904247045517 + ], + "credit_map": [ + 0.007481178268790245, + 0.004773282445967197, + 0.008857687003910542, + 0.0038885921239852905, + 0.006315670441836119, + 0.008179021067917347, + 0.004569239914417267, + 0.007330004125833511, + 0.007711779326200485, + 0.007335057016462088, + 0.025919031351804733, + 0.0073110503144562244, + 0.006694482639431953, + 0.0072722905315458775, + 0.0061719766817986965, + 0.02676893211901188, + 0.008249747566878796, + 0.003949676640331745, + 0.02120494470000267, + 0.0054415385238826275, + 0.004149177111685276, + 0.005469210911542177, + 0.0053218817338347435, + 0.021396899595856667, + 0.007381842937320471, + 0.005282937549054623, + 0.005304477643221617, + 0.0051148151978850365, + 0.005675078369677067, + 0.006421908736228943, + 0.007761992979794741, + 0.007340422831475735, + 0.0019286242313683033, + 0.0016609149752184749, + 0.004702570382505655, + 0.002602081745862961, + 0.0028540920466184616, + 0.0037733977660536766, + 0.0020032941829413176, + 0.0018621376948431134, + 0.0021470708306878805, + 0.00254953489638865, + 0.005217989441007376, + 0.003302348777651787, + 0.002261590212583542, + 0.0021511598024517298, + 0.001881053321994841, + 0.005355885252356529, + 0.014059118926525116, + 0.004216477740556002, + 0.03794756904244423, + 0.006376749370247126, + 0.006219486705958843, + 0.09767415374517441, + 0.00515918480232358, + 0.07032907754182816, + 0.0071715619415044785, + 0.002372810384258628, + 0.0033315112814307213, + 0.0021001966670155525, + 0.00203639711253345, + 0.004531591199338436, + 0.002703185426071286, + 0.01099215354770422, + 0.0367923267185688, + 0.04359420761466026, + 0.03577110916376114, + 0.03563394770026207, + 0.0345177948474884, + 0.042275890707969666, + 0.02323092520236969, + 0.04743388667702675, + 0.011966400779783726, + 0.017052022740244865, + 0.01614770106971264, + 0.013226822018623352, + 0.015193053521215916, + 0.013285977765917778, + 0.007637003902345896, + 0.010720055550336838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.008646076545119286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816283842548728 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816283842548728, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816283842548728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19951802492141724, + 0.10803317278623581, + 0.06029582396149635, + 0.08405350148677826, + 0.1374112367630005, + 0.04395638406276703, + 0.07429983466863632, + 0.05558522418141365, + 0.11625789850950241, + 0.12058894336223602 + ], + "credit_map": [ + 0.02298852801322937, + 0.02376057393848896, + 0.03067677654325962, + 0.01482449471950531, + 0.015964308753609657, + 0.023920146748423576, + 0.022274959832429886, + 0.045108240097761154, + 0.00960896909236908, + 0.005970847327262163, + 0.02993742562830448, + 0.009342042729258537, + 0.0075196195393800735, + 0.008338939398527145, + 0.005561821162700653, + 0.031753506511449814, + 0.0059616779908537865, + 0.008985788561403751, + 0.008961227722465992, + 0.00742171797901392, + 0.005749111529439688, + 0.005649841856211424, + 0.01049214880913496, + 0.00707431323826313, + 0.011635472066700459, + 0.012237627990543842, + 0.018171347677707672, + 0.0063862851820886135, + 0.006891381926834583, + 0.006804062053561211, + 0.005080078728497028, + 0.016847241669893265, + 0.023716647177934647, + 0.012828395701944828, + 0.015529206022620201, + 0.008641151711344719, + 0.013254338875412941, + 0.017028234899044037, + 0.009219318628311157, + 0.03719394654035568, + 0.003818346420302987, + 0.003658026224002242, + 0.010753372684121132, + 0.0034313106443732977, + 0.004522556904703379, + 0.0038507261779159307, + 0.004378753714263439, + 0.009543290361762047, + 0.006105212029069662, + 0.0044192541390657425, + 0.01245944481343031, + 0.005982834845781326, + 0.006768846418708563, + 0.022613124921917915, + 0.003994750790297985, + 0.011956368573009968, + 0.007272481452673674, + 0.004395151045173407, + 0.006151252426207066, + 0.0034079989418387413, + 0.0030622417107224464, + 0.013582968153059483, + 0.004776723217219114, + 0.012936405837535858, + 0.015007046051323414, + 0.016576142981648445, + 0.014579798094928265, + 0.01325403992086649, + 0.012259243056178093, + 0.013744777999818325, + 0.016753187403082848, + 0.01408365648239851, + 0.012128260917961597, + 0.0116758793592453, + 0.016451455652713776, + 0.008919944055378437, + 0.00911801215261221, + 0.03004666417837143, + 0.008013336919248104, + 0.024235405027866364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.002816283842548728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, shot handheld, creates a constant, slightly shaky feel, creating an unstable and immersive visual experience. The camera focuses on a vibrant red maple leaf suspended in mid-air, its color so saturated it seems to ignite with autumn's final vitality. Warm afternoon sunlight penetrates the leaf from behind, revealing its thin, wispy structure and network of crimson veins, its edges slightly curled from dryness. The background is completely blurred, a soft bokeh of golden and orange specks, suggestive of a forest in late autumn. An invisible breeze sweeps across the leaf, causing it to tremble slightly. Then, starting at its sharp corner, it doesn't shatter, but rather gracefully and gradually disintegrates, its material form transforming into energy. The bright red fiber structure first becomes transparent, and then turns into the first shining point of light. This process quickly spreads to the entire leaf. In an instant, the entire maple leaf decomposes into countless bright, warm, firefly-like golden and red points of light. They form a short, spiraling trajectory in the air, slowly drifting and flickering as the wind blows, and finally disappearing in the bright halo. The whole scene is full of dreamy, poetic and fleeting beauty.", + "seed": 13595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_6/index_2814_seed_13595_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": -0.14109676962252707, + "vlm_logprob": -0.14109676962252707, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.14109676962252707, + "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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041166334412992 + ], + "opposite_logprob_tokens": [ + -7.004116535186768 + ], + "logprob_reward": -0.0041166334412992, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0041166334412992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09958644956350327, + 0.07065541297197342, + 0.05619186908006668, + 0.09627345949411392, + 0.038967110216617584, + 0.13020780682563782, + 0.1437475085258484, + 0.17766128480434418, + 0.11321695148944855, + 0.0734921470284462 + ], + "credit_map": [ + 0.014491051435470581, + 0.012996080331504345, + 0.013581129722297192, + 0.010113099589943886, + 0.014379899017512798, + 0.013209301047027111, + 0.011772969737648964, + 0.009042923338711262, + 0.01121895294636488, + 0.008109875954687595, + 0.012525871396064758, + 0.007087051868438721, + 0.008731471374630928, + 0.007976272143423557, + 0.006577090360224247, + 0.008428828790783882, + 0.007822697050869465, + 0.005139216780662537, + 0.014183701947331429, + 0.005210504867136478, + 0.005692729260772467, + 0.006695126183331013, + 0.0032434347085654736, + 0.00820445828139782, + 0.015370785258710384, + 0.014084638096392155, + 0.015337401069700718, + 0.01345884334295988, + 0.009751449339091778, + 0.010224572382867336, + 0.007684008684009314, + 0.010361764580011368, + 0.005234848707914352, + 0.005463940091431141, + 0.00585173862054944, + 0.005722295958548784, + 0.005226087290793657, + 0.003902474883943796, + 0.003184160217642784, + 0.004381566774100065, + 0.01401964109390974, + 0.016181329265236855, + 0.020649010315537453, + 0.019240546971559525, + 0.02353677898645401, + 0.013247141614556313, + 0.012094465084373951, + 0.011238893494009972, + 0.0157594196498394, + 0.01457008346915245, + 0.03806714341044426, + 0.01031157374382019, + 0.014336459338665009, + 0.018576519563794136, + 0.009694397449493408, + 0.02243191935122013, + 0.020868295803666115, + 0.014981521293520927, + 0.06052994728088379, + 0.010741973295807838, + 0.01221909187734127, + 0.014894116669893265, + 0.009494265541434288, + 0.03393206000328064, + 0.01293359324336052, + 0.015875093638896942, + 0.014172891154885292, + 0.015711190178990364, + 0.013887128792703152, + 0.011574465781450272, + 0.019222097471356392, + 0.009840489365160465, + 0.005482700187712908, + 0.007395726628601551, + 0.009869210422039032, + 0.005785264540463686, + 0.010702809318900108, + 0.0063074808567762375, + 0.018316227942705154, + 0.009632728062570095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.0041166334412992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6933774948120117 + ], + "opposite_logprob_tokens": [ + -0.6933774948120117 + ], + "logprob_reward": -0.6933774948120117, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933774948120117, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17319487035274506, + 0.10050153732299805, + 0.05975114926695824, + 0.05213768035173416, + 0.03158414363861084, + 0.09548339247703552, + 0.12473652511835098, + 0.08341854065656662, + 0.1337626576423645, + 0.14542947709560394 + ], + "credit_map": [ + 0.01778734102845192, + 0.023655425757169724, + 0.0231812484562397, + 0.025591939687728882, + 0.027112489566206932, + 0.020421383902430534, + 0.022872494533658028, + 0.01257255021482706, + 0.014219417236745358, + 0.010754546150565147, + 0.01774848997592926, + 0.015465052798390388, + 0.011553400196135044, + 0.009275011718273163, + 0.012241307646036148, + 0.009244317188858986, + 0.006777371745556593, + 0.005009685643017292, + 0.017495637759566307, + 0.005849977023899555, + 0.0050870319828391075, + 0.00720103457570076, + 0.0036266420502215624, + 0.00870377104729414, + 0.005338605958968401, + 0.004381243605166674, + 0.01460026204586029, + 0.00588014954701066, + 0.004511113278567791, + 0.005524346139281988, + 0.0041786180809140205, + 0.007723345886915922, + 0.004569532815366983, + 0.0035757918376475573, + 0.005853353533893824, + 0.0034829818177968264, + 0.0028685294091701508, + 0.00458484934642911, + 0.002528195036575198, + 0.0041209119372069836, + 0.010382530279457569, + 0.008088478818535805, + 0.025941027328372, + 0.008225863799452782, + 0.01048409566283226, + 0.006432855501770973, + 0.007726476993411779, + 0.01820206083357334, + 0.012476598843932152, + 0.015029063448309898, + 0.022798040881752968, + 0.011219635605812073, + 0.016901593655347824, + 0.016483571380376816, + 0.010652863420546055, + 0.01917516626417637, + 0.008089193142950535, + 0.01038297638297081, + 0.01766533963382244, + 0.008861797861754894, + 0.01047577429562807, + 0.009472094476222992, + 0.006766360718756914, + 0.01170500461012125, + 0.010241487994790077, + 0.016776708886027336, + 0.024562539532780647, + 0.01424181368201971, + 0.01195545494556427, + 0.024062620475888252, + 0.015990175306797028, + 0.015931854024529457, + 0.007120651658624411, + 0.014579743146896362, + 0.019627120345830917, + 0.012904352508485317, + 0.031461361795663834, + 0.0099620521068573, + 0.029518216848373413, + 0.02025597356259823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.6933774948120117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019497923785820603 + ], + "opposite_logprob_tokens": [ + -6.251949787139893 + ], + "logprob_reward": -0.0019497923785820603, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019497923785820603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19951742887496948, + 0.0653696283698082, + 0.10959503799676895, + 0.11987141519784927, + 0.07943665236234665, + 0.05935090780258179, + 0.07804590463638306, + 0.04577017202973366, + 0.050506409257650375, + 0.19253648817539215 + ], + "credit_map": [ + 0.02286696434020996, + 0.0252907183021307, + 0.02020920254290104, + 0.026984035968780518, + 0.033123843371868134, + 0.03334353119134903, + 0.023633722215890884, + 0.014065409079194069, + 0.00758303701877594, + 0.005899390671402216, + 0.013768368400633335, + 0.006486199796199799, + 0.008057212457060814, + 0.006531915161758661, + 0.006402929313480854, + 0.010640573687851429, + 0.01859406754374504, + 0.015263826586306095, + 0.014232913963496685, + 0.0160047747194767, + 0.01288914866745472, + 0.009063614532351494, + 0.011184027418494225, + 0.012362661771476269, + 0.023872030898928642, + 0.0065475828014314175, + 0.04063200578093529, + 0.008585503324866295, + 0.007363817188888788, + 0.009609728120267391, + 0.0047198752872645855, + 0.018540868535637856, + 0.009899947792291641, + 0.010000419802963734, + 0.013347901403903961, + 0.010909334756433964, + 0.007303325459361076, + 0.011059407144784927, + 0.006774893030524254, + 0.010141425766050816, + 0.005124600138515234, + 0.008262901566922665, + 0.008702903054654598, + 0.009972643107175827, + 0.010016275569796562, + 0.005776242818683386, + 0.005182555876672268, + 0.006312781944870949, + 0.008959664031863213, + 0.01128321886062622, + 0.017651276662945747, + 0.006476749666035175, + 0.008166220970451832, + 0.010193023830652237, + 0.006367035675793886, + 0.00894871260970831, + 0.0046584727242589, + 0.0057790097780525684, + 0.009594526141881943, + 0.006260660942643881, + 0.0061264727264642715, + 0.0045026023872196674, + 0.003640789771452546, + 0.005207640118896961, + 0.0034369053319096565, + 0.007453107740730047, + 0.008222050033509731, + 0.004828650038689375, + 0.00575776444748044, + 0.008984181098639965, + 0.005670799873769283, + 0.006152953486889601, + 0.015884215012192726, + 0.015258990228176117, + 0.047451481223106384, + 0.015008612535893917, + 0.03204774484038353, + 0.013332352973520756, + 0.02530888468027115, + 0.028244202956557274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0019497923785820603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100962076336145 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100962076336145, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100962076336145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04769252613186836, + 0.048452410846948624, + 0.023500781506299973, + 0.04166467487812042, + 0.04233216866850853, + 0.10170568525791168, + 0.15150465071201324, + 0.20814859867095947, + 0.20332671701908112, + 0.13167181611061096 + ], + "credit_map": [ + 0.006525316741317511, + 0.007606636732816696, + 0.005847443826496601, + 0.0065177506767213345, + 0.006006027106195688, + 0.005611679051071405, + 0.0051691289991140366, + 0.004408535081893206, + 0.0076398360542953014, + 0.006185136269778013, + 0.0060729230754077435, + 0.006740374490618706, + 0.007024975027889013, + 0.004160583019256592, + 0.005186709575355053, + 0.005441870540380478, + 0.003022376913577318, + 0.0022710252087563276, + 0.003490133211016655, + 0.0033271354623138905, + 0.002595264697447419, + 0.0028689196333289146, + 0.0025982577353715897, + 0.0033276663161814213, + 0.007954675704240799, + 0.004079101141542196, + 0.009349348954856396, + 0.003776698373258114, + 0.0036981564480811357, + 0.005786168854683638, + 0.0025799681898206472, + 0.004440551158040762, + 0.006123718339949846, + 0.004842832684516907, + 0.006822997238487005, + 0.0065685175359249115, + 0.0036704798694700003, + 0.007077279966324568, + 0.003492020070552826, + 0.0037343185395002365, + 0.008818244561553001, + 0.013876927085220814, + 0.014322630129754543, + 0.01414898969233036, + 0.020578639581799507, + 0.010470828972756863, + 0.009744442068040371, + 0.0097449766471982, + 0.015886878594756126, + 0.016799258068203926, + 0.02865874208509922, + 0.014935188926756382, + 0.02060444839298725, + 0.01790311187505722, + 0.013374579139053822, + 0.0233424324542284, + 0.02065962180495262, + 0.026679793372750282, + 0.044982343912124634, + 0.03431273251771927, + 0.026619616895914078, + 0.014830484986305237, + 0.016451075673103333, + 0.02361289970576763, + 0.024854857474565506, + 0.02515675500035286, + 0.02514640800654888, + 0.02661716192960739, + 0.0238514244556427, + 0.03125617280602455, + 0.030713548883795738, + 0.015730371698737144, + 0.009357694536447525, + 0.013684733770787716, + 0.01271724235266447, + 0.019696466624736786, + 0.021579086780548096, + 0.008892235346138477, + 0.029911141842603683, + 0.015833202749490738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.004100962076336145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant red maple leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot handheld with a constant, slightly shaky feel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does warm afternoon sunlight penetrate the leaf from behind, revealing its thin structure and crimson veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the leaf disintegrate into countless bright, warm, firefly-like golden and red points of light that form a short, spiraling trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with a soft bokeh of golden and orange specks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00193896540440619 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.00193896540440619, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00193896540440619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15631729364395142, + 0.06942664831876755, + 0.06306978315114975, + 0.12435805797576904, + 0.11834356933832169, + 0.1457848995923996, + 0.06326985359191895, + 0.08774829655885696, + 0.056820616126060486, + 0.11486103385686874 + ], + "credit_map": [ + 0.029037026688456535, + 0.014508018270134926, + 0.023113632574677467, + 0.011297877877950668, + 0.014259048737585545, + 0.0349273718893528, + 0.01517281774431467, + 0.01400147844105959, + 0.009727252647280693, + 0.006491222884505987, + 0.015261036343872547, + 0.0056116473861038685, + 0.008079757913947105, + 0.00935289729386568, + 0.006882328074425459, + 0.008020498789846897, + 0.010293123312294483, + 0.005134230945259333, + 0.015823697671294212, + 0.006029952317476273, + 0.005077978130429983, + 0.0075933923944830894, + 0.0035520109813660383, + 0.009565395303070545, + 0.016577111557126045, + 0.009745313785970211, + 0.0366494357585907, + 0.010814021341502666, + 0.008586373180150986, + 0.01627812720835209, + 0.006702601443976164, + 0.019005054607987404, + 0.02139914408326149, + 0.007626640144735575, + 0.033767759799957275, + 0.009345526807010174, + 0.006997912190854549, + 0.01849558763206005, + 0.006461767945438623, + 0.014249218627810478, + 0.018609197810292244, + 0.017476551234722137, + 0.026744790375232697, + 0.018035991117358208, + 0.016740648075938225, + 0.01856011524796486, + 0.011350383050739765, + 0.018267212435603142, + 0.008128347806632519, + 0.011498264968395233, + 0.007951895706355572, + 0.0047919731587171555, + 0.008053121156990528, + 0.010477857664227486, + 0.006393177900463343, + 0.0059752073138952255, + 0.00870070792734623, + 0.0063957758247852325, + 0.027373844757676125, + 0.009214752353727818, + 0.007001899182796478, + 0.007758066523820162, + 0.005636262241750956, + 0.01566697470843792, + 0.005313295405358076, + 0.006891345605254173, + 0.010851575061678886, + 0.00721998605877161, + 0.006614869926124811, + 0.004330561030656099, + 0.007593901827931404, + 0.00800507515668869, + 0.007464286405593157, + 0.012048644945025444, + 0.009574474766850471, + 0.01142415776848793, + 0.01825190708041191, + 0.010116782039403915, + 0.03172483667731285, + 0.014255939051508904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.00193896540440619 + } + ], + "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/0000122/group_7/index_tensor([6530, 6530])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_tensor([6530, 6530])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eae76c887868c65e59a1f42f29f39e20146501c1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_tensor([6530, 6530])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a meticulously rendered fantasy animation style, the camera begins with an elegant, fluid, circular motion, circling the summit of a rugged mountain shrouded in dusk. Atop the peak, a massive, anthropomorphic pear reigns supreme, its form imbued with a stately, almost sentient presence. Its skin, a delicately velvety golden-green, is speckled with fine brown speckles. As the camera steadily pans, the pear's skin slowly and methodically unfurls, each petal gracefully curling back like a giant petal. This process reveals its breathtaking core—a vast geode-like structure nestled within countless shimmering, angular crystals. These crystals radiate a soft, ethereal blue-violet glow, filtering through their facets and casting a shifting, prismatic play of light across the unfolding pear's skin and the surrounding ancient, moss-covered rock, capturing every shimmering detail and imbuing the scene with a sense of quiet wonder and awe-inspiring grandeur.", + "seed": 20319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20319_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": -0.15565938828513026, + "vlm_logprob": -0.15565938828513026, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.15565938828513026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16753244400024414 + ], + "opposite_logprob_tokens": [ + -1.9175324440002441 + ], + "logprob_reward": -0.16753244400024414, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16753244400024414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0997798815369606, + 0.13321872055530548, + 0.09195980429649353, + 0.128315269947052, + 0.04102994501590729, + 0.07436338067054749, + 0.07082684338092804, + 0.08354652673006058, + 0.04131333529949188, + 0.2356462925672531 + ], + "credit_map": [ + 0.012164224870502949, + 0.006864877417683601, + 0.008503599092364311, + 0.008300167508423328, + 0.023639019578695297, + 0.021513571962714195, + 0.007507041096687317, + 0.011287381872534752, + 0.017975028604269028, + 0.008467172272503376, + 0.010217087343335152, + 0.010854599066078663, + 0.02943846397101879, + 0.040982116013765335, + 0.008802169002592564, + 0.006482087075710297, + 0.011216076090931892, + 0.01041307020932436, + 0.006957968696951866, + 0.01612512767314911, + 0.015506020747125149, + 0.015751924365758896, + 0.007922428660094738, + 0.00806718785315752, + 0.020706607028841972, + 0.011460762470960617, + 0.013875013217329979, + 0.014056661166250706, + 0.01761675626039505, + 0.015534733422100544, + 0.018228933215141296, + 0.016835805028676987, + 0.004982675891369581, + 0.003871448338031769, + 0.003420391818508506, + 0.004711226560175419, + 0.004678727593272924, + 0.005012502893805504, + 0.0044372533448040485, + 0.009915715083479881, + 0.006622385233640671, + 0.0047978488728404045, + 0.005610489286482334, + 0.0057470956817269325, + 0.005482269451022148, + 0.025868693366646767, + 0.006815530825406313, + 0.013419069349765778, + 0.00887637585401535, + 0.00813897792249918, + 0.007674074731767178, + 0.006734126713126898, + 0.006216887384653091, + 0.014483034610748291, + 0.008235497400164604, + 0.010467873886227608, + 0.008065961301326752, + 0.01890767738223076, + 0.008791693486273289, + 0.008903183043003082, + 0.007379413582384586, + 0.013684622012078762, + 0.005708870943635702, + 0.01210510078817606, + 0.0033697576727718115, + 0.0070160082541406155, + 0.004615709185600281, + 0.0035950015299022198, + 0.007920898497104645, + 0.005059052258729935, + 0.005360887851566076, + 0.004376018885523081, + 0.02896919846534729, + 0.038914863020181656, + 0.0314289890229702, + 0.014100183732807636, + 0.02494947426021099, + 0.041629210114479065, + 0.01851828023791313, + 0.037136100232601166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.16753244400024414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026378588750958443 + ], + "opposite_logprob_tokens": [ + -3.651378631591797 + ], + "logprob_reward": -0.026378588750958443, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026378588750958443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3077066242694855, + 0.10295704007148743, + 0.12710559368133545, + 0.06523134559392929, + 0.02671443112194538, + 0.04026107117533684, + 0.044739920645952225, + 0.07503446936607361, + 0.07537776976823807, + 0.134871706366539 + ], + "credit_map": [ + 0.05612770467996597, + 0.017180170863866806, + 0.026793166995048523, + 0.019344301894307137, + 0.09065630286931992, + 0.05787428468465805, + 0.01629754900932312, + 0.023433150723576546, + 0.018076380714774132, + 0.008107283152639866, + 0.013976404443383217, + 0.008731316775083542, + 0.014650640077888966, + 0.02085014060139656, + 0.008733170107007027, + 0.009831702336668968, + 0.02236863784492016, + 0.01021215133368969, + 0.010771035216748714, + 0.01516779325902462, + 0.01424962654709816, + 0.021523963660001755, + 0.015060787089169025, + 0.017751598730683327, + 0.010291880927979946, + 0.00525212287902832, + 0.0086996890604496, + 0.005974035710096359, + 0.008739850483834743, + 0.006557747256010771, + 0.010034460574388504, + 0.009681561030447483, + 0.0024445659946650267, + 0.002891330514103174, + 0.003976542968302965, + 0.0033033040817826986, + 0.0024241511709988117, + 0.0037010214291512966, + 0.0030311597511172295, + 0.004942356608808041, + 0.0027964364271610975, + 0.004576569423079491, + 0.004113500006496906, + 0.004937981255352497, + 0.004039937164634466, + 0.008694934658706188, + 0.005840161349624395, + 0.00526155810803175, + 0.00386473978869617, + 0.007850988768041134, + 0.004116196185350418, + 0.004924729932099581, + 0.0029030705336481333, + 0.007692158687859774, + 0.005638092756271362, + 0.007749946787953377, + 0.008930832147598267, + 0.012068052776157856, + 0.005922039970755577, + 0.011181104928255081, + 0.008424586616456509, + 0.010880108922719955, + 0.0060816798359155655, + 0.011546077206730843, + 0.0062080733478069305, + 0.01177714578807354, + 0.007078889757394791, + 0.007480166852474213, + 0.014626745134592056, + 0.008413881063461304, + 0.011625182814896107, + 0.008167692460119724, + 0.015344707295298576, + 0.02277962863445282, + 0.013940940611064434, + 0.010668964125216007, + 0.014903709292411804, + 0.02150808274745941, + 0.011986012570559978, + 0.02373967319726944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.026378588750958443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041033364832401276 + ], + "opposite_logprob_tokens": [ + -5.504103183746338 + ], + "logprob_reward": -0.0041033364832401276, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041033364832401276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10880207270383835, + 0.10645408928394318, + 0.07354175299406052, + 0.042631953954696655, + 0.03227094188332558, + 0.0697675421833992, + 0.09841957688331604, + 0.24641519784927368, + 0.13610130548477173, + 0.08559557795524597 + ], + "credit_map": [ + 0.012820185162127018, + 0.01285405084490776, + 0.010809754021465778, + 0.015609306283295155, + 0.012085072696208954, + 0.01437276042997837, + 0.014723080210387707, + 0.015527860261499882, + 0.013559398241341114, + 0.014463645406067371, + 0.01096448302268982, + 0.012721116654574871, + 0.015290186740458012, + 0.02275139093399048, + 0.008567913435399532, + 0.008135943673551083, + 0.011124765500426292, + 0.007068375125527382, + 0.006243726704269648, + 0.010369482450187206, + 0.013579156249761581, + 0.008219712413847446, + 0.006882640067487955, + 0.010053894482553005, + 0.005881982855498791, + 0.005819818004965782, + 0.004489607643336058, + 0.004324193112552166, + 0.005520160775631666, + 0.004939863923937082, + 0.005487531889230013, + 0.0061687929555773735, + 0.0032962486147880554, + 0.002899154322221875, + 0.003867080435156822, + 0.004566225688904524, + 0.004149886779487133, + 0.003852205816656351, + 0.003965160343796015, + 0.005674976855516434, + 0.0051580071449279785, + 0.008872357197105885, + 0.008373451419174671, + 0.007106862962245941, + 0.009234350174665451, + 0.012774575501680374, + 0.008320645429193974, + 0.009927290491759777, + 0.01061407569795847, + 0.01157297845929861, + 0.009294072166085243, + 0.014854907989501953, + 0.0054367003031075, + 0.016952944919466972, + 0.009424276649951935, + 0.020269615575671196, + 0.032566964626312256, + 0.03837041184306145, + 0.021341688930988312, + 0.035675298422575, + 0.029998155310750008, + 0.03144128620624542, + 0.02162865735590458, + 0.03539270907640457, + 0.01762092486023903, + 0.02867862395942211, + 0.010470869019627571, + 0.008122085593640804, + 0.025946972891688347, + 0.019598156213760376, + 0.012343041598796844, + 0.01332063414156437, + 0.010196320712566376, + 0.02194192260503769, + 0.010012982413172722, + 0.005550185684114695, + 0.010788495652377605, + 0.011018702760338783, + 0.007246716879308224, + 0.008840244263410568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0041033364832401276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096213262528181 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096213262528181, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096213262528181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1231410875916481, + 0.051963016390800476, + 0.1476476490497589, + 0.0973401814699173, + 0.09244053810834885, + 0.08030661195516586, + 0.05497434362769127, + 0.19479280710220337, + 0.0673290565609932, + 0.09006471931934357 + ], + "credit_map": [ + 0.02272615022957325, + 0.012268559075891972, + 0.012275592423975468, + 0.00913746003061533, + 0.017986657097935677, + 0.024272918701171875, + 0.009120142087340355, + 0.015353608876466751, + 0.007757536135613918, + 0.007963674142956734, + 0.00521230511367321, + 0.008192883804440498, + 0.005054537206888199, + 0.0064455801621079445, + 0.0067589315585792065, + 0.004577564541250467, + 0.01938484236598015, + 0.007659214548766613, + 0.007754093036055565, + 0.01024644821882248, + 0.04099652171134949, + 0.02937164157629013, + 0.012561958283185959, + 0.019672920927405357, + 0.011088875122368336, + 0.01147843524813652, + 0.00943735707551241, + 0.01000121794641018, + 0.012629779055714607, + 0.014181839302182198, + 0.01405325997620821, + 0.014469409361481667, + 0.010752781294286251, + 0.011358062736690044, + 0.008633915334939957, + 0.010729121044278145, + 0.011218138039112091, + 0.01020140666514635, + 0.009924895130097866, + 0.019622210413217545, + 0.008137259632349014, + 0.00755765475332737, + 0.007054055575281382, + 0.008060187101364136, + 0.006249242462217808, + 0.022530309855937958, + 0.0070536318235099316, + 0.013664266094565392, + 0.005793480202555656, + 0.006196560338139534, + 0.006875978317111731, + 0.006929130759090185, + 0.004213769920170307, + 0.011729312129318714, + 0.0060236514545977116, + 0.007212457247078419, + 0.02946474961936474, + 0.025231245905160904, + 0.01859908737242222, + 0.019499721005558968, + 0.01670936867594719, + 0.03875551372766495, + 0.01630939356982708, + 0.030223723500967026, + 0.006781186442822218, + 0.009719563648104668, + 0.00708384532481432, + 0.005201877560466528, + 0.013640329241752625, + 0.008970808237791061, + 0.006888024974614382, + 0.009043421596288681, + 0.01168232224881649, + 0.018324164673686028, + 0.012642618268728256, + 0.007081262301653624, + 0.010999943129718304, + 0.012141702696681023, + 0.007671038620173931, + 0.0095216641202569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.004096213262528181 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761863589286804 + ], + "opposite_logprob_tokens": [ + -0.8261863589286804 + ], + "logprob_reward": -0.5761863589286804, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761863589286804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17055347561836243, + 0.09821555763483047, + 0.1205340027809143, + 0.09706089645624161, + 0.08176436275243759, + 0.08189743757247925, + 0.05556419864296913, + 0.12757112085819244, + 0.07213473320007324, + 0.09470424056053162 + ], + "credit_map": [ + 0.028534172102808952, + 0.02485000155866146, + 0.01730015128850937, + 0.016186771914362907, + 0.027252594009041786, + 0.0277644544839859, + 0.01488527562469244, + 0.013780052773654461, + 0.012594821862876415, + 0.012009868398308754, + 0.008848138153553009, + 0.012782828882336617, + 0.015991119667887688, + 0.02144342102110386, + 0.008977322839200497, + 0.005568033549934626, + 0.014722523279488087, + 0.014718009158968925, + 0.010062498040497303, + 0.02129547856748104, + 0.01648033782839775, + 0.021553795784711838, + 0.010456169955432415, + 0.011245190165936947, + 0.014517663978040218, + 0.012385883368551731, + 0.010379870422184467, + 0.013900642283260822, + 0.01042250171303749, + 0.00989683996886015, + 0.011537011712789536, + 0.014020483009517193, + 0.010163825936615467, + 0.009764504618942738, + 0.006848109420388937, + 0.009243926033377647, + 0.009700488299131393, + 0.00880621187388897, + 0.008554523810744286, + 0.018682770431041718, + 0.008221294730901718, + 0.00806050468236208, + 0.008641336113214493, + 0.008413433097302914, + 0.007789653725922108, + 0.017804322764277458, + 0.01069337222725153, + 0.01227351650595665, + 0.00594358891248703, + 0.006755874026566744, + 0.006920935586094856, + 0.0062813651748001575, + 0.0038479759823530912, + 0.010136588476598263, + 0.004797373432666063, + 0.010880493558943272, + 0.017346428707242012, + 0.010440696030855179, + 0.01342847105115652, + 0.012426037341356277, + 0.012004153802990913, + 0.025602975860238075, + 0.012011655606329441, + 0.024310704320669174, + 0.009526769630610943, + 0.012103415094316006, + 0.006816096603870392, + 0.0048912446945905685, + 0.012829648330807686, + 0.01067833136767149, + 0.006278769113123417, + 0.009010461159050465, + 0.01092550903558731, + 0.01910395920276642, + 0.012947666458785534, + 0.005376704502850771, + 0.010303236544132233, + 0.015174327418208122, + 0.006621341221034527, + 0.014251497574150562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.5761863589286804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a meticulously rendered fantasy animation style, the camera begins with an elegant, fluid, circular motion, circling the summit of a rugged mountain shrouded in dusk. Atop the peak, a massive, anthropomorphic pear reigns supreme, its form imbued with a stately, almost sentient presence. Its skin, a delicately velvety golden-green, is speckled with fine brown speckles. As the camera steadily pans, the pear's skin slowly and methodically unfurls, each petal gracefully curling back like a giant petal. This process reveals its breathtaking core—a vast geode-like structure nestled within countless shimmering, angular crystals. These crystals radiate a soft, ethereal blue-violet glow, filtering through their facets and casting a shifting, prismatic play of light across the unfolding pear's skin and the surrounding ancient, moss-covered rock, capturing every shimmering detail and imbuing the scene with a sense of quiet wonder and awe-inspiring grandeur.", + "seed": 20320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20320_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": -0.0863558969926089, + "vlm_logprob": -0.0863558969926089, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.0863558969926089, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18433430790901184 + ], + "opposite_logprob_tokens": [ + -1.8093342781066895 + ], + "logprob_reward": -0.18433430790901184, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18433430790901184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20783421397209167, + 0.12112374603748322, + 0.03179538622498512, + 0.12908512353897095, + 0.04045206308364868, + 0.04597977548837662, + 0.057140327990055084, + 0.03928374871611595, + 0.30632391571998596, + 0.020981669425964355 + ], + "credit_map": [ + 0.02112543396651745, + 0.017048319801688194, + 0.01891530677676201, + 0.03453011438250542, + 0.014475801959633827, + 0.03513535112142563, + 0.03025319240987301, + 0.036350708454847336, + 0.007879032753407955, + 0.007989046163856983, + 0.013050009496510029, + 0.01681099273264408, + 0.01581183820962906, + 0.016868475824594498, + 0.026797091588377953, + 0.015917262062430382, + 0.0033920472487807274, + 0.00289544858969748, + 0.004372074268758297, + 0.005386450327932835, + 0.0022396838758140802, + 0.004526262171566486, + 0.003390525234863162, + 0.005592893343418837, + 0.009346728213131428, + 0.020618410781025887, + 0.01013561338186264, + 0.025711942464113235, + 0.012829714454710484, + 0.01989501714706421, + 0.015115219168365002, + 0.015432491898536682, + 0.0039023051504045725, + 0.007486968766897917, + 0.003760925028473139, + 0.005123731680214405, + 0.004739215597510338, + 0.005580976605415344, + 0.004666424822062254, + 0.0051915147341787815, + 0.004188231658190489, + 0.005401620175689459, + 0.006879317574203014, + 0.0076628802344202995, + 0.005175726488232613, + 0.005036745220422745, + 0.00522440392524004, + 0.006410850211977959, + 0.005772897973656654, + 0.006859831511974335, + 0.010562824085354805, + 0.006119357421994209, + 0.008697162382304668, + 0.006339275278151035, + 0.00686472374945879, + 0.0059242588467895985, + 0.003125865012407303, + 0.006114721763879061, + 0.004915244877338409, + 0.005772341508418322, + 0.00682052643969655, + 0.004204676020890474, + 0.0055489870719611645, + 0.0027813916094601154, + 0.024896834045648575, + 0.05198057368397713, + 0.019884487614035606, + 0.025478463619947433, + 0.03775300458073616, + 0.06738130748271942, + 0.029787905514240265, + 0.04916135221719742, + 0.0019095028983429074, + 0.0027233483269810677, + 0.0016086113173514605, + 0.002046039095148444, + 0.0032119224779307842, + 0.002708864165470004, + 0.0033528946805745363, + 0.003420487977564335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.18433430790901184 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209500506520271 + ], + "opposite_logprob_tokens": [ + -2.8120949268341064 + ], + "logprob_reward": -0.06209500506520271, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209500506520271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29347971081733704, + 0.1610388457775116, + 0.09461601078510284, + 0.10252721607685089, + 0.0437537357211113, + 0.025816796347498894, + 0.02244746685028076, + 0.05327022820711136, + 0.18030264973640442, + 0.022747322916984558 + ], + "credit_map": [ + 0.045316487550735474, + 0.018143586814403534, + 0.018167346715927124, + 0.02344329282641411, + 0.02801644429564476, + 0.06819286197423935, + 0.03400525450706482, + 0.05819443240761757, + 0.01863132230937481, + 0.010920248925685883, + 0.014886770397424698, + 0.016375066712498665, + 0.015204024501144886, + 0.027045102789998055, + 0.024679826572537422, + 0.03329649567604065, + 0.010765877552330494, + 0.010824756696820259, + 0.012090054340660572, + 0.011822904460132122, + 0.008310738950967789, + 0.014019136317074299, + 0.00964953564107418, + 0.01713300123810768, + 0.00706594530493021, + 0.015733659267425537, + 0.008496234193444252, + 0.01558252889662981, + 0.010559682734310627, + 0.01576429046690464, + 0.014653060585260391, + 0.014671813696622849, + 0.0043262033723294735, + 0.007887791842222214, + 0.004496838431805372, + 0.0063218423165380955, + 0.005561824422329664, + 0.006219419650733471, + 0.004389502573758364, + 0.004550314974039793, + 0.0038043444510549307, + 0.002937810029834509, + 0.004048313945531845, + 0.003476313780993223, + 0.0024283372331410646, + 0.0038200453855097294, + 0.0025981306098401546, + 0.002703501144424081, + 0.0020833215676248074, + 0.0038307972718030214, + 0.0032183160074055195, + 0.0020115587394684553, + 0.004602985456585884, + 0.002423889935016632, + 0.0022190127056092024, + 0.002057583536952734, + 0.004267939366400242, + 0.009252790361642838, + 0.006884678732603788, + 0.005206680856645107, + 0.012158957310020924, + 0.006580207962542772, + 0.005062001291662455, + 0.0038569739554077387, + 0.014183344319462776, + 0.03026009537279606, + 0.01171645987778902, + 0.015386315993964672, + 0.021915854886174202, + 0.037816036492586136, + 0.0188501738011837, + 0.03017435595393181, + 0.0019556276965886354, + 0.003842550329864025, + 0.0012501846067607403, + 0.0021938516292721033, + 0.002786242635920644, + 0.0038100574165582657, + 0.0034162711817771196, + 0.003492537420243025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.06209500506520271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013450870756059885 + ], + "opposite_logprob_tokens": [ + -6.626345157623291 + ], + "logprob_reward": -0.0013450870756059885, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013450870756059885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16590553522109985, + 0.12543325126171112, + 0.04499564319849014, + 0.06490492075681686, + 0.0933045744895935, + 0.1414639800786972, + 0.07577546685934067, + 0.11628768593072891, + 0.12990547716617584, + 0.04202338680624962 + ], + "credit_map": [ + 0.02008284069597721, + 0.015341429971158504, + 0.014195539988577366, + 0.02492850460112095, + 0.014588302932679653, + 0.032274674624204636, + 0.021867401897907257, + 0.022626861929893494, + 0.010972894728183746, + 0.008886940777301788, + 0.01267001312226057, + 0.017169520258903503, + 0.015463936142623425, + 0.018753690645098686, + 0.023424644023180008, + 0.018091624602675438, + 0.003966513089835644, + 0.004512527491897345, + 0.004472665023058653, + 0.006822718773037195, + 0.005849081557244062, + 0.0066970414482057095, + 0.005092103965580463, + 0.007582994177937508, + 0.005918943788856268, + 0.009895918890833855, + 0.006096215452998877, + 0.01371763925999403, + 0.006931397598236799, + 0.010019908659160137, + 0.006757074501365423, + 0.0055678244680166245, + 0.00840581301599741, + 0.013048043474555016, + 0.009281381964683533, + 0.011685415171086788, + 0.014627712778747082, + 0.013477005995810032, + 0.010972391813993454, + 0.011806818656623363, + 0.023392846807837486, + 0.01383472140878439, + 0.021273160353302956, + 0.020844127982854843, + 0.011911451816558838, + 0.01852351985871792, + 0.017373260110616684, + 0.01431090198457241, + 0.0077734654769301414, + 0.010051502846181393, + 0.013319332152605057, + 0.0096813989803195, + 0.010105307213962078, + 0.007649173028767109, + 0.009588774293661118, + 0.007606521714478731, + 0.010406485758721828, + 0.0181974358856678, + 0.013329644687473774, + 0.01115566398948431, + 0.020929891616106033, + 0.01675419509410858, + 0.014151645824313164, + 0.011362744495272636, + 0.010723561979830265, + 0.02173689566552639, + 0.009529639035463333, + 0.012692938558757305, + 0.015294802375137806, + 0.025517329573631287, + 0.012457999400794506, + 0.021952329203486443, + 0.003323792014271021, + 0.0058252932503819466, + 0.0025773418601602316, + 0.0035686399787664413, + 0.006287065334618092, + 0.00623587379232049, + 0.008870518766343594, + 0.005334867164492607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0013450870756059885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096213262528181 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096213262528181, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096213262528181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19945816695690155, + 0.10390244424343109, + 0.08547918498516083, + 0.10459515452384949, + 0.06651455909013748, + 0.11541003733873367, + 0.04834497720003128, + 0.09540145099163055, + 0.1449889987707138, + 0.03590506315231323 + ], + "credit_map": [ + 0.02868777886033058, + 0.011536991223692894, + 0.012940981425344944, + 0.017082350328564644, + 0.013657906092703342, + 0.05479699373245239, + 0.020294778048992157, + 0.04046037793159485, + 0.007046804763376713, + 0.006991956382989883, + 0.010683134198188782, + 0.012105780653655529, + 0.012370540760457516, + 0.01594035141170025, + 0.02205783687531948, + 0.01670604571700096, + 0.008438777178525925, + 0.008809194900095463, + 0.010211838409304619, + 0.01198909617960453, + 0.007829710841178894, + 0.01122992392629385, + 0.010946976020932198, + 0.01602366380393505, + 0.0104058301076293, + 0.017949683591723442, + 0.009625093080103397, + 0.009790221229195595, + 0.00868187565356493, + 0.017006337642669678, + 0.014134653843939304, + 0.01700146310031414, + 0.006215888075530529, + 0.011910898610949516, + 0.007238849997520447, + 0.009339923039078712, + 0.006582240108400583, + 0.009237481281161308, + 0.006652122363448143, + 0.009337157011032104, + 0.015544540248811245, + 0.014214404858648777, + 0.017786270007491112, + 0.01186505239456892, + 0.0107108224183321, + 0.016282904893159866, + 0.01475136261433363, + 0.01425467524677515, + 0.004909355193376541, + 0.005155019927769899, + 0.0069502065889537334, + 0.0065656085498631, + 0.004948971793055534, + 0.0071232737973332405, + 0.006386784836649895, + 0.006305753719061613, + 0.008965596556663513, + 0.012091660872101784, + 0.011463452130556107, + 0.0129628274589777, + 0.01157013326883316, + 0.015153693966567516, + 0.014327378943562508, + 0.008866703137755394, + 0.011754636652767658, + 0.026565158739686012, + 0.009281682781875134, + 0.010447122156620026, + 0.017990486696362495, + 0.03206941485404968, + 0.013743874616920948, + 0.02313663437962532, + 0.0030907385516911745, + 0.00495006050914526, + 0.0019124652026221156, + 0.003427691524848342, + 0.004734600428491831, + 0.006766754202544689, + 0.005080605857074261, + 0.005942143965512514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.004096213262528181 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799088716506958 + ], + "opposite_logprob_tokens": [ + -1.8049088716506958 + ], + "logprob_reward": -0.1799088716506958, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799088716506958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15290485322475433, + 0.10543990135192871, + 0.05800411477684975, + 0.09675217419862747, + 0.10716213285923004, + 0.06773996353149414, + 0.1507941037416458, + 0.04892456904053688, + 0.19352807104587555, + 0.018750090152025223 + ], + "credit_map": [ + 0.01884448155760765, + 0.013459487818181515, + 0.012902708724141121, + 0.022343605756759644, + 0.013268811628222466, + 0.02522139996290207, + 0.020573409274220467, + 0.02629096992313862, + 0.0073069799691438675, + 0.006849557626992464, + 0.010440172627568245, + 0.013935424387454987, + 0.013735909014940262, + 0.015305564738810062, + 0.022416792809963226, + 0.01544951181858778, + 0.00488940067589283, + 0.0045709796249866486, + 0.005457471590489149, + 0.009020619094371796, + 0.008185185492038727, + 0.008827653713524342, + 0.0089704105630517, + 0.008082401938736439, + 0.007782129105180502, + 0.018271027132868767, + 0.008766738697886467, + 0.010695657692849636, + 0.00868898630142212, + 0.015246433205902576, + 0.01214748714119196, + 0.01515372097492218, + 0.009570829570293427, + 0.02111053839325905, + 0.01029924862086773, + 0.0137979406863451, + 0.013065573759377003, + 0.0164780355989933, + 0.010961206629872322, + 0.011878767982125282, + 0.007552316412329674, + 0.008946768939495087, + 0.00951103400439024, + 0.009403962641954422, + 0.006285852752625942, + 0.009866520762443542, + 0.008851142600178719, + 0.007322374265640974, + 0.01621309109032154, + 0.014894182793796062, + 0.024878239259123802, + 0.0173505749553442, + 0.01806243322789669, + 0.020632339641451836, + 0.02025151625275612, + 0.01851172372698784, + 0.003986695781350136, + 0.008018926717340946, + 0.005511509254574776, + 0.005922218784689903, + 0.009146512486040592, + 0.0055839926935732365, + 0.006339291576296091, + 0.004415426403284073, + 0.015584130771458149, + 0.03275250270962715, + 0.013016710057854652, + 0.017562948167324066, + 0.024549927562475204, + 0.039141833782196045, + 0.019836997613310814, + 0.03108302690088749, + 0.0016069201519712806, + 0.0023422585800290108, + 0.0012450090143829584, + 0.0018975812708958983, + 0.0023726592771708965, + 0.00276876799762249, + 0.0030683099757879972, + 0.0034485862124711275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.1799088716506958 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a meticulously rendered fantasy animation style, the camera begins with an elegant, fluid, circular motion, circling the summit of a rugged mountain shrouded in dusk. Atop the peak, a massive, anthropomorphic pear reigns supreme, its form imbued with a stately, almost sentient presence. Its skin, a delicately velvety golden-green, is speckled with fine brown speckles. As the camera steadily pans, the pear's skin slowly and methodically unfurls, each petal gracefully curling back like a giant petal. This process reveals its breathtaking core—a vast geode-like structure nestled within countless shimmering, angular crystals. These crystals radiate a soft, ethereal blue-violet glow, filtering through their facets and casting a shifting, prismatic play of light across the unfolding pear's skin and the surrounding ancient, moss-covered rock, capturing every shimmering detail and imbuing the scene with a sense of quiet wonder and awe-inspiring grandeur.", + "seed": 20321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20321_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": -0.02490151980891824, + "vlm_logprob": -0.02490151980891824, + "vlm_yes_no_margin": 4.4499999523162845, + "vlm_token_credit": -0.02490151980891824, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020901652052998543 + ], + "opposite_logprob_tokens": [ + -4.270901679992676 + ], + "logprob_reward": -0.020901652052998543, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020901652052998543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27052557468414307, + 0.06796354800462723, + 0.16029377281665802, + 0.13202108442783356, + 0.06428073346614838, + 0.02591615915298462, + 0.03570269048213959, + 0.029286757111549377, + 0.039469003677368164, + 0.1745406836271286 + ], + "credit_map": [ + 0.025733226910233498, + 0.02233421988785267, + 0.027135590091347694, + 0.051186684519052505, + 0.04172232374548912, + 0.024138037115335464, + 0.03941597789525986, + 0.038859494030475616, + 0.008339879103004932, + 0.005773259326815605, + 0.0079203424975276, + 0.011953883804380894, + 0.008638083003461361, + 0.009424795396625996, + 0.007894250564277172, + 0.008019057102501392, + 0.016481034457683563, + 0.0234737079590559, + 0.013203546404838562, + 0.027199717238545418, + 0.018664266914129257, + 0.019794128835201263, + 0.020587118342518806, + 0.020890245214104652, + 0.011671080254018307, + 0.006353853270411491, + 0.014929857105016708, + 0.01686515286564827, + 0.014411918818950653, + 0.024514488875865936, + 0.023401157930493355, + 0.019873572513461113, + 0.010618225671350956, + 0.006064956076443195, + 0.006361919920891523, + 0.007136090192943811, + 0.007988960482180119, + 0.009560426697134972, + 0.008896633982658386, + 0.007653514388948679, + 0.0029811542481184006, + 0.00390442181378603, + 0.002930896822363138, + 0.003106490010395646, + 0.0028233921620994806, + 0.004034834913909435, + 0.002919998252764344, + 0.0032149706967175007, + 0.002828985918313265, + 0.006657496094703674, + 0.003166087670251727, + 0.005862611811608076, + 0.004727949388325214, + 0.003489136230200529, + 0.005059708375483751, + 0.003910713829100132, + 0.0019582693930715322, + 0.003357680980116129, + 0.0031949332915246487, + 0.004712293390184641, + 0.0035003963857889175, + 0.0029571796767413616, + 0.006114496849477291, + 0.0034915073774755, + 0.0023545678704977036, + 0.010699383914470673, + 0.003008697647601366, + 0.004085229244083166, + 0.003244894091039896, + 0.0042670713737607, + 0.004227716010063887, + 0.007581444922834635, + 0.012467291206121445, + 0.009374869056046009, + 0.005588510539382696, + 0.02951367385685444, + 0.013369020074605942, + 0.01169672328978777, + 0.06051329895853996, + 0.03201727569103241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.020901652052998543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820533163845539 + ], + "opposite_logprob_tokens": [ + -4.018205165863037 + ], + "logprob_reward": -0.01820533163845539, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01820533163845539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09098229557275772, + 0.17154842615127563, + 0.11733325570821762, + 0.09162947535514832, + 0.1685916930437088, + 0.07816103845834732, + 0.036019328981637955, + 0.08071286231279373, + 0.06399021297693253, + 0.10103139281272888 + ], + "credit_map": [ + 0.0092738326638937, + 0.006274039391428232, + 0.012807250954210758, + 0.013993743807077408, + 0.011217709630727768, + 0.008373946882784367, + 0.01463183294981718, + 0.014409945346415043, + 0.01754746213555336, + 0.013652478344738483, + 0.019156601279973984, + 0.035012610256671906, + 0.026588909327983856, + 0.015235865488648415, + 0.02142108604311943, + 0.02293342351913452, + 0.017787374556064606, + 0.007764917332679033, + 0.01180820818990469, + 0.014108133502304554, + 0.014221896417438984, + 0.02365105226635933, + 0.011453607119619846, + 0.016538061201572418, + 0.009711426682770252, + 0.0049259657971560955, + 0.009918604046106339, + 0.011533821932971478, + 0.011172972619533539, + 0.009976176545023918, + 0.020315108820796013, + 0.01407539565116167, + 0.026348667219281197, + 0.02040875516831875, + 0.01548423245549202, + 0.0244981050491333, + 0.02391001582145691, + 0.019498098641633987, + 0.019155453890562057, + 0.019288361072540283, + 0.011541188694536686, + 0.006123824045062065, + 0.006280058063566685, + 0.022309569641947746, + 0.007022399455308914, + 0.005587791558355093, + 0.013064267113804817, + 0.0062319450080394745, + 0.003495166776701808, + 0.007034934125840664, + 0.0034076038282364607, + 0.005220715422183275, + 0.004341252613812685, + 0.004260362591594458, + 0.004838562570512295, + 0.0034207291901111603, + 0.004810255952179432, + 0.01167974527925253, + 0.00776903098449111, + 0.022509317845106125, + 0.007873219437897205, + 0.007612584624439478, + 0.011050160974264145, + 0.007408547680824995, + 0.00531815504655242, + 0.010155336931347847, + 0.0030101477168500423, + 0.00510936975479126, + 0.007038445211946964, + 0.008465313352644444, + 0.00838122796267271, + 0.01651221700012684, + 0.0072728656232357025, + 0.00432360265403986, + 0.003379813162609935, + 0.016756871715188026, + 0.004530441015958786, + 0.006821343209594488, + 0.037495844066143036, + 0.020450608804821968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.01820533163845539 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036210936959832907 + ], + "opposite_logprob_tokens": [ + -5.6286211013793945 + ], + "logprob_reward": -0.0036210936959832907, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036210936959832907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12826067209243774, + 0.04189636558294296, + 0.1594519019126892, + 0.11511672288179398, + 0.10228186845779419, + 0.07098338752985, + 0.11714489012956619, + 0.14905047416687012, + 0.04342413321137428, + 0.07238955795764923 + ], + "credit_map": [ + 0.015981758013367653, + 0.008472071960568428, + 0.012928958050906658, + 0.013275898061692715, + 0.01870308443903923, + 0.011050380766391754, + 0.022663621231913567, + 0.02518489956855774, + 0.0046964241191744804, + 0.0025980020873248577, + 0.007309813518077135, + 0.0057209208607673645, + 0.005150971934199333, + 0.004342527128756046, + 0.005336540285497904, + 0.0067411670461297035, + 0.018581723794341087, + 0.020536046475172043, + 0.01653970777988434, + 0.020846901461482048, + 0.020635545253753662, + 0.028195349499583244, + 0.014022847637534142, + 0.02009379118680954, + 0.011558514088392258, + 0.006017148960381746, + 0.013174520805478096, + 0.013774086721241474, + 0.01123746857047081, + 0.022191425785422325, + 0.019935982301831245, + 0.017227575182914734, + 0.015706399455666542, + 0.00980935338884592, + 0.010939130559563637, + 0.013235199265182018, + 0.016265729442238808, + 0.01613205298781395, + 0.011195415630936623, + 0.008998597972095013, + 0.010493883863091469, + 0.006546212825924158, + 0.006595688872039318, + 0.016207536682486534, + 0.006248082499951124, + 0.009953761473298073, + 0.008990971371531487, + 0.005947255529463291, + 0.008999972604215145, + 0.017024312168359756, + 0.011397995986044407, + 0.020778516307473183, + 0.013371710665524006, + 0.012665626592934132, + 0.021356789395213127, + 0.011549977585673332, + 0.006761009804904461, + 0.021070145070552826, + 0.014671175740659237, + 0.04183102399110794, + 0.01467844843864441, + 0.012541234493255615, + 0.024370145052671432, + 0.013127310201525688, + 0.004555278457701206, + 0.004335404839366674, + 0.0045619141310453415, + 0.004817898850888014, + 0.004393124021589756, + 0.005318350624293089, + 0.006916239392012358, + 0.008525925688445568, + 0.005913810804486275, + 0.004712569992989302, + 0.0032641140278428793, + 0.013932484202086926, + 0.007012688089162111, + 0.006703771185129881, + 0.021344536915421486, + 0.009505582973361015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.0036210936959832907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817948115989566 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.002817948115989566, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817948115989566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09452584385871887, + 0.043796565383672714, + 0.04257126897573471, + 0.039325837045907974, + 0.09474210441112518, + 0.09358145296573639, + 0.12340150773525238, + 0.22195135056972504, + 0.1288379430770874, + 0.11726614087820053 + ], + "credit_map": [ + 0.013773423619568348, + 0.006926361937075853, + 0.011357602663338184, + 0.010648570954799652, + 0.012832676991820335, + 0.008879839442670345, + 0.01452481746673584, + 0.015582548454403877, + 0.00544539513066411, + 0.0031844971235841513, + 0.004501599818468094, + 0.008584379218518734, + 0.006863242015242577, + 0.003840208984911442, + 0.00580060388892889, + 0.005576637573540211, + 0.0036941692233085632, + 0.005043808836489916, + 0.005114600993692875, + 0.005286507774144411, + 0.006224931683391333, + 0.0058908709324896336, + 0.00528032798320055, + 0.006036052946001291, + 0.005068994592875242, + 0.002161680720746517, + 0.004120820201933384, + 0.0040586525574326515, + 0.0038166954182088375, + 0.009503338485956192, + 0.004560512490570545, + 0.006035142112523317, + 0.015537600964307785, + 0.0067320032976567745, + 0.011224109679460526, + 0.00949140079319477, + 0.010132305324077606, + 0.020106596872210503, + 0.009632100351154804, + 0.011885983869433403, + 0.011321887373924255, + 0.01091100461781025, + 0.009677919559180737, + 0.0173485167324543, + 0.012466050684452057, + 0.009789605624973774, + 0.015342033468186855, + 0.006724433042109013, + 0.008895539678633213, + 0.020817464217543602, + 0.011077623814344406, + 0.02077009342610836, + 0.014402828179299831, + 0.013344071805477142, + 0.018917730078101158, + 0.015176162123680115, + 0.009050358086824417, + 0.030346814543008804, + 0.022971421480178833, + 0.05641864240169525, + 0.020423904061317444, + 0.021655747666954994, + 0.0385640412569046, + 0.02252041921019554, + 0.0127713643014431, + 0.03488338366150856, + 0.008464338257908821, + 0.011866297572851181, + 0.015559608116745949, + 0.01017169002443552, + 0.020006723701953888, + 0.01511453464627266, + 0.00850529782474041, + 0.009046463295817375, + 0.004931142088025808, + 0.017132259905338287, + 0.01723405346274376, + 0.008914442732930183, + 0.032134637236595154, + 0.019367845728993416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.002817948115989566 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789615735411644 + ], + "opposite_logprob_tokens": [ + -2.5789616107940674 + ], + "logprob_reward": -0.0789615735411644, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789615735411644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07118110358715057, + 0.0771905779838562, + 0.11329897493124008, + 0.048200108110904694, + 0.11169812083244324, + 0.11500319093465805, + 0.09418706595897675, + 0.04252025485038757, + 0.2732349932193756, + 0.05348561704158783 + ], + "credit_map": [ + 0.007062894757837057, + 0.005487383808940649, + 0.007628023624420166, + 0.009506741538643837, + 0.013029719702899456, + 0.007448342628777027, + 0.010130107402801514, + 0.010887895710766315, + 0.008931877091526985, + 0.008562370203435421, + 0.008649928495287895, + 0.014519436284899712, + 0.009605536237359047, + 0.006726289168000221, + 0.009209523908793926, + 0.01098561380058527, + 0.011752401478588581, + 0.008093046024441719, + 0.012112805619835854, + 0.014583771117031574, + 0.012209005653858185, + 0.025977762416005135, + 0.012606285512447357, + 0.015963897109031677, + 0.004601013846695423, + 0.002561893779784441, + 0.005478686187416315, + 0.0065361857414245605, + 0.005680474452674389, + 0.00436506699770689, + 0.011406566947698593, + 0.0075702182948589325, + 0.01730959676206112, + 0.011842477135360241, + 0.011516308411955833, + 0.015538327395915985, + 0.015705745667219162, + 0.01906539686024189, + 0.009470230899751186, + 0.011250034905970097, + 0.015031807124614716, + 0.010933920741081238, + 0.009939865209162235, + 0.03364410623908043, + 0.010890001431107521, + 0.007566291373223066, + 0.019349316135048866, + 0.0076478831470012665, + 0.0073985755443573, + 0.014054759405553341, + 0.011649301275610924, + 0.014848198741674423, + 0.011132659390568733, + 0.009078720584511757, + 0.01546989195048809, + 0.010554969310760498, + 0.002520920243114233, + 0.006055282894521952, + 0.0039653051644563675, + 0.010219157673418522, + 0.004130586516112089, + 0.003825112711638212, + 0.007956432178616524, + 0.0038474570028483868, + 0.014440871775150299, + 0.07433141022920609, + 0.0076588476076722145, + 0.016927281394600868, + 0.015824109315872192, + 0.01719471998512745, + 0.09226065874099731, + 0.03459707275032997, + 0.004064232110977173, + 0.0032919461373239756, + 0.0025510515552014112, + 0.012107162736356258, + 0.003927440848201513, + 0.0026345355436205864, + 0.01994520239531994, + 0.004964047577232122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.0789615735411644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a meticulously rendered fantasy animation style, the camera begins with an elegant, fluid, circular motion, circling the summit of a rugged mountain shrouded in dusk. Atop the peak, a massive, anthropomorphic pear reigns supreme, its form imbued with a stately, almost sentient presence. Its skin, a delicately velvety golden-green, is speckled with fine brown speckles. As the camera steadily pans, the pear's skin slowly and methodically unfurls, each petal gracefully curling back like a giant petal. This process reveals its breathtaking core—a vast geode-like structure nestled within countless shimmering, angular crystals. These crystals radiate a soft, ethereal blue-violet glow, filtering through their facets and casting a shifting, prismatic play of light across the unfolding pear's skin and the surrounding ancient, moss-covered rock, capturing every shimmering detail and imbuing the scene with a sense of quiet wonder and awe-inspiring grandeur.", + "seed": 20322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_7/index_6530_seed_20322_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": -0.08395549384877085, + "vlm_logprob": -0.08395549384877085, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.08395549384877085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12128271907567978 + ], + "opposite_logprob_tokens": [ + -2.2462828159332275 + ], + "logprob_reward": -0.12128271907567978, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12128271907567978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15286977589130402, + 0.039224494248628616, + 0.06280962377786636, + 0.04562169313430786, + 0.05645809322595596, + 0.039614275097846985, + 0.08595689386129379, + 0.1495734602212906, + 0.07662994414567947, + 0.29124170541763306 + ], + "credit_map": [ + 0.011656234040856361, + 0.02294851653277874, + 0.022967621684074402, + 0.021758301183581352, + 0.013723368756473064, + 0.017449727281928062, + 0.018386952579021454, + 0.023979051038622856, + 0.0030658189207315445, + 0.003913304768502712, + 0.005894449073821306, + 0.004402937367558479, + 0.003717652056366205, + 0.006045373622328043, + 0.006352880969643593, + 0.005832078866660595, + 0.005375358741730452, + 0.004129685927182436, + 0.006527837365865707, + 0.00672939931973815, + 0.010353965684771538, + 0.009794999845325947, + 0.007811041083186865, + 0.012087337672710419, + 0.00437758257612586, + 0.0053762272000312805, + 0.00498233363032341, + 0.005762874148786068, + 0.006915312260389328, + 0.006721646524965763, + 0.005150499753654003, + 0.006335215177386999, + 0.005367147736251354, + 0.005929543636739254, + 0.004704893101006746, + 0.0040648384019732475, + 0.008365429006516933, + 0.01500165369361639, + 0.0050285286270082, + 0.007996056228876114, + 0.00536509370431304, + 0.005360778421163559, + 0.004819293972104788, + 0.003951867111027241, + 0.006455251947045326, + 0.005139040295034647, + 0.0034174758475273848, + 0.005105473566800356, + 0.010343934409320354, + 0.0125700943171978, + 0.007268570363521576, + 0.008225863799452782, + 0.011180533096194267, + 0.014337869361042976, + 0.007586201652884483, + 0.01444382593035698, + 0.023677820339798927, + 0.014249582774937153, + 0.025856006890535355, + 0.013751932419836521, + 0.012165650725364685, + 0.021221688017249107, + 0.012021373026072979, + 0.026629405096173286, + 0.012707924470305443, + 0.006032492034137249, + 0.010937198996543884, + 0.005343124736100435, + 0.008528681471943855, + 0.012064059264957905, + 0.007371317129582167, + 0.01364514697343111, + 0.030346449464559555, + 0.0243765190243721, + 0.04770146310329437, + 0.02194656804203987, + 0.024441279470920563, + 0.04568471014499664, + 0.01699131540954113, + 0.07975342869758606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.12128271907567978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310028254985809 + ], + "opposite_logprob_tokens": [ + -3.168100357055664 + ], + "logprob_reward": -0.04310028254985809, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310028254985809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1225053146481514, + 0.11047762632369995, + 0.067192442715168, + 0.07190496474504471, + 0.05206342041492462, + 0.02433798275887966, + 0.3061876893043518, + 0.03652249276638031, + 0.03566855564713478, + 0.17313948273658752 + ], + "credit_map": [ + 0.01015001442283392, + 0.022057490423321724, + 0.01468448806554079, + 0.013839908875524998, + 0.012158853001892567, + 0.014148938469588757, + 0.016084857285022736, + 0.019380776211619377, + 0.01025149691849947, + 0.013739265501499176, + 0.011694769375026226, + 0.017183639109134674, + 0.014218361116945744, + 0.016307368874549866, + 0.011393556371331215, + 0.015689164400100708, + 0.005667856894433498, + 0.009332569316029549, + 0.006171863526105881, + 0.008503120392560959, + 0.00807318463921547, + 0.010637025348842144, + 0.007719789631664753, + 0.011087031103670597, + 0.007121972274035215, + 0.00858151912689209, + 0.010474846698343754, + 0.00946138333529234, + 0.008578685112297535, + 0.010911844670772552, + 0.007654997520148754, + 0.00911971926689148, + 0.0067963930778205395, + 0.004739292431622744, + 0.003753568511456251, + 0.003568437183275819, + 0.0060211955569684505, + 0.015337587334215641, + 0.0033366656862199306, + 0.008510280400514603, + 0.0035671363584697247, + 0.0024292597081512213, + 0.0029134624637663364, + 0.002798340981826186, + 0.003050523577257991, + 0.003329532453790307, + 0.0023994450457394123, + 0.0038502824027091265, + 0.050651341676712036, + 0.042792174965143204, + 0.032973747700452805, + 0.0229082852602005, + 0.040984924882650375, + 0.04138197749853134, + 0.025816500186920166, + 0.04867873340845108, + 0.006945975124835968, + 0.0025879170279949903, + 0.006815659813582897, + 0.0019122635712847114, + 0.002190248342230916, + 0.005787879694253206, + 0.002396258059889078, + 0.007886290550231934, + 0.005430035758763552, + 0.0021996472496539354, + 0.008053545840084553, + 0.0018974470440298319, + 0.0031131834257394075, + 0.006407245062291622, + 0.0030067074112594128, + 0.0055607459507882595, + 0.01602143608033657, + 0.01657351292669773, + 0.0317460261285305, + 0.012693592347204685, + 0.016298238188028336, + 0.0253782756626606, + 0.01182632427662611, + 0.042602069675922394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.04310028254985809 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013446108205243945 + ], + "opposite_logprob_tokens": [ + -6.626344680786133 + ], + "logprob_reward": -0.0013446108205243945, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013446108205243945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08527704328298569, + 0.050114598125219345, + 0.035026565194129944, + 0.07653039693832397, + 0.0582088939845562, + 0.10729259252548218, + 0.3351743519306183, + 0.04508237913250923, + 0.07900666445493698, + 0.12828651070594788 + ], + "credit_map": [ + 0.006835301872342825, + 0.015821507200598717, + 0.01132313720881939, + 0.01077598799020052, + 0.00831435527652502, + 0.010253622196614742, + 0.008802630938589573, + 0.013150502927601337, + 0.006583287380635738, + 0.005975581239908934, + 0.005753257777541876, + 0.006953636649996042, + 0.005915333516895771, + 0.007653341628611088, + 0.006008251570165157, + 0.005271908361464739, + 0.002869802061468363, + 0.003162674605846405, + 0.003572758287191391, + 0.005502284970134497, + 0.004531837068498135, + 0.0048804497346282005, + 0.005753877107053995, + 0.004752877168357372, + 0.006765800062566996, + 0.008657999336719513, + 0.006889687851071358, + 0.00955668929964304, + 0.013511561788618565, + 0.012569032609462738, + 0.008865117095410824, + 0.00971450749784708, + 0.005721000488847494, + 0.007196876220405102, + 0.00536488089710474, + 0.005159804597496986, + 0.007004091050475836, + 0.014661001972854137, + 0.004339850507676601, + 0.008761389181017876, + 0.013826250098645687, + 0.014939422719180584, + 0.01149024348706007, + 0.012588955461978912, + 0.019729342311620712, + 0.013161220587790012, + 0.008354087360203266, + 0.013203072361648083, + 0.05574849247932434, + 0.04564408212900162, + 0.03729623556137085, + 0.025110865011811256, + 0.04399137198925018, + 0.04523501172661781, + 0.027757395058870316, + 0.054390911012887955, + 0.008353890851140022, + 0.004421079531311989, + 0.00764103839173913, + 0.0030474825762212276, + 0.003750068601220846, + 0.006433058995753527, + 0.005108287092298269, + 0.006327473558485508, + 0.010077813640236855, + 0.009273131377995014, + 0.013342387042939663, + 0.0055740089155733585, + 0.009692495688796043, + 0.011529412120580673, + 0.007964346557855606, + 0.011553064920008183, + 0.014160307124257088, + 0.014197137206792831, + 0.02442523092031479, + 0.007935196161270142, + 0.013195939362049103, + 0.01889040134847164, + 0.009710155427455902, + 0.025772128254175186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.0013446108205243945 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019426537910476327 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019426537910476327, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019426537910476327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059363096952438354, + 0.05426967144012451, + 0.021151745691895485, + 0.027174707502126694, + 0.028995037078857422, + 0.04512692242860794, + 0.3885367810726166, + 0.06515645235776901, + 0.058569326996803284, + 0.25165629386901855 + ], + "credit_map": [ + 0.0037785011809319258, + 0.012409538961946964, + 0.005929069593548775, + 0.008519868366420269, + 0.005815721116960049, + 0.007626482751220465, + 0.007481210399419069, + 0.007802708074450493, + 0.00515464274212718, + 0.005502611864358187, + 0.006905361544340849, + 0.009674004279077053, + 0.005808504298329353, + 0.006901695858687162, + 0.007469105068594217, + 0.006853744853287935, + 0.0017798278713598847, + 0.002118317410349846, + 0.003526589134708047, + 0.001971461344510317, + 0.0021533481776714325, + 0.0032313133124262094, + 0.0024410116020590067, + 0.0039298757910728455, + 0.0015813019126653671, + 0.0029519405215978622, + 0.0024464046582579613, + 0.0026068647857755423, + 0.0053183515556156635, + 0.004306957591325045, + 0.003790827700868249, + 0.004172061569988728, + 0.002408016240224242, + 0.0024887043982744217, + 0.0022436389699578285, + 0.0024457864928990602, + 0.003372188424691558, + 0.008644826710224152, + 0.002755767200142145, + 0.004636107012629509, + 0.007622692734003067, + 0.006072402000427246, + 0.004834021907299757, + 0.005944705102592707, + 0.004906798712909222, + 0.0061268252320587635, + 0.003992801997810602, + 0.005626673810184002, + 0.06270092725753784, + 0.05469951778650284, + 0.0409754253923893, + 0.03057447075843811, + 0.052535317838191986, + 0.05276257544755936, + 0.031988728791475296, + 0.06229979917407036, + 0.005971095059067011, + 0.005639484152197838, + 0.01066668052226305, + 0.004899249412119389, + 0.007444368209689856, + 0.008817663416266441, + 0.013103331439197063, + 0.00861458107829094, + 0.00811945740133524, + 0.004016530700027943, + 0.01111559383571148, + 0.003995182923972607, + 0.005632918328046799, + 0.00924489926546812, + 0.007359856273978949, + 0.009084884077310562, + 0.027080334722995758, + 0.0219065360724926, + 0.051269225776195526, + 0.015934444963932037, + 0.022334225475788116, + 0.0359879806637764, + 0.014560863375663757, + 0.06258266419172287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0019426537910476327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 anthropomorphic pear located on the summit of a rugged mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the pear's skin golden-green with fine brown speckles and a velvety texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pear's skin slowly unfurl to reveal a core containing a vast geode-like structure with shimmering angular crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the crystals inside the core radiate a soft blue-violet glow that casts prismatic light on the pear and surrounding moss-covered rocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a fluid circular motion circling the mountain summit?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521072030067444 + ], + "opposite_logprob_tokens": [ + -1.5021071434020996 + ], + "logprob_reward": -0.2521072030067444, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521072030067444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06754353642463684, + 0.10029043257236481, + 0.0434734933078289, + 0.030293915420770645, + 0.03863105550408363, + 0.05111076310276985, + 0.10961192846298218, + 0.040186457335948944, + 0.035535119473934174, + 0.48332321643829346 + ], + "credit_map": [ + 0.003916868474334478, + 0.010503074154257774, + 0.008746342733502388, + 0.007186237256973982, + 0.006309312302619219, + 0.008919927291572094, + 0.008405310101807117, + 0.013556476682424545, + 0.009778274223208427, + 0.010707739740610123, + 0.012851943261921406, + 0.017733322456479073, + 0.010843001306056976, + 0.012674526311457157, + 0.013090092688798904, + 0.012611539103090763, + 0.0029740547761321068, + 0.004403312690556049, + 0.005767041351646185, + 0.006017968524247408, + 0.004294415935873985, + 0.006870579905807972, + 0.004887582268565893, + 0.008258543908596039, + 0.0025868851225823164, + 0.0033378044608980417, + 0.005182357504963875, + 0.0040945555083453655, + 0.003810073249042034, + 0.0037399153225123882, + 0.003533722134307027, + 0.004008605610579252, + 0.0034460697788745165, + 0.004352344200015068, + 0.0030595387797802687, + 0.003960727248340845, + 0.004753128159791231, + 0.010114835575222969, + 0.0032622457947582006, + 0.005682169925421476, + 0.007229797542095184, + 0.004948580637574196, + 0.0048644146881997585, + 0.006078419275581837, + 0.005580398254096508, + 0.010151316411793232, + 0.0048292335122823715, + 0.007428611163049936, + 0.01936432346701622, + 0.01480991393327713, + 0.011572806164622307, + 0.010180295445024967, + 0.013761992566287518, + 0.014291455037891865, + 0.008666383102536201, + 0.016964765265583992, + 0.006682485807687044, + 0.003042304888367653, + 0.007471795659512281, + 0.0030945506878197193, + 0.003627370810136199, + 0.005807959008961916, + 0.0037696727085858583, + 0.006690321955829859, + 0.004694992210716009, + 0.00498352013528347, + 0.005536839831620455, + 0.00247993809171021, + 0.004832381848245859, + 0.004942032508552074, + 0.003792322240769863, + 0.004273099824786186, + 0.048154957592487335, + 0.043536193668842316, + 0.1004905253648758, + 0.0296767707914114, + 0.042434681206941605, + 0.06854049116373062, + 0.028404632583260536, + 0.12208501994609833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.2521072030067444 + } + ], + "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/0000122/group_8/index_tensor([2401, 2401])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_8/index_tensor([2401, 2401])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2729325f6ecf9a0744e8bfafe8cd7fb993175b1b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_8/index_tensor([2401, 2401])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth, slowly rotating close-up, arcing shot, with the intimacy of a home movie, focuses on a dark, solid wood workbench. An old-fashioned desk lamp casts a soft, concentrated halo of warm yellow light, centered on a pair of well-maintained, deft hands with neatly manicured nails. These hands carefully repair a partially disassembled, delicate brass mechanism, its interior comprised of variously sized, gleaming brass gears, tiny springs, and levers. One hand holds the mechanism's main frame steady with its fingertips, while the other, gripping a pair of delicate tweezers, delicately grasps a tiny gear and steadily and deliberately inserts it into a slot within the intricate mechanism. As the camera slowly rotates, the lamplight reflects off the polished surface of the brass gear, creating fluid, dotted highlights and casting complex, ever-shifting shadows across the wooden surface. In the background, the rest of the living room is blurred, and only the outline of the bookshelf and the edge of the soft sofa can be vaguely seen. The whole scene is wrapped in warm colors, full of concentration, tranquility and healing atmosphere.", + "seed": 30237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_8/index_2401_seed_30237_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": -0.0769200336188078, + "vlm_logprob": -0.0769200336188078, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.0769200336188078, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.21094009280204773 + ], + "opposite_logprob_tokens": [ + -1.7109401226043701 + ], + "logprob_reward": -0.21094009280204773, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21094009280204773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28160613775253296, + 0.0297201257199049, + 0.10206286609172821, + 0.033674102276563644, + 0.20693403482437134, + 0.08011899143457413, + 0.07351671904325485, + 0.052909303456544876, + 0.02282756380736828, + 0.1166301742196083 + ], + "credit_map": [ + 0.07604966312646866, + 0.01570005528628826, + 0.08413344621658325, + 0.022346854209899902, + 0.023056425154209137, + 0.02945677377283573, + 0.014659104868769646, + 0.01620381325483322, + 0.0034620838705450296, + 0.0030026466120034456, + 0.003470777301117778, + 0.006083185784518719, + 0.0038829660043120384, + 0.004575910046696663, + 0.0032234331592917442, + 0.0020191201474517584, + 0.014392953366041183, + 0.008764997124671936, + 0.0127760861068964, + 0.020313961431384087, + 0.009651613421738148, + 0.015420752577483654, + 0.011765833012759686, + 0.008976667188107967, + 0.005817237310111523, + 0.0033687788527458906, + 0.004766648635268211, + 0.005393461789935827, + 0.002822295529767871, + 0.0052916440181434155, + 0.0029696703422814608, + 0.0032443637028336525, + 0.04979854077100754, + 0.007591970264911652, + 0.022030333057045937, + 0.016416853293776512, + 0.009424017742276192, + 0.07690423727035522, + 0.00866228062659502, + 0.016105791553854942, + 0.01285178679972887, + 0.007154005579650402, + 0.012993664480745792, + 0.016170110553503036, + 0.0078479228541255, + 0.00912083126604557, + 0.006342090666294098, + 0.00763857364654541, + 0.01868123933672905, + 0.005463349632918835, + 0.017700472846627235, + 0.0075585320591926575, + 0.006641483400017023, + 0.007166455499827862, + 0.0048940167762339115, + 0.005411168094724417, + 0.007029895205050707, + 0.005535502452403307, + 0.006560398265719414, + 0.01074218936264515, + 0.005476361606270075, + 0.006343660410493612, + 0.005508708767592907, + 0.0057125878520309925, + 0.00414632773026824, + 0.001980892615392804, + 0.0022974859457463026, + 0.004857494030147791, + 0.001779354875907302, + 0.002837653737515211, + 0.0018634821753948927, + 0.0030648696701973677, + 0.031779058277606964, + 0.00603864248842001, + 0.038948819041252136, + 0.0064582121558487415, + 0.006950852926820517, + 0.01191437616944313, + 0.00402058893814683, + 0.010519614443182945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.21094009280204773 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.023327438160777092 + ], + "opposite_logprob_tokens": [ + -3.773327350616455 + ], + "logprob_reward": -0.023327438160777092, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023327438160777092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21097996830940247, + 0.15235690772533417, + 0.15847137570381165, + 0.09907040745019913, + 0.11724914610385895, + 0.04959876835346222, + 0.04494590312242508, + 0.05325933173298836, + 0.050085827708244324, + 0.06398236006498337 + ], + "credit_map": [ + 0.04698050022125244, + 0.012243888340890408, + 0.0655263215303421, + 0.02236623875796795, + 0.026346033439040184, + 0.015819355845451355, + 0.01151616033166647, + 0.010181467980146408, + 0.020382139831781387, + 0.012144372798502445, + 0.033141251653432846, + 0.0246448814868927, + 0.023692231625318527, + 0.01845473237335682, + 0.009792783297598362, + 0.010104512795805931, + 0.0189516581594944, + 0.014233281835913658, + 0.022417420521378517, + 0.04314189404249191, + 0.01142271887511015, + 0.016706136986613274, + 0.019364891573786736, + 0.012233370915055275, + 0.012926938943564892, + 0.007967151701450348, + 0.012661919929087162, + 0.020396091043949127, + 0.015328793786466122, + 0.013828018680214882, + 0.010078499093651772, + 0.005882996134459972, + 0.022124620154500008, + 0.009519267827272415, + 0.016762908548116684, + 0.022493984550237656, + 0.010782009921967983, + 0.015513247810304165, + 0.01085924357175827, + 0.00919385813176632, + 0.008190009742975235, + 0.0036272441502660513, + 0.005842185579240322, + 0.008661042898893356, + 0.003422454698011279, + 0.011193866841495037, + 0.003224399173632264, + 0.005437563639134169, + 0.007463059853762388, + 0.004771284759044647, + 0.0046971822157502174, + 0.008657106198370457, + 0.003422411857172847, + 0.007291811518371105, + 0.004059873055666685, + 0.004583172034472227, + 0.006621640641242266, + 0.006665145047008991, + 0.0040204161778092384, + 0.009238154627382755, + 0.006483210250735283, + 0.007652520667761564, + 0.006219033617526293, + 0.0063592116348445415, + 0.006238315720111132, + 0.005087147932499647, + 0.004012978635728359, + 0.009020590223371983, + 0.00613938644528389, + 0.006113023497164249, + 0.006744135171175003, + 0.00673025194555521, + 0.004512697923928499, + 0.0036757111083716154, + 0.019382018595933914, + 0.0029450086876749992, + 0.00430551590397954, + 0.0051538050174713135, + 0.004458720330148935, + 0.01954888179898262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.023327438160777092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.01821434497833252 + ], + "opposite_logprob_tokens": [ + -4.018214225769043 + ], + "logprob_reward": -0.01821434497833252, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821434497833252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45753955841064453, + 0.07231143116950989, + 0.15676747262477875, + 0.0570317767560482, + 0.051726628094911575, + 0.02916129305958748, + 0.053631510585546494, + 0.03371778130531311, + 0.040190190076828, + 0.04792232811450958 + ], + "credit_map": [ + 0.14584051072597504, + 0.01690722443163395, + 0.17803166806697845, + 0.023140523582696915, + 0.034944530576467514, + 0.031977396458387375, + 0.012612907215952873, + 0.014084807597100735, + 0.006875881925225258, + 0.006996587850153446, + 0.007963155396282673, + 0.017953813076019287, + 0.00897674635052681, + 0.009865989908576012, + 0.008154040202498436, + 0.005525220651179552, + 0.021950412541627884, + 0.016456369310617447, + 0.01689610816538334, + 0.04103996977210045, + 0.012975931167602539, + 0.016674082726240158, + 0.017811451107263565, + 0.012963160872459412, + 0.008844275958836079, + 0.005007374100387096, + 0.0089228805154562, + 0.006698834244161844, + 0.00944769661873579, + 0.008512772619724274, + 0.005757111124694347, + 0.0038408320397138596, + 0.006696814205497503, + 0.0036055538803339005, + 0.0076554701663553715, + 0.008820121176540852, + 0.005491424817591906, + 0.00887149665504694, + 0.0052626836113631725, + 0.005323063116520643, + 0.0042474218644201756, + 0.0027396816294640303, + 0.00328347971662879, + 0.006483418866991997, + 0.0028051233384758234, + 0.00372639880515635, + 0.003309248946607113, + 0.0025665205903351307, + 0.014263737946748734, + 0.004594920203089714, + 0.012123681604862213, + 0.0057303146459162235, + 0.003144570393487811, + 0.005533264949917793, + 0.0034509124234318733, + 0.004790110979229212, + 0.004581801127642393, + 0.004345499910414219, + 0.0031726155430078506, + 0.004881871864199638, + 0.00394348194822669, + 0.004743561614304781, + 0.0035434940364211798, + 0.0045054578222334385, + 0.005408352706581354, + 0.0045794276520609856, + 0.004062318708747625, + 0.006046077236533165, + 0.005659850779920816, + 0.005009740125387907, + 0.004944524262100458, + 0.004479899536818266, + 0.011016163043677807, + 0.0027034925296902657, + 0.0172682236880064, + 0.0031231697648763657, + 0.0034387342166155577, + 0.0050653754733502865, + 0.0018967033829540014, + 0.003410463687032461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01821434497833252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.08905123919248581 + ], + "opposite_logprob_tokens": [ + -2.4640512466430664 + ], + "logprob_reward": -0.08905123919248581, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905123919248581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2110258936882019, + 0.13893234729766846, + 0.07632803171873093, + 0.11221683770418167, + 0.19143274426460266, + 0.05776436999440193, + 0.04373207315802574, + 0.07878264039754868, + 0.05372588708996773, + 0.0360591895878315 + ], + "credit_map": [ + 0.04780307784676552, + 0.019797995686531067, + 0.04338568076491356, + 0.026297951117157936, + 0.032410942018032074, + 0.0210163164883852, + 0.011007730849087238, + 0.009306183084845543, + 0.01414358802139759, + 0.014383075758814812, + 0.014998570084571838, + 0.03378598392009735, + 0.020544959232211113, + 0.017503054812550545, + 0.012826889753341675, + 0.010746217332780361, + 0.01070624403655529, + 0.007126026786863804, + 0.006073860451579094, + 0.015386640094220638, + 0.00803354848176241, + 0.014518747106194496, + 0.009135776199400425, + 0.005347186233848333, + 0.014699436724185944, + 0.00846131145954132, + 0.012386412359774113, + 0.02462533675134182, + 0.017547547817230225, + 0.0163114033639431, + 0.011605506762862206, + 0.006579877808690071, + 0.06675811111927032, + 0.00920164491981268, + 0.011102289892733097, + 0.011700551025569439, + 0.007324560545384884, + 0.07002849131822586, + 0.006764211226254702, + 0.008552887476980686, + 0.007792788557708263, + 0.005051781889051199, + 0.0061846645548939705, + 0.012259186245501041, + 0.005761178210377693, + 0.009331104345619678, + 0.006457652430981398, + 0.004926009103655815, + 0.008128924295306206, + 0.0047933487221598625, + 0.007098562549799681, + 0.0075840274803340435, + 0.003648758865892887, + 0.0056027574464678764, + 0.003564660670235753, + 0.0033110296353697777, + 0.008937284350395203, + 0.011906771920621395, + 0.006185508333146572, + 0.0121284956112504, + 0.010132497176527977, + 0.009162712842226028, + 0.010380230844020844, + 0.009949137456715107, + 0.006837595719844103, + 0.0065373629331588745, + 0.005127132870256901, + 0.009252568706870079, + 0.007179421838372946, + 0.006581442430615425, + 0.006531411781907082, + 0.005678944755345583, + 0.0055692377500236034, + 0.0025234681088477373, + 0.012247567065060139, + 0.0029653895180672407, + 0.003152555553242564, + 0.003268847241997719, + 0.0021836746018379927, + 0.004148445557802916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.08905123919248581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.04306705296039581 + ], + "opposite_logprob_tokens": [ + -3.16806697845459 + ], + "logprob_reward": -0.04306705296039581, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306705296039581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.376592755317688, + 0.055670373141765594, + 0.03771758824586868, + 0.1517755091190338, + 0.12625057995319366, + 0.036177001893520355, + 0.06059226393699646, + 0.0582934133708477, + 0.030765444040298462, + 0.06616511195898056 + ], + "credit_map": [ + 0.10265028476715088, + 0.016615519300103188, + 0.13356156647205353, + 0.02502644620835781, + 0.03269779309630394, + 0.030286096036434174, + 0.019650552421808243, + 0.016104498878121376, + 0.0060565536841750145, + 0.00514852162450552, + 0.013787349686026573, + 0.007242913823574781, + 0.008485667407512665, + 0.005896782502532005, + 0.0054168300703167915, + 0.003635754808783531, + 0.0047113774344325066, + 0.0035729166120290756, + 0.0035430388525128365, + 0.009741563349962234, + 0.003178567625582218, + 0.005595841445028782, + 0.004417306277900934, + 0.0029569759499281645, + 0.021080145612359047, + 0.011849411763250828, + 0.01804807223379612, + 0.027347130700945854, + 0.02700444497168064, + 0.022655509412288666, + 0.014416645281016827, + 0.009374149143695831, + 0.043942730873823166, + 0.004303706809878349, + 0.008120466023683548, + 0.004543306306004524, + 0.0038100972305983305, + 0.052372194826602936, + 0.0029412235599011183, + 0.006216859444975853, + 0.004693691618740559, + 0.003420980181545019, + 0.003193932119756937, + 0.010081908665597439, + 0.002625358523800969, + 0.004645055625587702, + 0.003582522738724947, + 0.00393355218693614, + 0.01060273963958025, + 0.005955949425697327, + 0.012847322970628738, + 0.007735427003353834, + 0.006272174883633852, + 0.006271337158977985, + 0.005937953479588032, + 0.004969356581568718, + 0.01019746158272028, + 0.008347216993570328, + 0.00415667612105608, + 0.007664697710424662, + 0.007642161566764116, + 0.007496082689613104, + 0.00624749343842268, + 0.006541620474308729, + 0.0032938537187874317, + 0.0032230664510279894, + 0.0028692937921732664, + 0.004648545291274786, + 0.005124955903738737, + 0.004704548045992851, + 0.003296883078292012, + 0.0036042986903339624, + 0.01658361591398716, + 0.0034689432941377163, + 0.025132764130830765, + 0.0036264590453356504, + 0.004570586606860161, + 0.007149714510887861, + 0.002051806543022394, + 0.003581220982596278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.04306705296039581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, slowly rotating close-up, arcing shot, with the intimacy of a home movie, focuses on a dark, solid wood workbench. An old-fashioned desk lamp casts a soft, concentrated halo of warm yellow light, centered on a pair of well-maintained, deft hands with neatly manicured nails. These hands carefully repair a partially disassembled, delicate brass mechanism, its interior comprised of variously sized, gleaming brass gears, tiny springs, and levers. One hand holds the mechanism's main frame steady with its fingertips, while the other, gripping a pair of delicate tweezers, delicately grasps a tiny gear and steadily and deliberately inserts it into a slot within the intricate mechanism. As the camera slowly rotates, the lamplight reflects off the polished surface of the brass gear, creating fluid, dotted highlights and casting complex, ever-shifting shadows across the wooden surface. In the background, the rest of the living room is blurred, and only the outline of the bookshelf and the edge of the soft sofa can be vaguely seen. The whole scene is wrapped in warm colors, full of concentration, tranquility and healing atmosphere.", + "seed": 30238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_8/index_2401_seed_30238_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": -0.11948444209992885, + "vlm_logprob": -0.11948444209992885, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.11948444209992885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.23269839584827423 + ], + "opposite_logprob_tokens": [ + -1.6076984405517578 + ], + "logprob_reward": -0.23269839584827423, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23269839584827423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13006512820720673, + 0.19796276092529297, + 0.05255332961678505, + 0.1172809824347496, + 0.2146652340888977, + 0.10367105156183243, + 0.04326902702450752, + 0.06342300772666931, + 0.05094679445028305, + 0.026162754744291306 + ], + "credit_map": [ + 0.011650610715150833, + 0.007904089987277985, + 0.02342391386628151, + 0.02333012968301773, + 0.011143569834530354, + 0.01379603985697031, + 0.014771007001399994, + 0.02404576726257801, + 0.02470194175839424, + 0.012924988754093647, + 0.03127771243453026, + 0.04738575965166092, + 0.01676790602505207, + 0.022916067391633987, + 0.020889781415462494, + 0.021098600700497627, + 0.006682093720883131, + 0.004659289959818125, + 0.005217287223786116, + 0.008737553842365742, + 0.007883989252150059, + 0.00874190405011177, + 0.005348502192646265, + 0.0052827103063464165, + 0.019178472459316254, + 0.017566286027431488, + 0.011293928138911724, + 0.01857340708374977, + 0.00989984255284071, + 0.013861073181033134, + 0.012025821954011917, + 0.014882144518196583, + 0.0366700254380703, + 0.020531708374619484, + 0.022169023752212524, + 0.031734224408864975, + 0.01841493882238865, + 0.033637430518865585, + 0.02577465958893299, + 0.025733212009072304, + 0.017697764560580254, + 0.009215133264660835, + 0.012370580807328224, + 0.01422207523137331, + 0.009563872590661049, + 0.010377253405749798, + 0.015390118584036827, + 0.014834261499345303, + 0.009281986393034458, + 0.006101790815591812, + 0.004528248216956854, + 0.00419384241104126, + 0.002808813937008381, + 0.0040051294490695, + 0.007283227983862162, + 0.005065988749265671, + 0.013689154759049416, + 0.006487996317446232, + 0.00803285650908947, + 0.010855565778911114, + 0.005331987515091896, + 0.006053693126887083, + 0.00556211220100522, + 0.007409641053527594, + 0.005501411389559507, + 0.004469702951610088, + 0.0044128489680588245, + 0.004966873675584793, + 0.007313736714422703, + 0.0072554019279778, + 0.010942171327769756, + 0.006084648426622152, + 0.004700113087892532, + 0.0036008888855576515, + 0.0024994215928018093, + 0.0027952403761446476, + 0.0028964185621589422, + 0.0026456820778548717, + 0.004286125302314758, + 0.00273886532522738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.23269839584827423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.04311603680253029 + ], + "opposite_logprob_tokens": [ + -3.1681160926818848 + ], + "logprob_reward": -0.04311603680253029, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311603680253029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08711612224578857, + 0.06745448708534241, + 0.10393495857715607, + 0.10829649120569229, + 0.11866974085569382, + 0.09128785133361816, + 0.06856544315814972, + 0.15112194418907166, + 0.16514062881469727, + 0.03841228038072586 + ], + "credit_map": [ + 0.006633086130023003, + 0.008501752279698849, + 0.013640676625072956, + 0.012488288804888725, + 0.006456024944782257, + 0.017256872728466988, + 0.00807706918567419, + 0.014062355272471905, + 0.007488375995308161, + 0.005000794772058725, + 0.009654438123106956, + 0.010028727352619171, + 0.009831201285123825, + 0.00748848170042038, + 0.00858828704804182, + 0.00937418732792139, + 0.013152177445590496, + 0.012664705514907837, + 0.011829746887087822, + 0.018964070826768875, + 0.01205966155976057, + 0.011603848077356815, + 0.01158128585666418, + 0.012079467065632343, + 0.013135934248566628, + 0.011115132831037045, + 0.009551146999001503, + 0.014691390097141266, + 0.014460532926023006, + 0.01458553597331047, + 0.014690712094306946, + 0.016066113486886024, + 0.01931695081293583, + 0.011596918106079102, + 0.012711446732282639, + 0.016812732443213463, + 0.011699886992573738, + 0.014052348211407661, + 0.01634552702307701, + 0.016133945435285568, + 0.012956160120666027, + 0.008518029935657978, + 0.0101111875846982, + 0.014045803807675838, + 0.007954894565045834, + 0.009400459937751293, + 0.016211912035942078, + 0.012089414522051811, + 0.008095244877040386, + 0.006226851139217615, + 0.0075379894115030766, + 0.010168488137423992, + 0.0063069043681025505, + 0.008073275908827782, + 0.014135663397610188, + 0.008021031506359577, + 0.01757812313735485, + 0.013935834169387817, + 0.017142221331596375, + 0.021490588784217834, + 0.01702103018760681, + 0.021146710962057114, + 0.024802759289741516, + 0.018004676327109337, + 0.013596759177744389, + 0.015696482732892036, + 0.017877237871289253, + 0.0197328869253397, + 0.025374101474881172, + 0.015057426877319813, + 0.03934413939714432, + 0.018461616709828377, + 0.003863712539896369, + 0.003178080776706338, + 0.004313815850764513, + 0.006118678487837315, + 0.005204436369240284, + 0.0036897764075547457, + 0.007314278278499842, + 0.004729505628347397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.04311603680253029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.04311969131231308 + ], + "opposite_logprob_tokens": [ + -3.1681196689605713 + ], + "logprob_reward": -0.04311969131231308, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311969131231308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17336124181747437, + 0.15557065606117249, + 0.10891515761613846, + 0.06767243146896362, + 0.13453605771064758, + 0.06939969211816788, + 0.09002382308244705, + 0.10237228125333786, + 0.07332222163677216, + 0.024826405569911003 + ], + "credit_map": [ + 0.01880471222102642, + 0.011788957752287388, + 0.03321077302098274, + 0.019252236932516098, + 0.012744021601974964, + 0.029114214703440666, + 0.019757820293307304, + 0.028688514605164528, + 0.013368449173867702, + 0.01137086283415556, + 0.026799503713846207, + 0.03491251543164253, + 0.012360824272036552, + 0.017946774140000343, + 0.01888258010149002, + 0.019929155707359314, + 0.017317594960331917, + 0.008173039183020592, + 0.012492047622799873, + 0.015779318287968636, + 0.010556530207395554, + 0.01026014145463705, + 0.014934965409338474, + 0.019401537254452705, + 0.010583445429801941, + 0.005077004432678223, + 0.006875134538859129, + 0.006635913625359535, + 0.009944315999746323, + 0.012161342427134514, + 0.006875553168356419, + 0.009519726037979126, + 0.029547639191150665, + 0.015032639726996422, + 0.012115564197301865, + 0.019724972546100616, + 0.011084436438977718, + 0.012579570524394512, + 0.015809351578354836, + 0.018641900271177292, + 0.01100184116512537, + 0.0060868412256240845, + 0.008578949607908726, + 0.007604206446558237, + 0.006480555981397629, + 0.013050618581473827, + 0.007572464644908905, + 0.009024228900671005, + 0.01745867356657982, + 0.009603854268789291, + 0.00923139601945877, + 0.013334420509636402, + 0.0058356113731861115, + 0.00749824196100235, + 0.016987556591629982, + 0.010074072517454624, + 0.01479590404778719, + 0.009509643539786339, + 0.010682733729481697, + 0.014756448566913605, + 0.01156824640929699, + 0.012073207646608353, + 0.01757894456386566, + 0.011407172307372093, + 0.006032828241586685, + 0.007734479382634163, + 0.007897243835031986, + 0.010169556364417076, + 0.008710280992090702, + 0.006499026902019978, + 0.019017186015844345, + 0.007261622231453657, + 0.0031782498117536306, + 0.0027449382469058037, + 0.002538515254855156, + 0.0029797216411679983, + 0.003367763478308916, + 0.0026613816153258085, + 0.004757862538099289, + 0.002597975777462125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04311969131231308 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.2521339952945709 + ], + "opposite_logprob_tokens": [ + -1.5021339654922485 + ], + "logprob_reward": -0.2521339952945709, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521339952945709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0859968289732933, + 0.07714484632015228, + 0.0975489392876625, + 0.0873432382941246, + 0.18408901989459991, + 0.043585795909166336, + 0.12542784214019775, + 0.19737207889556885, + 0.07050652801990509, + 0.030984867364168167 + ], + "credit_map": [ + 0.009027627296745777, + 0.00901790615171194, + 0.015890508890151978, + 0.015222408808767796, + 0.007591013330966234, + 0.007187665905803442, + 0.009411520324647427, + 0.012648181989789009, + 0.012158200144767761, + 0.005607976578176022, + 0.010748131200671196, + 0.00795977097004652, + 0.00975361093878746, + 0.011460354551672935, + 0.008661815896630287, + 0.010794986970722675, + 0.012850199826061726, + 0.011098125949501991, + 0.011318158358335495, + 0.016694998368620872, + 0.012074101716279984, + 0.00980544276535511, + 0.009769723750650883, + 0.013938182033598423, + 0.013048837892711163, + 0.013837415724992752, + 0.00772786745801568, + 0.01127646118402481, + 0.006841063965111971, + 0.010863150469958782, + 0.010960496962070465, + 0.012787938117980957, + 0.035651106387376785, + 0.01921551488339901, + 0.018325863406062126, + 0.023513279855251312, + 0.014665557071566582, + 0.027180315926671028, + 0.0239882692694664, + 0.021549109369516373, + 0.009043938480317593, + 0.004941603634506464, + 0.00469937501475215, + 0.003533275332301855, + 0.0058189881965518, + 0.005395111627876759, + 0.0046094791032373905, + 0.005544023588299751, + 0.017798516899347305, + 0.010453849099576473, + 0.015136235393583775, + 0.015533920377492905, + 0.01101119164377451, + 0.01639123633503914, + 0.024681758135557175, + 0.014421134255826473, + 0.01974431425333023, + 0.01559398416429758, + 0.02078765630722046, + 0.026004021987318993, + 0.02679709903895855, + 0.028110364452004433, + 0.035677410662174225, + 0.024657219648361206, + 0.006678042002022266, + 0.008023087866604328, + 0.006856737192720175, + 0.008674520067870617, + 0.00703474972397089, + 0.009215367957949638, + 0.015411056578159332, + 0.008612966164946556, + 0.004922397434711456, + 0.002904596971347928, + 0.0033768073190003633, + 0.0038207860197871923, + 0.0035311123356223106, + 0.0034503485076129436, + 0.005821607541292906, + 0.003157213795930147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.2521339952945709 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.026354091241955757 + ], + "opposite_logprob_tokens": [ + -3.6513540744781494 + ], + "logprob_reward": -0.026354091241955757, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354091241955757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10082890093326569, + 0.09032131731510162, + 0.03318772837519646, + 0.12601664662361145, + 0.11536715179681778, + 0.1656915694475174, + 0.10260340571403503, + 0.08125050365924835, + 0.1501159369945526, + 0.03461688384413719 + ], + "credit_map": [ + 0.007854889146983624, + 0.009856531396508217, + 0.014537758193910122, + 0.011168880388140678, + 0.01151197124272585, + 0.019786570221185684, + 0.011852141469717026, + 0.014260155148804188, + 0.009858638979494572, + 0.00606776075437665, + 0.013645702973008156, + 0.01935197226703167, + 0.009031872265040874, + 0.011054642498493195, + 0.01121057104319334, + 0.010100145824253559, + 0.0033579333685338497, + 0.002595121506601572, + 0.003794866381213069, + 0.00511386152356863, + 0.004416298121213913, + 0.004111696034669876, + 0.00521978922188282, + 0.004578160587698221, + 0.012969166971743107, + 0.013413575477898121, + 0.01503605768084526, + 0.019362973049283028, + 0.015567079186439514, + 0.01148992870002985, + 0.016844255849719048, + 0.0213335994631052, + 0.026528937742114067, + 0.009703838266432285, + 0.011808275245130062, + 0.01618860475718975, + 0.008694526739418507, + 0.013005085289478302, + 0.01236781757324934, + 0.01707005873322487, + 0.024497978389263153, + 0.014620411209762096, + 0.019360121339559555, + 0.016974836587905884, + 0.0196183230727911, + 0.0323878638446331, + 0.019446048885583878, + 0.018785972148180008, + 0.02917371690273285, + 0.01715286262333393, + 0.009923949837684631, + 0.010039208456873894, + 0.0049475315026938915, + 0.0077424137853085995, + 0.009714562445878983, + 0.013909139670431614, + 0.010713385418057442, + 0.006012964062392712, + 0.010647616349160671, + 0.01153945829719305, + 0.009318648837506771, + 0.009998113848268986, + 0.012258509173989296, + 0.010761796496808529, + 0.01630474627017975, + 0.016292234882712364, + 0.01312131155282259, + 0.019283797591924667, + 0.02150885760784149, + 0.01192383747547865, + 0.03350590541958809, + 0.018175235018134117, + 0.004381345119327307, + 0.0024043135344982147, + 0.00340290367603302, + 0.004063216969370842, + 0.0042532021179795265, + 0.003916317131370306, + 0.007131093647330999, + 0.005064487457275391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.026354091241955757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, slowly rotating close-up, arcing shot, with the intimacy of a home movie, focuses on a dark, solid wood workbench. An old-fashioned desk lamp casts a soft, concentrated halo of warm yellow light, centered on a pair of well-maintained, deft hands with neatly manicured nails. These hands carefully repair a partially disassembled, delicate brass mechanism, its interior comprised of variously sized, gleaming brass gears, tiny springs, and levers. One hand holds the mechanism's main frame steady with its fingertips, while the other, gripping a pair of delicate tweezers, delicately grasps a tiny gear and steadily and deliberately inserts it into a slot within the intricate mechanism. As the camera slowly rotates, the lamplight reflects off the polished surface of the brass gear, creating fluid, dotted highlights and casting complex, ever-shifting shadows across the wooden surface. In the background, the rest of the living room is blurred, and only the outline of the bookshelf and the edge of the soft sofa can be vaguely seen. The whole scene is wrapped in warm colors, full of concentration, tranquility and healing atmosphere.", + "seed": 30239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_8/index_2401_seed_30239_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": -0.11038533709943295, + "vlm_logprob": -0.11038533709943295, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.11038533709943295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.3929271697998047 + ], + "opposite_logprob_tokens": [ + -1.1429271697998047 + ], + "logprob_reward": -0.3929271697998047, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3929271697998047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05059337988495827, + 0.055710792541503906, + 0.035515252500772476, + 0.04203984513878822, + 0.018128879368305206, + 0.022912457585334778, + 0.08749432116746902, + 0.10930919647216797, + 0.563320517539978, + 0.014975315891206264 + ], + "credit_map": [ + 0.006169229745864868, + 0.004300427157431841, + 0.005983833689242601, + 0.004810880869626999, + 0.005842849612236023, + 0.0055014826357364655, + 0.007595037575811148, + 0.010389645583927631, + 0.00728291179984808, + 0.005600645672529936, + 0.008646434172987938, + 0.00890507735311985, + 0.005790672264993191, + 0.007219858933240175, + 0.005245706997811794, + 0.007019487675279379, + 0.004229922313243151, + 0.003711310215294361, + 0.005853802897036076, + 0.006759925279766321, + 0.0028000117745250463, + 0.004735104739665985, + 0.0036053273361176252, + 0.0038198556285351515, + 0.0032287361100316048, + 0.005397462751716375, + 0.0054943677969276905, + 0.005389058962464333, + 0.005351129919290543, + 0.006394123658537865, + 0.005711635109037161, + 0.005073335021734238, + 0.0019109479617327452, + 0.003292589681223035, + 0.0021893326193094254, + 0.0032984623685479164, + 0.0014981001149863005, + 0.0027147517539560795, + 0.0014501381665468216, + 0.0017745587974786758, + 0.0021354418713599443, + 0.00277366628870368, + 0.003196300705894828, + 0.00480116531252861, + 0.0023202141746878624, + 0.0031774688977748156, + 0.001991491299122572, + 0.0025167108979076147, + 0.007738460320979357, + 0.006366279441863298, + 0.016302181407809258, + 0.019312353804707527, + 0.005567092448472977, + 0.01331691537052393, + 0.009134589694440365, + 0.009756462648510933, + 0.012232428416609764, + 0.006866671610623598, + 0.013692946173250675, + 0.012358599342405796, + 0.009917374700307846, + 0.012260680086910725, + 0.023544352501630783, + 0.01843615435063839, + 0.031055262312293053, + 0.012537085451185703, + 0.019694019109010696, + 0.01569453813135624, + 0.023110341280698776, + 0.22196705639362335, + 0.03020077385008335, + 0.20906153321266174, + 0.0015957958530634642, + 0.0012573335552588105, + 0.002080854494124651, + 0.0025952537544071674, + 0.00194133585318923, + 0.00152895029168576, + 0.001875394838862121, + 0.002100398764014244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.3929271697998047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.02333815209567547 + ], + "opposite_logprob_tokens": [ + -3.7733380794525146 + ], + "logprob_reward": -0.02333815209567547, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02333815209567547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06540565937757492, + 0.07996910065412521, + 0.03820309787988663, + 0.1410844922065735, + 0.04001331329345703, + 0.08915580064058304, + 0.07630883902311325, + 0.17472156882286072, + 0.2672220766544342, + 0.027916066348552704 + ], + "credit_map": [ + 0.009981595911085606, + 0.005682848393917084, + 0.007991624064743519, + 0.007504943758249283, + 0.006644879467785358, + 0.008201361633837223, + 0.009421014226973057, + 0.009977390058338642, + 0.0084438631311059, + 0.00872807390987873, + 0.011506883427500725, + 0.013597947545349598, + 0.007068576291203499, + 0.009156160056591034, + 0.01029014028608799, + 0.011177455075085163, + 0.0047152042388916016, + 0.00437211012467742, + 0.0057524205185472965, + 0.0051972344517707825, + 0.0036419497337192297, + 0.0028412528336048126, + 0.00803288072347641, + 0.0036500452551990747, + 0.009274368174374104, + 0.022294051945209503, + 0.020231593400239944, + 0.02619270049035549, + 0.013760306872427464, + 0.01521178800612688, + 0.019795561209321022, + 0.014324123971164227, + 0.004654449876397848, + 0.006644320208579302, + 0.004938003607094288, + 0.008346516638994217, + 0.0030581834726035595, + 0.004768625367432833, + 0.0031620380468666553, + 0.004441177472472191, + 0.008770288899540901, + 0.010577787645161152, + 0.01759544387459755, + 0.014478287659585476, + 0.009828809648752213, + 0.01407062727957964, + 0.0059987944550812244, + 0.007835759781301022, + 0.005328938364982605, + 0.0051580253057181835, + 0.013228049501776695, + 0.01865474320948124, + 0.0038337958976626396, + 0.012397288344800472, + 0.008934736251831055, + 0.0087732607498765, + 0.01570548675954342, + 0.014931106008589268, + 0.020920496433973312, + 0.02609611302614212, + 0.014752313494682312, + 0.01957455836236477, + 0.03594502806663513, + 0.02679646760225296, + 0.017810631543397903, + 0.007978218607604504, + 0.017255084589123726, + 0.012349949218332767, + 0.013868504203855991, + 0.09252814203500748, + 0.020619332790374756, + 0.0848122388124466, + 0.003497539320960641, + 0.0022914099972695112, + 0.003225062508136034, + 0.006557741202414036, + 0.002558692591264844, + 0.003027055412530899, + 0.0028883356135338545, + 0.003870232030749321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.02333815209567547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.023314161226153374 + ], + "opposite_logprob_tokens": [ + -3.7733142375946045 + ], + "logprob_reward": -0.023314161226153374, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314161226153374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0825648382306099, + 0.048662006855010986, + 0.06777290254831314, + 0.22011932730674744, + 0.0741557627916336, + 0.12896527349948883, + 0.10994496196508408, + 0.17832033336162567, + 0.05582144483923912, + 0.033673208206892014 + ], + "credit_map": [ + 0.012351316399872303, + 0.008391600102186203, + 0.009330050088465214, + 0.008709351532161236, + 0.00954359583556652, + 0.010895319283008575, + 0.010298424400389194, + 0.013045179657638073, + 0.005789302755147219, + 0.0047152056358754635, + 0.006658088881522417, + 0.006963609252125025, + 0.005310604348778725, + 0.005313403904438019, + 0.0071700261905789375, + 0.0067417616955935955, + 0.006665293127298355, + 0.00830091442912817, + 0.01231300551444292, + 0.01702304370701313, + 0.004747659899294376, + 0.00413934700191021, + 0.009428371675312519, + 0.0051552606746554375, + 0.01277691125869751, + 0.032732993364334106, + 0.036092184484004974, + 0.05457395315170288, + 0.017325114458799362, + 0.024669520556926727, + 0.023705774918198586, + 0.01824287138879299, + 0.007242472842335701, + 0.011959278024733067, + 0.010427645407617092, + 0.010893510654568672, + 0.004458538256585598, + 0.009586144238710403, + 0.006444796919822693, + 0.013143368996679783, + 0.011258823797106743, + 0.01541790273040533, + 0.02318151295185089, + 0.02387378178536892, + 0.015066763386130333, + 0.019570650532841682, + 0.009882022626698017, + 0.01071380265057087, + 0.007592365611344576, + 0.007715174928307533, + 0.018519202247262, + 0.027195502072572708, + 0.005688078235834837, + 0.017711320891976357, + 0.011286931112408638, + 0.014236373826861382, + 0.016934089362621307, + 0.016171399503946304, + 0.025940798223018646, + 0.024146653711795807, + 0.018453391268849373, + 0.019481591880321503, + 0.03166363388299942, + 0.025528758764266968, + 0.008559063076972961, + 0.0037063653580844402, + 0.009121090173721313, + 0.006105185020714998, + 0.005374955013394356, + 0.0061400760896503925, + 0.008238500915467739, + 0.008576205931603909, + 0.0044255331158638, + 0.00280049629509449, + 0.004004647023975849, + 0.007999401539564133, + 0.0038499156944453716, + 0.0035646005999296904, + 0.003130838740617037, + 0.0038977705407887697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.023314161226153374 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.08905342221260071 + ], + "opposite_logprob_tokens": [ + -2.4640533924102783 + ], + "logprob_reward": -0.08905342221260071, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905342221260071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09402251243591309, + 0.06499059498310089, + 0.08706784248352051, + 0.16743424534797668, + 0.17207545042037964, + 0.05990619957447052, + 0.05376888066530228, + 0.12793661653995514, + 0.13906563818454742, + 0.03373204171657562 + ], + "credit_map": [ + 0.012157648801803589, + 0.012457162141799927, + 0.01057303138077259, + 0.010746270418167114, + 0.011118552647531033, + 0.009891653433442116, + 0.014114669524133205, + 0.012963528744876385, + 0.007115877699106932, + 0.007171976380050182, + 0.009801548905670643, + 0.010966107249259949, + 0.006922608241438866, + 0.007850795052945614, + 0.005872005596756935, + 0.009289676323533058, + 0.005570556037127972, + 0.009120160713791847, + 0.015427087433636189, + 0.021028120070695877, + 0.007193068508058786, + 0.006414908450096846, + 0.015161458402872086, + 0.0071524824015796185, + 0.011778648942708969, + 0.029334770515561104, + 0.022975454106926918, + 0.02269546128809452, + 0.017976894974708557, + 0.018182603642344475, + 0.027915090322494507, + 0.016575312241911888, + 0.015415786765515804, + 0.024255990982055664, + 0.02125234156847, + 0.021514730527997017, + 0.023132581263780594, + 0.02772177755832672, + 0.016938824206590652, + 0.02184341289103031, + 0.004892449360340834, + 0.006357318256050348, + 0.010206067934632301, + 0.007922078482806683, + 0.0057001574896276, + 0.009988879784941673, + 0.006754293572157621, + 0.00808495469391346, + 0.006638773716986179, + 0.005049687810242176, + 0.006444933824241161, + 0.00858605932444334, + 0.005157605744898319, + 0.007676165085285902, + 0.006304491311311722, + 0.007911162450909615, + 0.014321496710181236, + 0.01224525086581707, + 0.012294331565499306, + 0.01902998611330986, + 0.01172203104943037, + 0.014209451153874397, + 0.02082330547273159, + 0.02329077012836933, + 0.019599076360464096, + 0.009157446213066578, + 0.02150978520512581, + 0.01446851622313261, + 0.012882702983915806, + 0.01720559224486351, + 0.020048048347234726, + 0.024194475263357162, + 0.00406500743702054, + 0.002586121903732419, + 0.003940008580684662, + 0.006177187897264957, + 0.005568151827901602, + 0.003518025390803814, + 0.003855020273476839, + 0.004022518638521433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.08905342221260071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.02329378016293049 + ], + "opposite_logprob_tokens": [ + -3.7732937335968018 + ], + "logprob_reward": -0.02329378016293049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329378016293049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10760920494794846, + 0.06457201391458511, + 0.13431602716445923, + 0.26724064350128174, + 0.08043081313371658, + 0.04954615607857704, + 0.07232192903757095, + 0.05574194714426994, + 0.12802676856517792, + 0.04019450396299362 + ], + "credit_map": [ + 0.013796081766486168, + 0.013596140779554844, + 0.01251910999417305, + 0.014793056063354015, + 0.013851628638803959, + 0.010629435069859028, + 0.013892462477087975, + 0.014531283639371395, + 0.007943099364638329, + 0.008623706176877022, + 0.008764104917645454, + 0.01018792949616909, + 0.005708467680960894, + 0.010365186259150505, + 0.007510815281420946, + 0.005468703806400299, + 0.016432451084256172, + 0.011533953249454498, + 0.021039064973592758, + 0.02330753020942211, + 0.012960977852344513, + 0.013393412344157696, + 0.020826591178774834, + 0.014822050929069519, + 0.016748109832406044, + 0.026437558233737946, + 0.04368855804204941, + 0.06671030819416046, + 0.02064349129796028, + 0.02275625802576542, + 0.03807421773672104, + 0.03218214213848114, + 0.010334212332963943, + 0.010844625532627106, + 0.01063877996057272, + 0.011381857097148895, + 0.004925339482724667, + 0.009022781625390053, + 0.010633588768541813, + 0.012649629265069962, + 0.005405377596616745, + 0.005710051394999027, + 0.005925863515585661, + 0.007346577476710081, + 0.002987084910273552, + 0.006686779670417309, + 0.006068771705031395, + 0.009415649808943272, + 0.005789646413177252, + 0.006704998668283224, + 0.007594458758831024, + 0.01254450436681509, + 0.004596926271915436, + 0.010723202489316463, + 0.012444835156202316, + 0.011923359706997871, + 0.004518439993262291, + 0.004298236686736345, + 0.008099203929305077, + 0.006684585474431515, + 0.005426728632301092, + 0.007093763444572687, + 0.01149967685341835, + 0.008121312595903873, + 0.00901133194565773, + 0.0039342609234154224, + 0.007047345861792564, + 0.005379189271479845, + 0.005697519984096289, + 0.04744834080338478, + 0.005711800418794155, + 0.04379698634147644, + 0.004397916607558727, + 0.0019513017032295465, + 0.00610710121691227, + 0.003887233790010214, + 0.005432674661278725, + 0.005781727377325296, + 0.005307122599333525, + 0.007329428102821112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.02329378016293049 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth, slowly rotating close-up, arcing shot, with the intimacy of a home movie, focuses on a dark, solid wood workbench. An old-fashioned desk lamp casts a soft, concentrated halo of warm yellow light, centered on a pair of well-maintained, deft hands with neatly manicured nails. These hands carefully repair a partially disassembled, delicate brass mechanism, its interior comprised of variously sized, gleaming brass gears, tiny springs, and levers. One hand holds the mechanism's main frame steady with its fingertips, while the other, gripping a pair of delicate tweezers, delicately grasps a tiny gear and steadily and deliberately inserts it into a slot within the intricate mechanism. As the camera slowly rotates, the lamplight reflects off the polished surface of the brass gear, creating fluid, dotted highlights and casting complex, ever-shifting shadows across the wooden surface. In the background, the rest of the living room is blurred, and only the outline of the bookshelf and the edge of the soft sofa can be vaguely seen. The whole scene is wrapped in warm colors, full of concentration, tranquility and healing atmosphere.", + "seed": 30240, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_8/index_2401_seed_30240_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": -0.08287250325083732, + "vlm_logprob": -0.08287250325083732, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.08287250325083732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.32217738032341003 + ], + "opposite_logprob_tokens": [ + -1.3221774101257324 + ], + "logprob_reward": -0.32217738032341003, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32217738032341003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22042691707611084, + 0.22805756330490112, + 0.06512229144573212, + 0.03545354679226875, + 0.1200810968875885, + 0.03364246338605881, + 0.12411649525165558, + 0.08975330740213394, + 0.03925642371177673, + 0.04408993199467659 + ], + "credit_map": [ + 0.024766171351075172, + 0.01719067059457302, + 0.03511301800608635, + 0.03526374697685242, + 0.012325718067586422, + 0.019098959863185883, + 0.031192738562822342, + 0.045475877821445465, + 0.02972642332315445, + 0.005826687905937433, + 0.07864291965961456, + 0.014027279801666737, + 0.003210099646821618, + 0.02468704804778099, + 0.006300443317741156, + 0.06563667207956314, + 0.005921369884163141, + 0.00853249616920948, + 0.007845843210816383, + 0.008984246291220188, + 0.006053614430129528, + 0.00940424483269453, + 0.010323971509933472, + 0.008056504651904106, + 0.004264089744538069, + 0.004204895813018084, + 0.005628963932394981, + 0.005534622352570295, + 0.0024617828894406557, + 0.0058942572213709354, + 0.003966326359659433, + 0.003498609410598874, + 0.01187201403081417, + 0.01723344810307026, + 0.023509245365858078, + 0.02238309010863304, + 0.005510222166776657, + 0.007443674840033054, + 0.012285196222364902, + 0.019844217225909233, + 0.006570786237716675, + 0.004177902825176716, + 0.005218368023633957, + 0.003613304113969207, + 0.0018836905946955085, + 0.002537884283810854, + 0.003405748400837183, + 0.00623477715998888, + 0.03258776664733887, + 0.011052966117858887, + 0.029581541195511818, + 0.012461553327739239, + 0.004922686144709587, + 0.011254736222326756, + 0.010033591650426388, + 0.01222164835780859, + 0.010739577002823353, + 0.014764174818992615, + 0.0105158481746912, + 0.01504872553050518, + 0.006275561172515154, + 0.008812150917947292, + 0.010369835421442986, + 0.013227432034909725, + 0.005248958710581064, + 0.003439814317971468, + 0.005791505798697472, + 0.0077123502269387245, + 0.0031010524835437536, + 0.004000979475677013, + 0.004333541728556156, + 0.005628220271319151, + 0.005851515103131533, + 0.002722916193306446, + 0.007062897086143494, + 0.008020495064556599, + 0.0035166344605386257, + 0.003906853962689638, + 0.005561962723731995, + 0.00744665740057826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.32217738032341003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.012577733024954796 + ], + "opposite_logprob_tokens": [ + -4.387577533721924 + ], + "logprob_reward": -0.012577733024954796, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012577733024954796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10122288763523102, + 0.268653005361557, + 0.04820086434483528, + 0.041016992181539536, + 0.08655079454183578, + 0.10621574521064758, + 0.10249481350183487, + 0.08739124983549118, + 0.13049550354480743, + 0.027758143842220306 + ], + "credit_map": [ + 0.009207135997712612, + 0.009257633239030838, + 0.018654581159353256, + 0.01150772999972105, + 0.006373423617333174, + 0.012107240036129951, + 0.015180268324911594, + 0.01893487386405468, + 0.03465820848941803, + 0.006803542375564575, + 0.08572201430797577, + 0.016286011785268784, + 0.0042197140865027905, + 0.031254641711711884, + 0.007567926775664091, + 0.0821409672498703, + 0.005383637268096209, + 0.0044949050061404705, + 0.007671258877962828, + 0.00932635460048914, + 0.0037980028428137302, + 0.006303663365542889, + 0.005932055413722992, + 0.005290984641760588, + 0.006604460533708334, + 0.005505613051354885, + 0.005724565126001835, + 0.005672156345099211, + 0.003066264558583498, + 0.004763517063111067, + 0.004104577004909515, + 0.005575839430093765, + 0.01104398351162672, + 0.015750184655189514, + 0.013127277605235577, + 0.016053682193160057, + 0.005543883889913559, + 0.00657958397641778, + 0.010124404914677143, + 0.008327796123921871, + 0.013349809683859348, + 0.017843041568994522, + 0.01518392376601696, + 0.018219642341136932, + 0.005902757402509451, + 0.014904925599694252, + 0.008091805502772331, + 0.012719840742647648, + 0.030024453997612, + 0.0074822441674768925, + 0.029007580131292343, + 0.009331456385552883, + 0.004052550997585058, + 0.008870771154761314, + 0.007212869357317686, + 0.006512896623462439, + 0.011044380255043507, + 0.009507464244961739, + 0.011434926651418209, + 0.011821716092526913, + 0.009044136852025986, + 0.009761258959770203, + 0.009109223261475563, + 0.015668146312236786, + 0.01917986385524273, + 0.010066750459372997, + 0.019412921741604805, + 0.021843837574124336, + 0.01161905936896801, + 0.015346897765994072, + 0.015693116933107376, + 0.017333045601844788, + 0.004316752310842276, + 0.002162849297747016, + 0.0041031804867088795, + 0.003436029190197587, + 0.002851103665307164, + 0.003654093947261572, + 0.0032335950527340174, + 0.00400053896009922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.012577733024954796 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.01609194651246071 + ], + "opposite_logprob_tokens": [ + -4.141091823577881 + ], + "logprob_reward": -0.01609194651246071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609194651246071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16095253825187683, + 0.14107689261436462, + 0.10229352116584778, + 0.044851772487163544, + 0.24939504265785217, + 0.08075520396232605, + 0.04192862659692764, + 0.06264472007751465, + 0.07350309938192368, + 0.04259853810071945 + ], + "credit_map": [ + 0.014131761156022549, + 0.017917465418577194, + 0.02330358326435089, + 0.02557356096804142, + 0.01318039745092392, + 0.022565387189388275, + 0.023232877254486084, + 0.021047521382570267, + 0.032784029841423035, + 0.008098754100501537, + 0.010567628778517246, + 0.012206394225358963, + 0.007317769806832075, + 0.018865916877985, + 0.012428744696080685, + 0.03880766034126282, + 0.011768441647291183, + 0.015988832339644432, + 0.012104030698537827, + 0.012449664995074272, + 0.00980188138782978, + 0.012977554462850094, + 0.01214656699448824, + 0.015056564472615719, + 0.006827136036008596, + 0.005543161183595657, + 0.006745382212102413, + 0.007962765172123909, + 0.0035402807407081127, + 0.004803481511771679, + 0.0050149294547736645, + 0.004414638504385948, + 0.02672695741057396, + 0.036904312670230865, + 0.05170426517724991, + 0.03566089645028114, + 0.011479990556836128, + 0.013923948630690575, + 0.0206082072108984, + 0.05238650366663933, + 0.009943398647010326, + 0.010497705079615116, + 0.0119602857157588, + 0.013444196432828903, + 0.005413298029452562, + 0.013524742797017097, + 0.006512840278446674, + 0.009458745829761028, + 0.009548609144985676, + 0.003839476266875863, + 0.008389338850975037, + 0.004894389305263758, + 0.003016443457454443, + 0.004501082003116608, + 0.0033930297940969467, + 0.0043462663888931274, + 0.006792236119508743, + 0.005512098781764507, + 0.01512226928025484, + 0.008980292826890945, + 0.003948905970901251, + 0.005384817253798246, + 0.007123289629817009, + 0.009780815802514553, + 0.00791481789201498, + 0.005713104270398617, + 0.011549204587936401, + 0.012554197572171688, + 0.006204430479556322, + 0.0067090787924826145, + 0.010084243491292, + 0.012774026952683926, + 0.004389549605548382, + 0.0029119281098246574, + 0.007390488870441914, + 0.006998866330832243, + 0.0033049818594008684, + 0.005182951223105192, + 0.005747810006141663, + 0.006671966053545475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.01609194651246071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.033712390810251236 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.033712390810251236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712390810251236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04707916080951691, + 0.11471376568078995, + 0.05471598356962204, + 0.09209085255861282, + 0.18402628600597382, + 0.14961442351341248, + 0.023671792820096016, + 0.06781921535730362, + 0.22194455564022064, + 0.04432396590709686 + ], + "credit_map": [ + 0.006121819838881493, + 0.004196244291961193, + 0.00628502294421196, + 0.006356896832585335, + 0.003639775328338146, + 0.010506969876587391, + 0.0055721234530210495, + 0.004400305915623903, + 0.03036166913807392, + 0.005328626371920109, + 0.01267249695956707, + 0.010615229606628418, + 0.004071233328431845, + 0.013714739121496677, + 0.005876824259757996, + 0.032072946429252625, + 0.0057311467826366425, + 0.004654750227928162, + 0.009924177080392838, + 0.009690088219940662, + 0.003789166221395135, + 0.00678538391366601, + 0.006081045605242252, + 0.008060223422944546, + 0.007560367695987225, + 0.012202122248709202, + 0.019052481278777122, + 0.01908961869776249, + 0.004977732431143522, + 0.010337447747588158, + 0.009285122156143188, + 0.009585962630808353, + 0.013041423633694649, + 0.010538816452026367, + 0.04921573027968407, + 0.016566025093197823, + 0.0050649698823690414, + 0.010813266038894653, + 0.015352899208664894, + 0.06343316286802292, + 0.018690558150410652, + 0.014375332742929459, + 0.0240169744938612, + 0.015378737822175026, + 0.009899703785777092, + 0.02186545729637146, + 0.014069563709199429, + 0.03131808340549469, + 0.0026657774578779936, + 0.002900617429986596, + 0.0030015469528734684, + 0.0032639442943036556, + 0.001833657268434763, + 0.002872074255719781, + 0.0034113903529942036, + 0.003722784109413624, + 0.009870371781289577, + 0.009854745119810104, + 0.008534633554518223, + 0.011989821679890156, + 0.00556692760437727, + 0.0069028702564537525, + 0.008196552284061909, + 0.0069032879546284676, + 0.03558669239282608, + 0.017121585085988045, + 0.03239889070391655, + 0.03332505747675896, + 0.020826300606131554, + 0.02780376747250557, + 0.026957521215081215, + 0.027924751862883568, + 0.005995399318635464, + 0.002660169033333659, + 0.00805925764143467, + 0.006458446849137545, + 0.0038873658049851656, + 0.0038893637247383595, + 0.00693229166790843, + 0.0064416746608912945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.033712390810251236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a close-up shot of a dark, solid wood workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of well-maintained hands with neatly manicured nails visible repairing a brass mechanism?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the brass mechanism comprised of gleaming gears, tiny springs, and levers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does one hand hold the mechanism steady while the other uses tweezers to insert a tiny gear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a smooth, slowly rotating close-up with an arcing 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": [ + -0.029803065583109856 + ], + "opposite_logprob_tokens": [ + -3.5298030376434326 + ], + "logprob_reward": -0.029803065583109856, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803065583109856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1190425306558609, + 0.15894609689712524, + 0.05724918097257614, + 0.05374490097165108, + 0.1952664852142334, + 0.05174297094345093, + 0.06910745799541473, + 0.08463262766599655, + 0.18769097328186035, + 0.022576769813895226 + ], + "credit_map": [ + 0.00958910770714283, + 0.015074992552399635, + 0.01932433433830738, + 0.01849295012652874, + 0.009374192915856838, + 0.019250251352787018, + 0.012552623637020588, + 0.01538407988846302, + 0.025161072611808777, + 0.014912270940840244, + 0.014026261866092682, + 0.012872322462499142, + 0.016824422404170036, + 0.02018130011856556, + 0.022373709827661514, + 0.032594744116067886, + 0.006781099364161491, + 0.007197542116045952, + 0.008268495090305805, + 0.00920585636049509, + 0.004659973084926605, + 0.009119781665503979, + 0.0051584807224571705, + 0.006857953500002623, + 0.007244559470564127, + 0.006052491255104542, + 0.008244325406849384, + 0.00850755162537098, + 0.0041835252195596695, + 0.007730158977210522, + 0.005877417512238026, + 0.005904871504753828, + 0.021472156047821045, + 0.020904971286654472, + 0.04459845647215843, + 0.022467590868473053, + 0.007121780887246132, + 0.011119022034108639, + 0.013365061953663826, + 0.05421743541955948, + 0.005700264126062393, + 0.005218519363552332, + 0.008992746472358704, + 0.007083560805767775, + 0.002541813300922513, + 0.00457767816260457, + 0.005964153911918402, + 0.011664235964417458, + 0.008905195631086826, + 0.0071503352373838425, + 0.009459075517952442, + 0.009456501342356205, + 0.005267563741654158, + 0.006789809092879295, + 0.010068091563880444, + 0.01201088447123766, + 0.008162513375282288, + 0.00776353245601058, + 0.013931144028902054, + 0.012458490207791328, + 0.006061739753931761, + 0.013068648986518383, + 0.009524187073111534, + 0.013662376441061497, + 0.020564470440149307, + 0.013744737021625042, + 0.03076358325779438, + 0.046761538833379745, + 0.012629241682589054, + 0.016599291935563087, + 0.021367957815527916, + 0.025260163471102715, + 0.002247060649096966, + 0.001476527308113873, + 0.0035491189919412136, + 0.005221468396484852, + 0.0021581214386969805, + 0.0021070484071969986, + 0.002717373427003622, + 0.003100050613284111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.029803065583109856 + } + ], + "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/0000122/group_9/index_tensor([6890, 6890])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_9/index_tensor([6890, 6890])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..669eebf11394eed1b2168e0ea910874f336b2ad8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_9/index_tensor([6890, 6890])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a high angle, begins, slowly and steadily moving closer to the central figure. In a quaint, elegant study, imbued with a scholarly atmosphere, the only light comes from the flickering flame of a bronze candlestick on the corner of the table. A slender, ancient strategist, approximately fifty years old, sits solemnly at a heavy rosewood desk. He wears a deep blue, wide-sleeved, finely textured silk robe, his long, jet-black hair meticulously tied atop his head with an antique wooden hairpin. The candlelight casts warm, pulsating shadows across his deep features, illuminating his high nose and the slight furrow of his brow, drawn in intense concentration. His right hand, steady and composed, holds a wolf-hair brush, the shaft softened by the warmth of his fingers. The tip, saturated with thick ink, moves smoothly across the spread of rice paper, leaving behind each powerful, vibrant, and inky Chinese character. As the camera moves forward extremely slowly, the picture gradually tightens into a close-up of his face. We can clearly see the wisdom flashing in his deep eyes and his total commitment to the strategy of his pen. His tightly pursed lips reveal his inner determination and deep consideration. The whole space is filled with a quiet, solemn and intellectually powerful atmosphere. It seems that the only sounds in the air are the faint crackling of the candlelight and the rustling of the pen tip across the paper.", + "seed": 27465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_9/index_6890_seed_27465_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": -0.07055819714441895, + "vlm_logprob": -0.07055819714441895, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.07055819714441895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.19156603515148163 + ], + "opposite_logprob_tokens": [ + -1.941565990447998 + ], + "logprob_reward": -0.19156603515148163, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.19156603515148163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2237144112586975, + 0.03792637959122658, + 0.024179117754101753, + 0.048986803740262985, + 0.030639586970210075, + 0.03496776148676872, + 0.3942880630493164, + 0.04597784951329231, + 0.07647562026977539, + 0.08284438401460648 + ], + "credit_map": [ + 0.023101532831788063, + 0.02918964996933937, + 0.015809671953320503, + 0.02467743493616581, + 0.03483983874320984, + 0.038391973823308945, + 0.022715816274285316, + 0.03498849645256996, + 0.005228900816291571, + 0.0047248052433133125, + 0.004145283717662096, + 0.0025471088010817766, + 0.0056448644027113914, + 0.005303990561515093, + 0.004347288049757481, + 0.005984139628708363, + 0.0015327362343668938, + 0.004477365408092737, + 0.0016276852693408728, + 0.0018886661855503917, + 0.0035039593931287527, + 0.0022309652995318174, + 0.005437887739390135, + 0.0034798509441316128, + 0.003418617183342576, + 0.006707636173814535, + 0.003987264819443226, + 0.004477431066334248, + 0.007278358563780785, + 0.004188999999314547, + 0.010906917974352837, + 0.008021580055356026, + 0.003941776230931282, + 0.0034892153926193714, + 0.00314223300665617, + 0.0029743558261543512, + 0.004318490158766508, + 0.0031563343945890665, + 0.004672606475651264, + 0.004944575484842062, + 0.0029266998171806335, + 0.003930708859115839, + 0.0033356717322021723, + 0.004213392734527588, + 0.005506687331944704, + 0.002428004750981927, + 0.007427355740219355, + 0.0051992409862577915, + 0.010204270482063293, + 0.005877029616385698, + 0.01030108705163002, + 0.012222304940223694, + 0.016184624284505844, + 0.024118071421980858, + 0.17714983224868774, + 0.1382308453321457, + 0.0054504601284861565, + 0.004138994961977005, + 0.003009992651641369, + 0.006644899956882, + 0.007911614142358303, + 0.004974485374987125, + 0.00644432520493865, + 0.00740307942032814, + 0.005532275419682264, + 0.006828265264630318, + 0.00553027493879199, + 0.0068948231637477875, + 0.013369050808250904, + 0.006157759111374617, + 0.015390128828585148, + 0.01677304320037365, + 0.00867520272731781, + 0.007998173125088215, + 0.0061913528479635715, + 0.005367247387766838, + 0.02017260156571865, + 0.00553555553779006, + 0.004324817564338446, + 0.024579443037509918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.19156603515148163 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.07008183747529984 + ], + "opposite_logprob_tokens": [ + -2.695081949234009 + ], + "logprob_reward": -0.07008183747529984, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008183747529984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1208578422665596, + 0.06616080552339554, + 0.02507673017680645, + 0.016876494511961937, + 0.00869271345436573, + 0.014516521245241165, + 0.6211085319519043, + 0.02491125836968422, + 0.06978750973939896, + 0.03201163187623024 + ], + "credit_map": [ + 0.014120860956609249, + 0.014563539996743202, + 0.0109496945515275, + 0.011574278585612774, + 0.015549628995358944, + 0.024572765454649925, + 0.009599036537110806, + 0.01992802508175373, + 0.010220798663794994, + 0.0050794510170817375, + 0.012365814298391342, + 0.0055941008031368256, + 0.005997365340590477, + 0.010374932549893856, + 0.0051661464385688305, + 0.011362189427018166, + 0.0028577339835464954, + 0.002097785472869873, + 0.0039839427918195724, + 0.0018499601865187287, + 0.002630793023854494, + 0.005353120155632496, + 0.002675920957699418, + 0.00362747092731297, + 0.0024581621401011944, + 0.0014389302814379334, + 0.0020917998626828194, + 0.0019320317078381777, + 0.0027986548375338316, + 0.0016485616797581315, + 0.0024584964849054813, + 0.0020498554222285748, + 0.001340119750238955, + 0.0008180086733773351, + 0.0009739742963574827, + 0.0007181154796853662, + 0.0012137491721659899, + 0.0009646991384215653, + 0.001451972289942205, + 0.0012120750034227967, + 0.0017799322959035635, + 0.0017106907907873392, + 0.0018481172155588865, + 0.0015209021512418985, + 0.0019539031200110912, + 0.0014749814290553331, + 0.0021292532328516245, + 0.0020987396128475666, + 0.016596728935837746, + 0.00907371286302805, + 0.016137385740876198, + 0.019597189500927925, + 0.02447664365172386, + 0.04310496896505356, + 0.2798651158809662, + 0.21225672960281372, + 0.002921360544860363, + 0.00258619780652225, + 0.0019024420762434602, + 0.003791981842368841, + 0.003998834639787674, + 0.0030175242573022842, + 0.0025606942363083363, + 0.004132220521569252, + 0.003453637706115842, + 0.0028458889573812485, + 0.0029954840429127216, + 0.003913385793566704, + 0.009294664487242699, + 0.0033840120304375887, + 0.02295200526714325, + 0.02094842493534088, + 0.003612122032791376, + 0.003239206736907363, + 0.002939851488918066, + 0.0026624328456819057, + 0.005932431668043137, + 0.0028487013187259436, + 0.0033907978795468807, + 0.007386083714663982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.07008183747529984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.07012395560741425 + ], + "opposite_logprob_tokens": [ + -2.6951239109039307 + ], + "logprob_reward": -0.07012395560741425, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07012395560741425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5304996967315674, + 0.05452584847807884, + 0.023634109646081924, + 0.01902359165251255, + 0.034304868429899216, + 0.027729295194149017, + 0.1939280927181244, + 0.03499094396829605, + 0.03559822961688042, + 0.04576536640524864 + ], + "credit_map": [ + 0.05183838680386543, + 0.06470256298780441, + 0.03775133937597275, + 0.05774267762899399, + 0.09330643713474274, + 0.08773517608642578, + 0.05356897786259651, + 0.08385413140058517, + 0.006283654365688562, + 0.003974447958171368, + 0.012794164009392262, + 0.0064360154792666435, + 0.00517083378508687, + 0.0051936316303908825, + 0.006753633264452219, + 0.007919466122984886, + 0.002122999867424369, + 0.0038485112600028515, + 0.0022931171115487814, + 0.002282883273437619, + 0.003869772655889392, + 0.0018492292147129774, + 0.004176503978669643, + 0.0031910925172269344, + 0.0013654975919052958, + 0.0018447588663548231, + 0.0019092862494289875, + 0.002190887928009033, + 0.0031301204580813646, + 0.002963119652122259, + 0.0029777607414871454, + 0.0026421595830470324, + 0.004328974988311529, + 0.003292200854048133, + 0.0035703328903764486, + 0.0027986629866063595, + 0.004950147587805986, + 0.003030288266018033, + 0.006192950531840324, + 0.006141309626400471, + 0.003966079093515873, + 0.003897853894159198, + 0.002139827236533165, + 0.0026778122410178185, + 0.002992859110236168, + 0.0022012873087078333, + 0.00565355084836483, + 0.004200025927275419, + 0.007096193730831146, + 0.0040305727161467075, + 0.005389435216784477, + 0.007199110928922892, + 0.008194842375814915, + 0.014222797937691212, + 0.08568967878818512, + 0.0621054507791996, + 0.0030998799484223127, + 0.004177260212600231, + 0.0030727924313396215, + 0.0063662901520729065, + 0.0059875319711863995, + 0.0035368772223591805, + 0.0040203481912612915, + 0.004729964304715395, + 0.002501058392226696, + 0.0018966273637488484, + 0.0028922446072101593, + 0.002913579810410738, + 0.004640966188162565, + 0.0029800247866660357, + 0.00818334799259901, + 0.009590382687747478, + 0.004562051966786385, + 0.0049433596432209015, + 0.0048515684902668, + 0.006534439977258444, + 0.008199348114430904, + 0.004414568189531565, + 0.006510790903121233, + 0.0057492367923259735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.07012395560741425 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.018195146694779396 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018195146694779396, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195146694779396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023167021572589874, + 0.016474951058626175, + 0.010958034545183182, + 0.007719567511230707, + 0.022139383479952812, + 0.018431568518280983, + 0.7301536202430725, + 0.0552653931081295, + 0.09496520459651947, + 0.020725276321172714 + ], + "credit_map": [ + 0.0027656913734972477, + 0.002115529263392091, + 0.003505945671349764, + 0.0022835012059658766, + 0.0030918957199901342, + 0.002984051825478673, + 0.0024682197254151106, + 0.0039521860890090466, + 0.002105231862515211, + 0.0021700183860957623, + 0.002058975398540497, + 0.0013109067222103477, + 0.002471314975991845, + 0.001342199044302106, + 0.002634849399328232, + 0.0023814556188881397, + 0.0009645833633840084, + 0.0011757340980693698, + 0.0014052924234420061, + 0.0011428240686655045, + 0.0014374763704836369, + 0.0011817381018772721, + 0.002200661925598979, + 0.0014497251249849796, + 0.0009834603406488895, + 0.0008820935618132353, + 0.0006876499392092228, + 0.0005741827189922333, + 0.001479970058426261, + 0.0009321072138845921, + 0.0011640893062576652, + 0.0010160143719986081, + 0.003481742227450013, + 0.0019928987603634596, + 0.0026614817325025797, + 0.0013336872216314077, + 0.002215359825640917, + 0.002250783145427704, + 0.0032685324549674988, + 0.004934899974614382, + 0.003757200203835964, + 0.0015236614271998405, + 0.003385929623618722, + 0.0012870350619778037, + 0.001496469834819436, + 0.0017607054905965924, + 0.0024931004736572504, + 0.0027274671010673046, + 0.02019992470741272, + 0.0100223608314991, + 0.01880556531250477, + 0.023297609761357307, + 0.02869379334151745, + 0.060089096426963806, + 0.32763969898223877, + 0.24140559136867523, + 0.0068765608593821526, + 0.004788914229720831, + 0.005810566712170839, + 0.010177191346883774, + 0.007494935300201178, + 0.006032243836671114, + 0.005827166605740786, + 0.00825781561434269, + 0.006711812689900398, + 0.005112479440867901, + 0.005342292599380016, + 0.007488444913178682, + 0.009854571893811226, + 0.008186565712094307, + 0.022864945232868195, + 0.029404088854789734, + 0.00252008019015193, + 0.0026574963703751564, + 0.0019690601620823145, + 0.0022148059215396643, + 0.004048495553433895, + 0.0026597646065056324, + 0.0020400811918079853, + 0.0026154934894293547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.018195146694779396 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.002824010793119669 + ], + "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": [ + 0.026528680697083473, + 0.012742682360112667, + 0.012188773602247238, + 0.006656095385551453, + 0.023647533729672432, + 0.021273910999298096, + 0.5198131799697876, + 0.10361764580011368, + 0.2071417272090912, + 0.06638975441455841 + ], + "credit_map": [ + 0.0042485627345740795, + 0.0027360240928828716, + 0.002583458786830306, + 0.002448944142088294, + 0.003355243243277073, + 0.003682773094624281, + 0.003272966481745243, + 0.0042007071897387505, + 0.0017498721135780215, + 0.001575920730829239, + 0.0022440291941165924, + 0.001084933988749981, + 0.0015032985247671604, + 0.0009857747936621308, + 0.001942540518939495, + 0.0016563119133934379, + 0.0011624731123447418, + 0.001770392176695168, + 0.001634719199500978, + 0.001111973193474114, + 0.0015226881951093674, + 0.0011456792708486319, + 0.0021970912348479033, + 0.001643757801502943, + 0.0007870941190049052, + 0.0006453308742493391, + 0.0007685077143833041, + 0.0007563217077404261, + 0.0008340374915860593, + 0.00063991395290941, + 0.0012412209762260318, + 0.0009836682584136724, + 0.0032182326540350914, + 0.0021911156363785267, + 0.002589454408735037, + 0.001688797608949244, + 0.002609978662803769, + 0.002519484143704176, + 0.003686974523589015, + 0.005143496207892895, + 0.0031872414983808994, + 0.0025471567641943693, + 0.003048714017495513, + 0.0024138905573636293, + 0.002071955008432269, + 0.00220477138645947, + 0.002589412732049823, + 0.0032107692677527666, + 0.01435115747153759, + 0.007069956045597792, + 0.01325591653585434, + 0.0162962656468153, + 0.02037801221013069, + 0.04393705353140831, + 0.23296864330768585, + 0.17155617475509644, + 0.009741829708218575, + 0.011538572609424591, + 0.008449394255876541, + 0.013603187166154385, + 0.018434422090649605, + 0.012374946847558022, + 0.014726433902978897, + 0.014748862013220787, + 0.012126046232879162, + 0.010510343126952648, + 0.007176132872700691, + 0.013936247676610947, + 0.02712840586900711, + 0.009987073950469494, + 0.0689212754368782, + 0.05735620856285095, + 0.00490765692666173, + 0.008573410101234913, + 0.003870275802910328, + 0.006642738822847605, + 0.014310899190604687, + 0.008705489337444305, + 0.00846984051167965, + 0.010909441858530045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.002824010793119669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a high angle, begins, slowly and steadily moving closer to the central figure. In a quaint, elegant study, imbued with a scholarly atmosphere, the only light comes from the flickering flame of a bronze candlestick on the corner of the table. A slender, ancient strategist, approximately fifty years old, sits solemnly at a heavy rosewood desk. He wears a deep blue, wide-sleeved, finely textured silk robe, his long, jet-black hair meticulously tied atop his head with an antique wooden hairpin. The candlelight casts warm, pulsating shadows across his deep features, illuminating his high nose and the slight furrow of his brow, drawn in intense concentration. His right hand, steady and composed, holds a wolf-hair brush, the shaft softened by the warmth of his fingers. The tip, saturated with thick ink, moves smoothly across the spread of rice paper, leaving behind each powerful, vibrant, and inky Chinese character. As the camera moves forward extremely slowly, the picture gradually tightens into a close-up of his face. We can clearly see the wisdom flashing in his deep eyes and his total commitment to the strategy of his pen. His tightly pursed lips reveal his inner determination and deep consideration. The whole space is filled with a quiet, solemn and intellectually powerful atmosphere. It seems that the only sounds in the air are the faint crackling of the candlelight and the rustling of the pen tip across the paper.", + "seed": 27466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_9/index_6890_seed_27466_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": -0.06973418425768614, + "vlm_logprob": -0.06973418425768614, + "vlm_yes_no_margin": 3.199999976158142, + "vlm_token_credit": -0.06973418425768614, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.15620000660419464 + ], + "opposite_logprob_tokens": [ + -2.0311999320983887 + ], + "logprob_reward": -0.15620000660419464, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15620000660419464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2913527190685272, + 0.06681987643241882, + 0.06434601545333862, + 0.09044499695301056, + 0.0764034315943718, + 0.04110000282526016, + 0.09976879507303238, + 0.0960163101553917, + 0.04129752516746521, + 0.13245035707950592 + ], + "credit_map": [ + 0.029449356719851494, + 0.02049429900944233, + 0.02102609910070896, + 0.024428969249129295, + 0.0228920578956604, + 0.036245040595531464, + 0.1057974174618721, + 0.031019434332847595, + 0.005980591755360365, + 0.005964390467852354, + 0.005672405939549208, + 0.004491234663873911, + 0.006263989955186844, + 0.015943173319101334, + 0.017855636775493622, + 0.004648453090339899, + 0.01928267627954483, + 0.005396045278757811, + 0.0035100060049444437, + 0.0022060538176447153, + 0.003815309377387166, + 0.021114051342010498, + 0.005119395907968283, + 0.003902471624314785, + 0.0180986188352108, + 0.006405742838978767, + 0.005170705262571573, + 0.0046541402116417885, + 0.006180181633681059, + 0.009591394104063511, + 0.011302986182272434, + 0.029041221365332603, + 0.015280643478035927, + 0.006190290208905935, + 0.007810249924659729, + 0.006722892634570599, + 0.008094350807368755, + 0.005782349966466427, + 0.008471227250993252, + 0.018051423132419586, + 0.005816714838147163, + 0.004807527642697096, + 0.003460484789684415, + 0.0036125946789979935, + 0.003939379006624222, + 0.003997650928795338, + 0.008045721799135208, + 0.0074199289083480835, + 0.016900327056646347, + 0.012384351342916489, + 0.009823836386203766, + 0.008575146086513996, + 0.009970339015126228, + 0.010253222659230232, + 0.015054947696626186, + 0.01680661551654339, + 0.007757192011922598, + 0.008588607423007488, + 0.01096436008810997, + 0.013364363461732864, + 0.017706993967294693, + 0.007854967378079891, + 0.013737118802964687, + 0.01604270190000534, + 0.005573289003223181, + 0.0054421755485236645, + 0.004941657185554504, + 0.005217429716140032, + 0.0063354335725307465, + 0.004088172689080238, + 0.0034606580156832933, + 0.006238704547286034, + 0.011160947382450104, + 0.009692785330116749, + 0.009706340730190277, + 0.019931791350245476, + 0.021631214767694473, + 0.012216473929584026, + 0.023509036749601364, + 0.024601757526397705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.15620000660419464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.10033771395683289 + ], + "opposite_logprob_tokens": [ + -2.3503377437591553 + ], + "logprob_reward": -0.10033771395683289, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033771395683289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15387748181819916, + 0.21837319433689117, + 0.15204225480556488, + 0.04313095286488533, + 0.03945491462945938, + 0.05114712938666344, + 0.10051891952753067, + 0.12172240763902664, + 0.05624549463391304, + 0.0634872242808342 + ], + "credit_map": [ + 0.014985966496169567, + 0.010652806609869003, + 0.015324815176427364, + 0.016080176457762718, + 0.014752903021872044, + 0.016267800703644753, + 0.04933135211467743, + 0.016481664031744003, + 0.024288728833198547, + 0.011959233321249485, + 0.009590917266905308, + 0.009359353221952915, + 0.013273131102323532, + 0.044789206236600876, + 0.08587577193975449, + 0.0192368533462286, + 0.04127702862024307, + 0.012079126201570034, + 0.009220738895237446, + 0.008065467700362206, + 0.007943744771182537, + 0.053417690098285675, + 0.01086853165179491, + 0.009169929660856724, + 0.007554810959845781, + 0.003594148438423872, + 0.0032772687263786793, + 0.002700838027521968, + 0.003080840455368161, + 0.004687334410846233, + 0.005946950055658817, + 0.012288762256503105, + 0.00559382326900959, + 0.0047965035773813725, + 0.006432147230952978, + 0.00638221949338913, + 0.003533625975251198, + 0.0036244590301066637, + 0.004319299478083849, + 0.004772835876792669, + 0.0053915525786578655, + 0.006347701419144869, + 0.008062123320996761, + 0.005665967240929604, + 0.0057940250262618065, + 0.005246902350336313, + 0.0064568123780190945, + 0.008182046934962273, + 0.012212189845740795, + 0.009761761873960495, + 0.009721792303025723, + 0.009893255308270454, + 0.01454425323754549, + 0.010326911695301533, + 0.013236934319138527, + 0.0208218265324831, + 0.01626790687441826, + 0.011688445694744587, + 0.014336202293634415, + 0.013598596677184105, + 0.021201487630605698, + 0.012387972325086594, + 0.012407711707055569, + 0.019834084436297417, + 0.007535828277468681, + 0.00880107656121254, + 0.005772164557129145, + 0.007040509022772312, + 0.007376909721642733, + 0.004126579966396093, + 0.0066929832100868225, + 0.008899441920220852, + 0.004310965072363615, + 0.004532813560217619, + 0.0038425845559686422, + 0.009979097172617912, + 0.012794542126357555, + 0.006796246860176325, + 0.008303203620016575, + 0.012927774339914322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.10033771395683289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.0701131746172905 + ], + "opposite_logprob_tokens": [ + -2.695113182067871 + ], + "logprob_reward": -0.0701131746172905, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0701131746172905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47831669449806213, + 0.2352418303489685, + 0.09322668612003326, + 0.024195728823542595, + 0.02693163976073265, + 0.03241296857595444, + 0.02894457057118416, + 0.025546275079250336, + 0.020057890564203262, + 0.03512575104832649 + ], + "credit_map": [ + 0.048252951353788376, + 0.034978676587343216, + 0.03960650414228439, + 0.045197583734989166, + 0.03536166250705719, + 0.05509558692574501, + 0.16878142952919006, + 0.051042310893535614, + 0.026854567229747772, + 0.011892253533005714, + 0.006670007016509771, + 0.009484568610787392, + 0.011896057054400444, + 0.04451296851038933, + 0.09637385606765747, + 0.027557553723454475, + 0.02209707535803318, + 0.006121661514043808, + 0.005669502541422844, + 0.010663001798093319, + 0.004999871831387281, + 0.027296720072627068, + 0.00873320922255516, + 0.007645642384886742, + 0.0035447468981146812, + 0.002420094795525074, + 0.00276132021099329, + 0.0018732385942712426, + 0.0029574567452073097, + 0.003531780792400241, + 0.0027349675074219704, + 0.004372123628854752, + 0.005992656107991934, + 0.0017144950106739998, + 0.0025645503774285316, + 0.0030649343971163034, + 0.0026934631168842316, + 0.0018502037273719907, + 0.0024144186172634363, + 0.006636917591094971, + 0.004778857342898846, + 0.0037607108242809772, + 0.004256043583154678, + 0.0031497743912041187, + 0.00284940074197948, + 0.0030636072624474764, + 0.005350207444280386, + 0.005204367451369762, + 0.00408173119649291, + 0.0026048265863209963, + 0.003858419833704829, + 0.003332495456561446, + 0.0036816035863012075, + 0.002500907750800252, + 0.0038851837161928415, + 0.004999401047825813, + 0.0029035345651209354, + 0.00261740293353796, + 0.002385444473475218, + 0.002841361565515399, + 0.0040070186369121075, + 0.0032420256175100803, + 0.003439445048570633, + 0.004110041540116072, + 0.0019261169945821166, + 0.002029897877946496, + 0.0015637364704161882, + 0.0026578973047435284, + 0.004106990527361631, + 0.0016745205502957106, + 0.0026111321058124304, + 0.0034875976853072643, + 0.0030744927935302258, + 0.0024944094475358725, + 0.004132754635065794, + 0.0055731055326759815, + 0.004302186891436577, + 0.003837520256638527, + 0.005765961948782206, + 0.005945321638137102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0701131746172905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.016068367287516594 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068367287516594, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068367287516594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15738223493099213, + 0.11680570989847183, + 0.0912175253033638, + 0.04174124822020531, + 0.05584796145558357, + 0.12282498180866241, + 0.046128448098897934, + 0.06250345706939697, + 0.12149053066968918, + 0.18405789136886597 + ], + "credit_map": [ + 0.014214885421097279, + 0.010231892578303814, + 0.014421211555600166, + 0.016392966732382774, + 0.010616753250360489, + 0.017619986087083817, + 0.056401949375867844, + 0.01748259738087654, + 0.011739188805222511, + 0.0074552129954099655, + 0.007573171053081751, + 0.005672828759998083, + 0.007501621730625629, + 0.02547401376068592, + 0.04208244010806084, + 0.00930723361670971, + 0.02596471831202507, + 0.0063040717504918575, + 0.006085053086280823, + 0.01977301947772503, + 0.006217575166374445, + 0.01196445431560278, + 0.009387535974383354, + 0.00552110280841589, + 0.006816655397415161, + 0.004567687399685383, + 0.003778789658099413, + 0.0037212746683508158, + 0.004963426850736141, + 0.0036474766675382853, + 0.0054990993812680244, + 0.008746837265789509, + 0.008617664687335491, + 0.005470362026244402, + 0.004375511314719915, + 0.006082348991185427, + 0.0068879337050020695, + 0.007066000252962112, + 0.008007504045963287, + 0.009340637363493443, + 0.013047858141362667, + 0.018575171008706093, + 0.008543914183974266, + 0.012337508611381054, + 0.01141633465886116, + 0.013343936763703823, + 0.02551155723631382, + 0.02004869654774666, + 0.008340712636709213, + 0.006926152389496565, + 0.003300541779026389, + 0.005563005805015564, + 0.005024963524192572, + 0.0034365742467343807, + 0.005103968549519777, + 0.00843252707272768, + 0.006090420763939619, + 0.005354490131139755, + 0.006656806915998459, + 0.007891596294939518, + 0.009699839167296886, + 0.006082851439714432, + 0.009639016352593899, + 0.011088432744145393, + 0.01508193276822567, + 0.011095775291323662, + 0.009049770422279835, + 0.021952932700514793, + 0.013767344877123833, + 0.011128675192594528, + 0.018916551023721695, + 0.020497547462582588, + 0.022323908284306526, + 0.016630705446004868, + 0.018203314393758774, + 0.029234841465950012, + 0.017887648195028305, + 0.02803361602127552, + 0.02095918171107769, + 0.030784673988819122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.016068367287516594 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.005951658822596073 + ], + "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": [ + 0.09293171018362045, + 0.08151394873857498, + 0.24335362017154694, + 0.04073948785662651, + 0.07173871248960495, + 0.1147458404302597, + 0.033003028482198715, + 0.03126274049282074, + 0.1360516995191574, + 0.1546592265367508 + ], + "credit_map": [ + 0.007597235031425953, + 0.006158254574984312, + 0.0076104276813566685, + 0.009456929750740528, + 0.006253072991967201, + 0.010603677481412888, + 0.034854400902986526, + 0.010397711768746376, + 0.006622681859880686, + 0.005172695964574814, + 0.0030344268307089806, + 0.0031266966834664345, + 0.005096894688904285, + 0.017327794805169106, + 0.03134215250611305, + 0.009790599346160889, + 0.07747471332550049, + 0.01587095484137535, + 0.010703068226575851, + 0.01344728097319603, + 0.011454865336418152, + 0.08784053474664688, + 0.014853710308670998, + 0.011708498932421207, + 0.005933071486651897, + 0.004582023248076439, + 0.0047576213255524635, + 0.0053991880267858505, + 0.004100120160728693, + 0.004816478118300438, + 0.004432255867868662, + 0.006718730088323355, + 0.014460766687989235, + 0.005919921677559614, + 0.005335246212780476, + 0.007851202972233295, + 0.007802080828696489, + 0.005388838239014149, + 0.00813880655914545, + 0.016841856762766838, + 0.011965247802436352, + 0.0164031945168972, + 0.008409000933170319, + 0.009310062043368816, + 0.014965803362429142, + 0.011180059984326363, + 0.023498818278312683, + 0.019013648852705956, + 0.004075867123901844, + 0.0035165196750313044, + 0.003287917235866189, + 0.00344935804605484, + 0.004363948944956064, + 0.004100273363292217, + 0.004115622490644455, + 0.006093519739806652, + 0.004676945507526398, + 0.003963538445532322, + 0.0023513000924140215, + 0.0036165539640933275, + 0.004886874929070473, + 0.004296067636460066, + 0.00262579508125782, + 0.004845665767788887, + 0.008930294774472713, + 0.009210056625306606, + 0.008706778287887573, + 0.02816309779882431, + 0.018705159425735474, + 0.013702341355383396, + 0.023864971473813057, + 0.024768993258476257, + 0.01496775634586811, + 0.012654917314648628, + 0.014716565608978271, + 0.02517285756766796, + 0.01616326905786991, + 0.020008550956845284, + 0.023387311026453972, + 0.0275880079716444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.005951658822596073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a high angle, begins, slowly and steadily moving closer to the central figure. In a quaint, elegant study, imbued with a scholarly atmosphere, the only light comes from the flickering flame of a bronze candlestick on the corner of the table. A slender, ancient strategist, approximately fifty years old, sits solemnly at a heavy rosewood desk. He wears a deep blue, wide-sleeved, finely textured silk robe, his long, jet-black hair meticulously tied atop his head with an antique wooden hairpin. The candlelight casts warm, pulsating shadows across his deep features, illuminating his high nose and the slight furrow of his brow, drawn in intense concentration. His right hand, steady and composed, holds a wolf-hair brush, the shaft softened by the warmth of his fingers. The tip, saturated with thick ink, moves smoothly across the spread of rice paper, leaving behind each powerful, vibrant, and inky Chinese character. As the camera moves forward extremely slowly, the picture gradually tightens into a close-up of his face. We can clearly see the wisdom flashing in his deep eyes and his total commitment to the strategy of his pen. His tightly pursed lips reveal his inner determination and deep consideration. The whole space is filled with a quiet, solemn and intellectually powerful atmosphere. It seems that the only sounds in the air are the faint crackling of the candlelight and the rustling of the pen tip across the paper.", + "seed": 27467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_9/index_6890_seed_27467_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": -0.05046719750389457, + "vlm_logprob": -0.05046719750389457, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.05046719750389457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.10900739580392838 + ], + "opposite_logprob_tokens": [ + -2.4840073585510254 + ], + "logprob_reward": -0.10900739580392838, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10900739580392838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09910237789154053, + 0.06178488954901695, + 0.0816458985209465, + 0.03416653722524643, + 0.08227752894163132, + 0.10546921193599701, + 0.13098116219043732, + 0.0815838873386383, + 0.11249397695064545, + 0.21049447357654572 + ], + "credit_map": [ + 0.011127294041216373, + 0.00912968348711729, + 0.01018375251442194, + 0.015309169888496399, + 0.011509040370583534, + 0.014230914413928986, + 0.016017284244298935, + 0.011595253832638264, + 0.006106621120125055, + 0.006151564419269562, + 0.007118711713701487, + 0.008101871237158775, + 0.010217701084911823, + 0.010009572841227055, + 0.007288176100701094, + 0.006790678948163986, + 0.0073208874091506, + 0.008320723660290241, + 0.008150911889970303, + 0.012045794166624546, + 0.017091866582632065, + 0.013847447000443935, + 0.005491065792739391, + 0.009377215057611465, + 0.0028460067696869373, + 0.003744023386389017, + 0.0033192001283168793, + 0.003034341149032116, + 0.0057888696901500225, + 0.005346870515495539, + 0.003347417339682579, + 0.006739812903106213, + 0.009591297246515751, + 0.007777503691613674, + 0.008929289877414703, + 0.01008727215230465, + 0.014262393116950989, + 0.013100957497954369, + 0.010572344064712524, + 0.007956483401358128, + 0.01492169126868248, + 0.009483990259468555, + 0.008162366226315498, + 0.010999447666108608, + 0.015771299600601196, + 0.014046435244381428, + 0.01607840694487095, + 0.016005586832761765, + 0.018093401566147804, + 0.010096094571053982, + 0.025414321571588516, + 0.009502624161541462, + 0.017840644344687462, + 0.008862425573170185, + 0.020028436556458473, + 0.021143222227692604, + 0.010133273899555206, + 0.008002889342606068, + 0.010521604679524899, + 0.007828762754797935, + 0.011089022271335125, + 0.007601948454976082, + 0.013428359292447567, + 0.01297803781926632, + 0.006752104964107275, + 0.01120084524154663, + 0.017908776178956032, + 0.013171120546758175, + 0.01243662741035223, + 0.010915813967585564, + 0.01659727841615677, + 0.02351141721010208, + 0.02911309152841568, + 0.017177598550915718, + 0.04081425070762634, + 0.018643397837877274, + 0.03369821608066559, + 0.022989869117736816, + 0.02980995364487171, + 0.018248112872242928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.10900739580392838 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.07006093859672546 + ], + "opposite_logprob_tokens": [ + -2.695060968399048 + ], + "logprob_reward": -0.07006093859672546, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006093859672546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16180557012557983, + 0.06992532312870026, + 0.09133505076169968, + 0.06297721713781357, + 0.03907552361488342, + 0.039340790361166, + 0.18943418562412262, + 0.1955355405807495, + 0.06101146340370178, + 0.0895593911409378 + ], + "credit_map": [ + 0.015856418758630753, + 0.016627652570605278, + 0.013426665216684341, + 0.020241687074303627, + 0.03263156861066818, + 0.018427783623337746, + 0.023899996653199196, + 0.02069380134344101, + 0.006506538484245539, + 0.009083079174160957, + 0.005698295775800943, + 0.007584645878523588, + 0.012884378433227539, + 0.011838517151772976, + 0.0071620093658566475, + 0.009167857468128204, + 0.006392930634319782, + 0.012799728661775589, + 0.007221173029392958, + 0.013810350559651852, + 0.018154093995690346, + 0.01340828463435173, + 0.010810130275785923, + 0.008738361299037933, + 0.005365433171391487, + 0.00489449966698885, + 0.004667430184781551, + 0.005930944811552763, + 0.015067476779222488, + 0.011117693036794662, + 0.006885314825922251, + 0.009048431180417538, + 0.003554639173671603, + 0.00400535250082612, + 0.006817380897700787, + 0.004178301896899939, + 0.004444432444870472, + 0.006263658870011568, + 0.00415089912712574, + 0.005660857539623976, + 0.005837198346853256, + 0.003336843801662326, + 0.0034021781757473946, + 0.004324482288211584, + 0.004918286576867104, + 0.005043079145252705, + 0.006330927833914757, + 0.006147793959826231, + 0.053261883556842804, + 0.016299987211823463, + 0.03086923062801361, + 0.01392307411879301, + 0.01944241300225258, + 0.009774218313395977, + 0.01579166017472744, + 0.030071713030338287, + 0.027163930237293243, + 0.018229497596621513, + 0.04097486287355423, + 0.017827553674578667, + 0.028128335252404213, + 0.013373826630413532, + 0.02131115458905697, + 0.02852637879550457, + 0.006447586230933666, + 0.007181379944086075, + 0.008372859098017216, + 0.006730812136083841, + 0.009092527441680431, + 0.007090686354786158, + 0.007126717362552881, + 0.008968896232545376, + 0.013017362914979458, + 0.009996950626373291, + 0.015943383798003197, + 0.007207931485027075, + 0.01694115810096264, + 0.008830512873828411, + 0.010802754200994968, + 0.006819333881139755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.07006093859672546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.06212950870394707 + ], + "opposite_logprob_tokens": [ + -2.812129497528076 + ], + "logprob_reward": -0.06212950870394707, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06212950870394707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29102057218551636, + 0.11924010515213013, + 0.05811915174126625, + 0.057199977338314056, + 0.062301088124513626, + 0.05826307833194733, + 0.11450223624706268, + 0.11296891421079636, + 0.06957521289587021, + 0.05680966004729271 + ], + "credit_map": [ + 0.02893328107893467, + 0.03385502099990845, + 0.026183145120739937, + 0.03932596743106842, + 0.04103038087487221, + 0.035514798015356064, + 0.050378747284412384, + 0.035799238830804825, + 0.010790049098432064, + 0.013625448569655418, + 0.014127073809504509, + 0.01714487187564373, + 0.01783258281648159, + 0.018174665048718452, + 0.013891329057514668, + 0.013654077425599098, + 0.005426024552434683, + 0.009451779536902905, + 0.004804051015526056, + 0.007170156575739384, + 0.009695651941001415, + 0.010303606279194355, + 0.004308717790991068, + 0.006959161721169949, + 0.006108712870627642, + 0.005690302234143019, + 0.005857712123543024, + 0.008370086550712585, + 0.008785531856119633, + 0.008191597647964954, + 0.006406156346201897, + 0.007789883762598038, + 0.008425011299550533, + 0.007669059559702873, + 0.007351605221629143, + 0.007941621355712414, + 0.009575443342328072, + 0.008113794028759003, + 0.0064194281585514545, + 0.006805127486586571, + 0.010478424839675426, + 0.005338970571756363, + 0.0047654081135988235, + 0.005081533920019865, + 0.008254208602011204, + 0.007617061957716942, + 0.007918432354927063, + 0.008809037506580353, + 0.015504089184105396, + 0.010121376253664494, + 0.022687582299113274, + 0.00944962166249752, + 0.01728811301290989, + 0.008422253653407097, + 0.012173285707831383, + 0.018855910748243332, + 0.013090492226183414, + 0.010515080764889717, + 0.022148635238409042, + 0.010448566637933254, + 0.015572074800729752, + 0.008265756070613861, + 0.01570242829620838, + 0.017225883901119232, + 0.005313721485435963, + 0.0075382208451628685, + 0.010895569808781147, + 0.007601532153785229, + 0.011794676072895527, + 0.008257263340055943, + 0.00850139930844307, + 0.009672830812633038, + 0.0062409136444330215, + 0.006664350628852844, + 0.008458038792014122, + 0.004941883962601423, + 0.007314085494726896, + 0.012228358536958694, + 0.0056487214751541615, + 0.0053133065812289715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.06212950870394707 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.008645131252706051 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645131252706051, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645131252706051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1316918283700943, + 0.1017063707113266, + 0.1447419673204422, + 0.09853232651948929, + 0.08251640200614929, + 0.08275729417800903, + 0.06703056395053864, + 0.11022932082414627, + 0.12422217428684235, + 0.05657171830534935 + ], + "credit_map": [ + 0.01329770777374506, + 0.015236634761095047, + 0.010435220785439014, + 0.018005939200520515, + 0.0211821086704731, + 0.016327431425452232, + 0.021679630503058434, + 0.015527164563536644, + 0.007399903144687414, + 0.012966391630470753, + 0.010231991298496723, + 0.01167875062674284, + 0.018721358850598335, + 0.01378156989812851, + 0.012589436955749989, + 0.014336973428726196, + 0.013223151676356792, + 0.015639333054423332, + 0.010879790410399437, + 0.021411625668406487, + 0.03049078956246376, + 0.026326781138777733, + 0.011559020727872849, + 0.015211471356451511, + 0.00884249433875084, + 0.011917050927877426, + 0.010720541700720787, + 0.012296440079808235, + 0.019826842471957207, + 0.015602816827595234, + 0.008152053691446781, + 0.011174087412655354, + 0.010438243858516216, + 0.010113022290170193, + 0.01041320152580738, + 0.010503491386771202, + 0.010578811168670654, + 0.012504377402365208, + 0.008405251428484917, + 0.009560002014040947, + 0.014851147308945656, + 0.00894530862569809, + 0.009289862588047981, + 0.00802294909954071, + 0.011342828162014484, + 0.007502794731408358, + 0.012065454386174679, + 0.010736952535808086, + 0.011035357601940632, + 0.004447358660399914, + 0.007663739379495382, + 0.007356041111052036, + 0.009163575246930122, + 0.005238684359937906, + 0.009170183911919594, + 0.012955628335475922, + 0.022458957508206367, + 0.009254919365048409, + 0.01969684101641178, + 0.012742104940116405, + 0.0116150276735425, + 0.008173927664756775, + 0.011218630708754063, + 0.01506891194730997, + 0.022034918889403343, + 0.01435728371143341, + 0.01820470206439495, + 0.011853300966322422, + 0.015389963984489441, + 0.01047070324420929, + 0.015308084897696972, + 0.016603218391537666, + 0.009770388714969158, + 0.0061323209665715694, + 0.012454623356461525, + 0.0044324337504804134, + 0.007475307676941156, + 0.004929668270051479, + 0.007020089775323868, + 0.004356889519840479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.008645131252706051 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.00249301316216588 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.00249301316216588, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249301316216588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0873735100030899, + 0.02819063887000084, + 0.04721982777118683, + 0.055633753538131714, + 0.05380197986960411, + 0.1348249614238739, + 0.10596374422311783, + 0.1877681463956833, + 0.17080649733543396, + 0.1284169852733612 + ], + "credit_map": [ + 0.0097091905772686, + 0.009844711981713772, + 0.008363116532564163, + 0.011001954786479473, + 0.015470776706933975, + 0.009417830035090446, + 0.013447191566228867, + 0.010118726640939713, + 0.0027044459711760283, + 0.00298853637650609, + 0.00329969497397542, + 0.0029850334394723177, + 0.005556113086640835, + 0.0037258299998939037, + 0.003513143863528967, + 0.003417837666347623, + 0.0036595952697098255, + 0.005504587199538946, + 0.0035331575199961662, + 0.006923884153366089, + 0.009578256867825985, + 0.008273568004369736, + 0.0043404889293015, + 0.0054062833078205585, + 0.0034885138738900423, + 0.004952080547809601, + 0.004534448962658644, + 0.006547356955707073, + 0.010411521419882774, + 0.00895342230796814, + 0.006649173330515623, + 0.010097231715917587, + 0.004517060704529285, + 0.006760811433196068, + 0.0060054040513932705, + 0.007649255450814962, + 0.008853279985487461, + 0.00870052445679903, + 0.005615629721432924, + 0.005700009409338236, + 0.017717307433485985, + 0.012721056118607521, + 0.011987244710326195, + 0.014187658205628395, + 0.018377449363470078, + 0.015392362140119076, + 0.023417318239808083, + 0.021024556830525398, + 0.01713990606367588, + 0.007845286279916763, + 0.0199307631701231, + 0.008303839713335037, + 0.01682978682219982, + 0.005956626031547785, + 0.012794656679034233, + 0.0171628687530756, + 0.020233606919646263, + 0.014807320199906826, + 0.04061724245548248, + 0.020836524665355682, + 0.027680568397045135, + 0.01206416916102171, + 0.02386690303683281, + 0.02766178920865059, + 0.010225790552794933, + 0.013245404697954655, + 0.022728191688656807, + 0.02002589777112007, + 0.027404649183154106, + 0.020868612453341484, + 0.02536533586680889, + 0.030942590907216072, + 0.01675862818956375, + 0.01504280511289835, + 0.026565812528133392, + 0.009956804104149342, + 0.02201792225241661, + 0.012939861044287682, + 0.01744190976023674, + 0.007693234831094742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.00249301316216588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a high angle, begins, slowly and steadily moving closer to the central figure. In a quaint, elegant study, imbued with a scholarly atmosphere, the only light comes from the flickering flame of a bronze candlestick on the corner of the table. A slender, ancient strategist, approximately fifty years old, sits solemnly at a heavy rosewood desk. He wears a deep blue, wide-sleeved, finely textured silk robe, his long, jet-black hair meticulously tied atop his head with an antique wooden hairpin. The candlelight casts warm, pulsating shadows across his deep features, illuminating his high nose and the slight furrow of his brow, drawn in intense concentration. His right hand, steady and composed, holds a wolf-hair brush, the shaft softened by the warmth of his fingers. The tip, saturated with thick ink, moves smoothly across the spread of rice paper, leaving behind each powerful, vibrant, and inky Chinese character. As the camera moves forward extremely slowly, the picture gradually tightens into a close-up of his face. We can clearly see the wisdom flashing in his deep eyes and his total commitment to the strategy of his pen. His tightly pursed lips reveal his inner determination and deep consideration. The whole space is filled with a quiet, solemn and intellectually powerful atmosphere. It seems that the only sounds in the air are the faint crackling of the candlelight and the rustling of the pen tip across the paper.", + "seed": 27468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000122/group_9/index_6890_seed_27468_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": -0.05253855725750327, + "vlm_logprob": -0.05253855725750327, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05253855725750327, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.11584562808275223 + ], + "opposite_logprob_tokens": [ + -2.3658456802368164 + ], + "logprob_reward": -0.11584562808275223, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11584562808275223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1245003193616867, + 0.03775936737656593, + 0.21757297217845917, + 0.027273889631032944, + 0.04659957066178322, + 0.07320541143417358, + 0.07731419056653976, + 0.1959119439125061, + 0.114052914083004, + 0.08580942451953888 + ], + "credit_map": [ + 0.019410936161875725, + 0.02149021066725254, + 0.014150296337902546, + 0.009162240661680698, + 0.021680055186152458, + 0.017532113939523697, + 0.009495074860751629, + 0.011579392477869987, + 0.007536393590271473, + 0.005496467463672161, + 0.004174463450908661, + 0.0026018747594207525, + 0.005922327283769846, + 0.004682900384068489, + 0.0034114925656467676, + 0.0039334469474852085, + 0.028434397652745247, + 0.06176779419183731, + 0.017288707196712494, + 0.012962095439434052, + 0.01702188514173031, + 0.06263144314289093, + 0.009341124445199966, + 0.00812552496790886, + 0.005955238360911608, + 0.0038647660985589027, + 0.003509598085656762, + 0.0017663674661889672, + 0.00418851338326931, + 0.0034337902907282114, + 0.001998373307287693, + 0.0025572432205080986, + 0.006138835102319717, + 0.006151063367724419, + 0.00475277379155159, + 0.0040526981465518475, + 0.007581495679914951, + 0.0071930838748812675, + 0.004827036522328854, + 0.005902584176510572, + 0.006387252360582352, + 0.011398077011108398, + 0.008808932267129421, + 0.007467339280992746, + 0.009891950525343418, + 0.00753073301166296, + 0.009610427543520927, + 0.012110699899494648, + 0.025311972945928574, + 0.01392320916056633, + 0.00627759238705039, + 0.0072670672088861465, + 0.006005160976201296, + 0.006977062672376633, + 0.004591254983097315, + 0.006960873957723379, + 0.021434202790260315, + 0.0352274551987648, + 0.011940982192754745, + 0.01677289977669716, + 0.013843155466020107, + 0.047164201736450195, + 0.0208217054605484, + 0.02870735339820385, + 0.01361523475497961, + 0.019404850900173187, + 0.014578192494809628, + 0.02656053937971592, + 0.014388632960617542, + 0.01069210097193718, + 0.006163665559142828, + 0.008649695664644241, + 0.013749610632658005, + 0.02220914512872696, + 0.009781594388186932, + 0.007814149372279644, + 0.010746583342552185, + 0.008131040260195732, + 0.006116137839853764, + 0.007261157035827637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.11584562808275223 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.07902513444423676 + ], + "opposite_logprob_tokens": [ + -2.5790250301361084 + ], + "logprob_reward": -0.07902513444423676, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902513444423676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15935568511486053, + 0.047740936279296875, + 0.1724197119474411, + 0.056031517684459686, + 0.06385213881731033, + 0.05848228186368942, + 0.0669976994395256, + 0.12231084704399109, + 0.06479252129793167, + 0.1880166381597519 + ], + "credit_map": [ + 0.027147380635142326, + 0.0325218103826046, + 0.018438952043652534, + 0.013312420807778835, + 0.01947184093296528, + 0.022931143641471863, + 0.013185914605855942, + 0.012346229515969753, + 0.009094472043216228, + 0.00745546305552125, + 0.0061119319871068, + 0.003792576491832733, + 0.00721246050670743, + 0.005818434990942478, + 0.004148096311837435, + 0.004107498098164797, + 0.020655253902077675, + 0.05581676587462425, + 0.008710270747542381, + 0.006802636198699474, + 0.010803303681313992, + 0.05687102675437927, + 0.005857300944626331, + 0.006903151515871286, + 0.013962069526314735, + 0.00791184138506651, + 0.009574768133461475, + 0.005134365521371365, + 0.006361047271639109, + 0.0053087095730006695, + 0.003871629247441888, + 0.003907082136720419, + 0.011162450537085533, + 0.009230152703821659, + 0.008209691382944584, + 0.0052077798172831535, + 0.009535042569041252, + 0.006581554654985666, + 0.006469255313277245, + 0.007456211838871241, + 0.010672427713871002, + 0.009321856312453747, + 0.010385613888502121, + 0.006433932110667229, + 0.006536212284117937, + 0.0051979441195726395, + 0.004786025267094374, + 0.005148269236087799, + 0.018075143918395042, + 0.011200199835002422, + 0.006413353141397238, + 0.0068469843827188015, + 0.007580718025565147, + 0.00717098917812109, + 0.004341276828199625, + 0.005369032267481089, + 0.013342713937163353, + 0.028366155922412872, + 0.018126076087355614, + 0.009814843535423279, + 0.010040645487606525, + 0.029426956549286842, + 0.006008322350680828, + 0.007185125257819891, + 0.010866804979741573, + 0.00849193800240755, + 0.009835169650614262, + 0.014901864342391491, + 0.006604466121643782, + 0.006294882390648127, + 0.003626604564487934, + 0.004170787986367941, + 0.034375742077827454, + 0.03676488623023033, + 0.03954388201236725, + 0.025390980765223503, + 0.012747655622661114, + 0.014857020229101181, + 0.011896801181137562, + 0.012439675629138947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.07902513444423676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.043133050203323364 + ], + "opposite_logprob_tokens": [ + -3.168133020401001 + ], + "logprob_reward": -0.043133050203323364, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043133050203323364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1769084483385086, + 0.05487978830933571, + 0.20539994537830353, + 0.02669834904372692, + 0.030865898355841637, + 0.04190908372402191, + 0.07909222692251205, + 0.12952208518981934, + 0.18865783512592316, + 0.06606633216142654 + ], + "credit_map": [ + 0.03526698797941208, + 0.029039163142442703, + 0.020158952102065086, + 0.021450037136673927, + 0.023165661841630936, + 0.02042844519019127, + 0.016028329730033875, + 0.011370879597961903, + 0.011696656234562397, + 0.010373508557677269, + 0.005287512671202421, + 0.0038036101032048464, + 0.006095828488469124, + 0.008034469559788704, + 0.003348925616592169, + 0.006239280570298433, + 0.021110257133841515, + 0.0702708289027214, + 0.006432511378079653, + 0.009194101206958294, + 0.012018885463476181, + 0.07266072183847427, + 0.005879876669496298, + 0.007832769304513931, + 0.004792052321135998, + 0.0033931504003703594, + 0.004821665585041046, + 0.0030944710597395897, + 0.00301365926861763, + 0.002560492604970932, + 0.0022756082471460104, + 0.002747251419350505, + 0.005494170822203159, + 0.003910773433744907, + 0.005165640730410814, + 0.002926045563071966, + 0.004232192877680063, + 0.0027169811073690653, + 0.002785322954878211, + 0.0036347699351608753, + 0.007509833667427301, + 0.006850400008261204, + 0.008318300358951092, + 0.004171798471361399, + 0.004997542593628168, + 0.003830177243798971, + 0.002937998855486512, + 0.0032930339220911264, + 0.021177440881729126, + 0.0134419621899724, + 0.009244905784726143, + 0.007847542874515057, + 0.007265109568834305, + 0.00861066859215498, + 0.004361522383987904, + 0.007143077440559864, + 0.0148976631462574, + 0.032462432980537415, + 0.009213032200932503, + 0.008257458917796612, + 0.007021443918347359, + 0.03976918011903763, + 0.006403109058737755, + 0.011497771367430687, + 0.030456991866230965, + 0.025722548365592957, + 0.0203502606600523, + 0.06519103795289993, + 0.010287310928106308, + 0.016471771523356438, + 0.009286385960876942, + 0.010891522280871868, + 0.013286176137626171, + 0.013699330389499664, + 0.012004095129668713, + 0.006990211084485054, + 0.005554031580686569, + 0.0050383638590574265, + 0.004503711126744747, + 0.004990406800061464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.043133050203323364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.020589079707860947 + ], + "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": [ + 0.11997385323047638, + 0.047261953353881836, + 0.0884699672460556, + 0.04560580477118492, + 0.039622485637664795, + 0.057779647409915924, + 0.07686527818441391, + 0.11132839322090149, + 0.3507605493068695, + 0.06233213096857071 + ], + "credit_map": [ + 0.02481020987033844, + 0.019731678068637848, + 0.012321943417191505, + 0.011969474144279957, + 0.013304423540830612, + 0.01697354204952717, + 0.008660055696964264, + 0.01220251340419054, + 0.005961602088063955, + 0.006917704362422228, + 0.005321080796420574, + 0.006041313521564007, + 0.006519000511616468, + 0.005560924764722586, + 0.0049475799314677715, + 0.005992741324007511, + 0.016474299132823944, + 0.02545108087360859, + 0.005857323762029409, + 0.005086203571408987, + 0.007476284634321928, + 0.017604604363441467, + 0.005211514886468649, + 0.005308648571372032, + 0.008740048855543137, + 0.007564091123640537, + 0.005889137741178274, + 0.004208274185657501, + 0.007312951609492302, + 0.005547268316149712, + 0.003196494420990348, + 0.0031475357245653868, + 0.007054523099213839, + 0.005492280703037977, + 0.00538021931424737, + 0.003579192329198122, + 0.005867388565093279, + 0.004380363970994949, + 0.0039013901259750128, + 0.00396712264046073, + 0.010723122395575047, + 0.009086478501558304, + 0.008538445457816124, + 0.005945113021880388, + 0.007416606415063143, + 0.005199786741286516, + 0.005113727413117886, + 0.00575636699795723, + 0.010773245245218277, + 0.011319574899971485, + 0.0065048267133533955, + 0.009158805012702942, + 0.007757978513836861, + 0.010027098469436169, + 0.00798526406288147, + 0.013338482938706875, + 0.015285096131265163, + 0.023686593398451805, + 0.006142962723970413, + 0.008264540694653988, + 0.008402439765632153, + 0.028762422502040863, + 0.008568847551941872, + 0.012215481139719486, + 0.05383182317018509, + 0.0465802326798439, + 0.03543749079108238, + 0.1261521577835083, + 0.0165568757802248, + 0.02984185703098774, + 0.018991556018590927, + 0.023368556052446365, + 0.009406492114067078, + 0.011576838791370392, + 0.008086557500064373, + 0.0090878214687109, + 0.007461403496563435, + 0.0060787806287407875, + 0.005057972855865955, + 0.005576257593929768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.020589079707860947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central figure a slender, ancient strategist wearing a deep blue silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a study illuminated only by the flickering flame of a bronze candlestick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the strategist holding a wolf-hair brush and writing Chinese characters on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin as a medium shot from a high angle and slowly move closer to the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera eventually tighten into a close-up of the strategist'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": [ + -0.0040998938493430614 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040998938493430614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040998938493430614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04763946682214737, + 0.022613538429141045, + 0.10663797706365585, + 0.02584662102162838, + 0.022004200145602226, + 0.020407484844326973, + 0.0524820014834404, + 0.11099991947412491, + 0.4336203336715698, + 0.15774843096733093 + ], + "credit_map": [ + 0.011295824311673641, + 0.009888654574751854, + 0.00439497921615839, + 0.003288735868409276, + 0.005647104699164629, + 0.005726602394133806, + 0.003641912480816245, + 0.0037556560710072517, + 0.005547293461859226, + 0.003008943283930421, + 0.001758509548380971, + 0.0017652856186032295, + 0.003522005630657077, + 0.002597354119643569, + 0.002278973814100027, + 0.002135174348950386, + 0.008953721262514591, + 0.032037876546382904, + 0.006453356239944696, + 0.007083195261657238, + 0.007527635898441076, + 0.03519149869680405, + 0.0047790901735424995, + 0.004611600656062365, + 0.003659759648144245, + 0.003726776922121644, + 0.0032019345089793205, + 0.0032775322906672955, + 0.004317166283726692, + 0.0030500981956720352, + 0.002453297609463334, + 0.002160056494176388, + 0.002641826868057251, + 0.0029051965102553368, + 0.002572287106886506, + 0.0024633537977933884, + 0.0034775191452354193, + 0.0023974510841071606, + 0.0027116418350487947, + 0.0028349244967103004, + 0.0029992754571139812, + 0.00301371980458498, + 0.0018890590872615576, + 0.0021868874318897724, + 0.0035097189247608185, + 0.002542095957323909, + 0.0023453556932508945, + 0.0019213767955079675, + 0.011898813769221306, + 0.008287813514471054, + 0.0039411443285644054, + 0.0055728270672261715, + 0.005537920631468296, + 0.00655323313549161, + 0.004387014079838991, + 0.006303239148110151, + 0.011865627951920033, + 0.024219060316681862, + 0.006236152723431587, + 0.008284059353172779, + 0.009160722605884075, + 0.030430013313889503, + 0.008634372614324093, + 0.012169914320111275, + 0.06481814384460449, + 0.060176730155944824, + 0.03367098793387413, + 0.16625523567199707, + 0.025478864088654518, + 0.03574555367231369, + 0.02102578803896904, + 0.026449063792824745, + 0.020590906962752342, + 0.020461007952690125, + 0.025155389681458473, + 0.021875347942113876, + 0.023688161745667458, + 0.01456595491617918, + 0.013673802837729454, + 0.017737865447998047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0040998938493430614 + } + ], + "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/0000128/group_11/index_tensor([4780, 4780])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_11/index_tensor([4780, 4780])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..029ba27893a9ba17dd9b1ffd7b3a38e0990d3ec6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_11/index_tensor([4780, 4780])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow-motion panning shot, shot from a slightly upward angle, slowly approaches Zhuge Liang, standing at the bow of a massive ancient warship. He wears a wide-sleeved, dark blue silk robe, its hem intricately embroidered with auspicious cloud patterns in silver thread, which flutters gently in the river breeze. In one hand, he calmly holds a feather fan made of pure white goose feathers, its handle made of warm, dark wood. The other hand rests naturally behind his back, his posture erect and steady as a pine tree, perched on the wet wooden deck, gently heaving with the waves. His face is lean, with a neatly trimmed long beard. His deep, piercing eyes gaze at the distant spot where the water meets the sky, their expression filled with the wisdom and tranquility of strategic planning. In the background, the vast, turbulent river, a deep gray-green hue, is visible. The massive hull of the warship plows through the waves with unwavering determination, stirring up white spray. The sky is filled with thick, layered clouds, with only a faint, divinely inspired golden light piercing the horizon, framing his silhouette with a bright rim of light. The entire painting is characterized by a cold, solemn and majestic tone. The combination of the figure's composed posture and the vast natural landscape creates an epic, historically rich atmosphere.", + "seed": 28152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_11/index_4780_seed_28152_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": -0.13384713996201753, + "vlm_logprob": -0.13384713996201753, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.13384713996201753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.5826702117919922 + ], + "opposite_logprob_tokens": [ + -0.8326702117919922 + ], + "logprob_reward": -0.5826702117919922, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5826702117919922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18570882081985474, + 0.07881049811840057, + 0.0491817332804203, + 0.0788346529006958, + 0.07729095965623856, + 0.043829530477523804, + 0.04375891014933586, + 0.07771056145429611, + 0.3002587556838989, + 0.06461551785469055 + ], + "credit_map": [ + 0.026281051337718964, + 0.02577715553343296, + 0.024457281455397606, + 0.020247332751750946, + 0.024212412536144257, + 0.021773621439933777, + 0.02014116197824478, + 0.02281881682574749, + 0.01153377816081047, + 0.011710385791957378, + 0.00992159266024828, + 0.011542541906237602, + 0.010533913038671017, + 0.006983333267271519, + 0.009387128986418247, + 0.007197829894721508, + 0.00796803180128336, + 0.008721780963242054, + 0.004769379738718271, + 0.005316770635545254, + 0.0066548739559948444, + 0.004506123252213001, + 0.004997749347239733, + 0.006247025914490223, + 0.010333369486033916, + 0.008847599849104881, + 0.006041363347321749, + 0.00626731663942337, + 0.017298851162195206, + 0.009510916657745838, + 0.008729049004614353, + 0.01180618442595005, + 0.009610124863684177, + 0.005179899744689465, + 0.004738692659884691, + 0.0034527110401540995, + 0.019577952101826668, + 0.020257072523236275, + 0.006622422020882368, + 0.00785208959132433, + 0.0051728286780416965, + 0.005144401919096708, + 0.004433983471244574, + 0.0030963136814534664, + 0.005430432502180338, + 0.0041264803148806095, + 0.0098140062764287, + 0.006611083168536425, + 0.0068698786199092865, + 0.006289731711149216, + 0.004088844172656536, + 0.0028066791128367186, + 0.007481147535145283, + 0.004410041030496359, + 0.006107226479798555, + 0.005705362651497126, + 0.012965432368218899, + 0.01136019453406334, + 0.008146717213094234, + 0.007295502349734306, + 0.009659169241786003, + 0.005918802693486214, + 0.012171410955488682, + 0.01019332930445671, + 0.07937002927064896, + 0.02564225345849991, + 0.03045918047428131, + 0.018657350912690163, + 0.027019739151000977, + 0.03314662724733353, + 0.02686822973191738, + 0.0590953566133976, + 0.011095094494521618, + 0.00692936172708869, + 0.008262674324214458, + 0.007334322202950716, + 0.006292533595114946, + 0.006195214111357927, + 0.006634590215981007, + 0.011871733702719212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.5826702117919922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.05497235059738159 + ], + "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": [ + 0.28941500186920166, + 0.10377447307109833, + 0.05526687204837799, + 0.01921893283724785, + 0.051863569766283035, + 0.023279856890439987, + 0.04082665964961052, + 0.12690728902816772, + 0.18258821964263916, + 0.10685906559228897 + ], + "credit_map": [ + 0.04237353429198265, + 0.03559120371937752, + 0.02791575901210308, + 0.02009839378297329, + 0.06615850329399109, + 0.04887745529413223, + 0.02512715384364128, + 0.023273009806871414, + 0.01940125599503517, + 0.015480807982385159, + 0.011786445043981075, + 0.009383407421410084, + 0.01319078914821148, + 0.009477929212152958, + 0.011112366802990437, + 0.013941474258899689, + 0.007271191105246544, + 0.011569231748580933, + 0.006693549454212189, + 0.006090030074119568, + 0.008196810260415077, + 0.004939529579132795, + 0.004838825203478336, + 0.005667709279805422, + 0.0026919632218778133, + 0.0023959593381732702, + 0.001731586060486734, + 0.0020029342267662287, + 0.0040170419961214066, + 0.002013648860156536, + 0.0025492070708423853, + 0.0018165931105613708, + 0.004670248366892338, + 0.006096331402659416, + 0.0030076084658503532, + 0.0032768987584859133, + 0.014573352411389351, + 0.012547275982797146, + 0.004607402253895998, + 0.0030844546854496002, + 0.0024024704471230507, + 0.004182990174740553, + 0.0031019756570458412, + 0.0020243891049176455, + 0.003367387456819415, + 0.0030985071789473295, + 0.0025245859287679195, + 0.0025775518734008074, + 0.00885738991200924, + 0.006668576505035162, + 0.00463342247530818, + 0.0040015047416090965, + 0.0029770806431770325, + 0.003250840352848172, + 0.0048467484302818775, + 0.005591096356511116, + 0.016288259997963905, + 0.01916663907468319, + 0.009302827529609203, + 0.009720983915030956, + 0.02272181399166584, + 0.012238775379955769, + 0.018438631668686867, + 0.01902935653924942, + 0.04974284768104553, + 0.017286082729697227, + 0.01683671586215496, + 0.008659669198095798, + 0.019321827217936516, + 0.02416876330971718, + 0.01172629650682211, + 0.034846026450395584, + 0.023526189848780632, + 0.014756571501493454, + 0.013972596265375614, + 0.011235176585614681, + 0.00961714331060648, + 0.011543363332748413, + 0.00812035147100687, + 0.014087681658565998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.05497235059738159 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.00978719163686037 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.00978719163686037, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978719163686037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12918372452259064, + 0.0882926806807518, + 0.0861060619354248, + 0.06637681275606155, + 0.22878429293632507, + 0.0359266959130764, + 0.021467143669724464, + 0.18580427765846252, + 0.08784148097038269, + 0.07021677494049072 + ], + "credit_map": [ + 0.016653668135404587, + 0.031505417078733444, + 0.011474338360130787, + 0.011149240657687187, + 0.014458599500358105, + 0.01148965023458004, + 0.013375949114561081, + 0.019076861441135406, + 0.011055507697165012, + 0.014686775393784046, + 0.011147195473313332, + 0.012340402230620384, + 0.008609898388385773, + 0.009335123002529144, + 0.009258382953703403, + 0.011859394609928131, + 0.01003448199480772, + 0.018736330792307854, + 0.009352795779705048, + 0.01120754424482584, + 0.012410116381943226, + 0.008081590756773949, + 0.007122647482901812, + 0.00916055217385292, + 0.008433275856077671, + 0.008597154170274734, + 0.005361203569918871, + 0.004785197786986828, + 0.016143741086125374, + 0.01026812382042408, + 0.006373133510351181, + 0.006414984352886677, + 0.022532189264893532, + 0.025378696620464325, + 0.018668999895453453, + 0.01494623627513647, + 0.05083289369940758, + 0.05324552208185196, + 0.021308934316039085, + 0.02187083289027214, + 0.004780345130711794, + 0.0063764494843780994, + 0.0034115046728402376, + 0.003285860875621438, + 0.003757265629246831, + 0.006518010515719652, + 0.003777098376303911, + 0.004020160995423794, + 0.0031075309962034225, + 0.0026947157457470894, + 0.0019289426272735, + 0.0014678954612463713, + 0.003137152176350355, + 0.0020990099292248487, + 0.0032526592258363962, + 0.0037792364601045847, + 0.028606873005628586, + 0.02059428207576275, + 0.01214359700679779, + 0.01356121152639389, + 0.03159908950328827, + 0.02596515789628029, + 0.01818234845995903, + 0.035151731222867966, + 0.02322302758693695, + 0.009636052884161472, + 0.007850119844079018, + 0.006773966830223799, + 0.009025074541568756, + 0.008047782815992832, + 0.008217220194637775, + 0.015068236738443375, + 0.007680494338274002, + 0.00874602235853672, + 0.006984056904911995, + 0.00923830270767212, + 0.009336761198937893, + 0.010970912873744965, + 0.00834088958799839, + 0.008919339627027512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.00978719163686037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.01819257251918316 + ], + "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": [ + 0.17800036072731018, + 0.135341614484787, + 0.08272825926542282, + 0.038128577172756195, + 0.050676751881837845, + 0.05060930177569389, + 0.029427126049995422, + 0.23692147433757782, + 0.1077471598982811, + 0.09041937440633774 + ], + "credit_map": [ + 0.032435379922389984, + 0.02804163657128811, + 0.01864483579993248, + 0.021214639768004417, + 0.02873830869793892, + 0.01612117514014244, + 0.019220085814595222, + 0.013584312051534653, + 0.019218456000089645, + 0.01904069073498249, + 0.01959400437772274, + 0.0171467624604702, + 0.014983980916440487, + 0.012423558160662651, + 0.018784508109092712, + 0.014149653725326061, + 0.0086794039234519, + 0.01534200832247734, + 0.010337534360587597, + 0.01197985839098692, + 0.009520662017166615, + 0.006813547108322382, + 0.010592779144644737, + 0.009462472051382065, + 0.004345844965428114, + 0.00505704153329134, + 0.003544545266777277, + 0.004427616018801928, + 0.007453366182744503, + 0.004501994699239731, + 0.0050640953704714775, + 0.0037340740673244, + 0.004807545803487301, + 0.00715177645906806, + 0.004109654109925032, + 0.004006425850093365, + 0.010255277156829834, + 0.010149316862225533, + 0.0061099533922970295, + 0.004086801782250404, + 0.006342780776321888, + 0.0076706502586603165, + 0.003670969046652317, + 0.0029219388961791992, + 0.0055016446858644485, + 0.009536032564938068, + 0.0073088835924863815, + 0.00765640614554286, + 0.00433091726154089, + 0.0050651514902710915, + 0.0032559679821133614, + 0.005246847867965698, + 0.0032183171715587378, + 0.002781921299174428, + 0.0029090214520692825, + 0.002618982922285795, + 0.035739727318286896, + 0.02801043540239334, + 0.016300929710268974, + 0.01835782825946808, + 0.041545569896698, + 0.03261831775307655, + 0.024714095517992973, + 0.03963456302881241, + 0.028894590213894844, + 0.009541863575577736, + 0.010660829953849316, + 0.006171447224915028, + 0.011418323963880539, + 0.015114843845367432, + 0.009527844376862049, + 0.016417421400547028, + 0.012508945539593697, + 0.00967040378600359, + 0.009341267868876457, + 0.008564941585063934, + 0.014664067886769772, + 0.01675890013575554, + 0.008161529898643494, + 0.01074932236224413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.01819257251918316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.003613373264670372 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613373264670372, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613373264670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2563808262348175, + 0.056896988302469254, + 0.04356618598103523, + 0.03598573058843613, + 0.04761273413896561, + 0.19998688995838165, + 0.025386610999703407, + 0.030183346942067146, + 0.23177339136600494, + 0.07222732901573181 + ], + "credit_map": [ + 0.04869023710489273, + 0.03319854661822319, + 0.02994024008512497, + 0.021189123392105103, + 0.033296819776296616, + 0.025502784177660942, + 0.03263494372367859, + 0.031928107142448425, + 0.008430144749581814, + 0.009106159210205078, + 0.007816658355295658, + 0.006837733555585146, + 0.007237516809254885, + 0.004769055172801018, + 0.005429151002317667, + 0.00727057084441185, + 0.007291274610906839, + 0.0068261874839663506, + 0.005594085436314344, + 0.00311523606069386, + 0.006435920484364033, + 0.0037093835417181253, + 0.005214081611484289, + 0.0053800176829099655, + 0.005767705384641886, + 0.004031268879771233, + 0.0028670108877122402, + 0.001986089628189802, + 0.0065380423329770565, + 0.0051477993838489056, + 0.00346962152980268, + 0.006178190466016531, + 0.005726006347686052, + 0.004722853656858206, + 0.004073477350175381, + 0.0027294305618852377, + 0.010610392317175865, + 0.009525983594357967, + 0.004503169097006321, + 0.005721418187022209, + 0.032353177666664124, + 0.014482986181974411, + 0.024600770324468613, + 0.008650562725961208, + 0.032692283391952515, + 0.020421449095010757, + 0.021849948912858963, + 0.04493570700287819, + 0.002918486250564456, + 0.004270724020898342, + 0.002922286745160818, + 0.002609085990116, + 0.0036102510057389736, + 0.0022231703624129295, + 0.004003745038062334, + 0.0028288608882576227, + 0.006228909827768803, + 0.0038713000249117613, + 0.0027017742395401, + 0.0030644519720226526, + 0.002614726545289159, + 0.003662119386717677, + 0.002767962869256735, + 0.005272102076560259, + 0.06536894291639328, + 0.021389054134488106, + 0.02180643379688263, + 0.013353206217288971, + 0.018886445090174675, + 0.022681264206767082, + 0.01910974830389023, + 0.049178287386894226, + 0.012650671415030956, + 0.0065480670891702175, + 0.009747316129505634, + 0.005413425620645285, + 0.008011842146515846, + 0.008466925472021103, + 0.006831923499703407, + 0.014557155780494213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.003613373264670372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion panning shot, shot from a slightly upward angle, slowly approaches Zhuge Liang, standing at the bow of a massive ancient warship. He wears a wide-sleeved, dark blue silk robe, its hem intricately embroidered with auspicious cloud patterns in silver thread, which flutters gently in the river breeze. In one hand, he calmly holds a feather fan made of pure white goose feathers, its handle made of warm, dark wood. The other hand rests naturally behind his back, his posture erect and steady as a pine tree, perched on the wet wooden deck, gently heaving with the waves. His face is lean, with a neatly trimmed long beard. His deep, piercing eyes gaze at the distant spot where the water meets the sky, their expression filled with the wisdom and tranquility of strategic planning. In the background, the vast, turbulent river, a deep gray-green hue, is visible. The massive hull of the warship plows through the waves with unwavering determination, stirring up white spray. The sky is filled with thick, layered clouds, with only a faint, divinely inspired golden light piercing the horizon, framing his silhouette with a bright rim of light. The entire painting is characterized by a cold, solemn and majestic tone. The combination of the figure's composed posture and the vast natural landscape creates an epic, historically rich atmosphere.", + "seed": 28153, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_11/index_4780_seed_28153_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": -0.055141985416412354, + "vlm_logprob": -0.055141985416412354, + "vlm_yes_no_margin": 3.350000023841858, + "vlm_token_credit": -0.055141985416412354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.15340128540992737 + ], + "opposite_logprob_tokens": [ + -2.0284013748168945 + ], + "logprob_reward": -0.15340128540992737, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.15340128540992737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05503775179386139, + 0.07143528014421463, + 0.06131216883659363, + 0.05909610539674759, + 0.15715129673480988, + 0.27280110120773315, + 0.12071938812732697, + 0.047626838088035583, + 0.13203948736190796, + 0.022780640050768852 + ], + "credit_map": [ + 0.007305507082492113, + 0.006013036705553532, + 0.008373492397367954, + 0.005594927817583084, + 0.006620888132601976, + 0.006436005234718323, + 0.006048967130482197, + 0.00864492729306221, + 0.006214896682649851, + 0.013968237675726414, + 0.007070006802678108, + 0.009141001850366592, + 0.010775054804980755, + 0.008886095136404037, + 0.008809207938611507, + 0.0065707736648619175, + 0.006748449523001909, + 0.0063399504870176315, + 0.006092451512813568, + 0.00717600341886282, + 0.011103304103016853, + 0.007874056696891785, + 0.006808329373598099, + 0.009169621393084526, + 0.006360647734254599, + 0.013455627486109734, + 0.004471585154533386, + 0.009004971943795681, + 0.007764800917357206, + 0.006090291775763035, + 0.005974872969090939, + 0.005973303224891424, + 0.011712105944752693, + 0.011692707426846027, + 0.0169226061552763, + 0.010126035660505295, + 0.03515482321381569, + 0.03871922567486763, + 0.012859770096838474, + 0.019964033737778664, + 0.02719770558178425, + 0.02786986343562603, + 0.031064337119460106, + 0.01902952790260315, + 0.06792246550321579, + 0.034411489963531494, + 0.024005422368645668, + 0.04130028188228607, + 0.0138246463611722, + 0.02092655375599861, + 0.010953197255730629, + 0.011914546601474285, + 0.022327125072479248, + 0.017772285267710686, + 0.0077432869002223015, + 0.015257749706506729, + 0.005637700203806162, + 0.006069542840123177, + 0.006819807458668947, + 0.006042142398655415, + 0.006863035727292299, + 0.006326308939605951, + 0.004731164779514074, + 0.0051371329464018345, + 0.009505493566393852, + 0.0280892476439476, + 0.019187621772289276, + 0.016460543498396873, + 0.014244653284549713, + 0.017991717904806137, + 0.010791972279548645, + 0.01576824299991131, + 0.0025258464738726616, + 0.004683670587837696, + 0.0021533649414777756, + 0.0022450063843280077, + 0.0038485373370349407, + 0.002597354119643569, + 0.001468365779146552, + 0.003258496057242155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.15340128540992737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.07004804909229279 + ], + "opposite_logprob_tokens": [ + -2.6950480937957764 + ], + "logprob_reward": -0.07004804909229279, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004804909229279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15471647679805756, + 0.10343866795301437, + 0.0945611223578453, + 0.1549617052078247, + 0.17883282899856567, + 0.09391413629055023, + 0.10044265538454056, + 0.04028948396444321, + 0.044017255306243896, + 0.034825630486011505 + ], + "credit_map": [ + 0.020076263695955276, + 0.013653015717864037, + 0.029319416731595993, + 0.012511762790381908, + 0.018264228478074074, + 0.017659951001405716, + 0.024348925799131393, + 0.018882915377616882, + 0.009022767655551434, + 0.015120440162718296, + 0.008959070779383183, + 0.012409621849656105, + 0.019028298556804657, + 0.017282089218497276, + 0.01048686821013689, + 0.011129513382911682, + 0.01063145138323307, + 0.015271873213350773, + 0.009322541765868664, + 0.013766498304903507, + 0.013132955878973007, + 0.01241455040872097, + 0.010754057206213474, + 0.009267195127904415, + 0.014856378547847271, + 0.02788599021732807, + 0.02033851481974125, + 0.019519980996847153, + 0.025292927399277687, + 0.020814314484596252, + 0.015057909302413464, + 0.011195683851838112, + 0.012572657316923141, + 0.041584502905607224, + 0.011570527218282223, + 0.031194353476166725, + 0.025623846799135208, + 0.030032381415367126, + 0.013110174797475338, + 0.013144385069608688, + 0.00868276972323656, + 0.01208079606294632, + 0.010845409706234932, + 0.006838702131062746, + 0.019858330488204956, + 0.011587819084525108, + 0.009007332846522331, + 0.015012978576123714, + 0.010507222265005112, + 0.018422376364469528, + 0.010521677322685719, + 0.011155630461871624, + 0.017682194709777832, + 0.013477916829288006, + 0.007757016457617283, + 0.010918614454567432, + 0.004790087230503559, + 0.0037632975727319717, + 0.0041169351898133755, + 0.0039485725574195385, + 0.005995371378958225, + 0.005118993576616049, + 0.004476924892514944, + 0.008079302497208118, + 0.005383227486163378, + 0.0061635421589016914, + 0.004669416230171919, + 0.003132072975859046, + 0.009579510428011417, + 0.007128968834877014, + 0.002813993487507105, + 0.005146527197211981, + 0.004704453516751528, + 0.0032642600126564503, + 0.003392293816432357, + 0.002926324727013707, + 0.007383247371762991, + 0.004939471837133169, + 0.0036235011648386717, + 0.004592078272253275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.07004804909229279 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.011085351929068565 + ], + "opposite_logprob_tokens": [ + -4.511085510253906 + ], + "logprob_reward": -0.011085351929068565, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085351929068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25527921319007874, + 0.12627919018268585, + 0.06681521236896515, + 0.09357857704162598, + 0.15801319479942322, + 0.07360787689685822, + 0.10356854647397995, + 0.04368842393159866, + 0.03189912438392639, + 0.04727058857679367 + ], + "credit_map": [ + 0.034299641847610474, + 0.020188745111227036, + 0.03568282350897789, + 0.01829349622130394, + 0.04199376702308655, + 0.033794697374105453, + 0.030235078185796738, + 0.04079100117087364, + 0.016740424558520317, + 0.019883418455719948, + 0.011265991255640984, + 0.015653537586331367, + 0.020028935745358467, + 0.013763685710728168, + 0.01596854254603386, + 0.012974671088159084, + 0.0071814339607954025, + 0.008634915575385094, + 0.006635456345975399, + 0.00665663368999958, + 0.012666446156799793, + 0.01078296173363924, + 0.007313196547329426, + 0.006944173946976662, + 0.009267615154385567, + 0.010176469571888447, + 0.011930715292692184, + 0.007493933197110891, + 0.015424977988004684, + 0.014010836370289326, + 0.008221938274800777, + 0.017052102833986282, + 0.016311926767230034, + 0.02888789400458336, + 0.013717885129153728, + 0.026625584810972214, + 0.021934278309345245, + 0.023910824209451675, + 0.012162907980382442, + 0.014461902901530266, + 0.006496665999293327, + 0.011441674083471298, + 0.005396233405917883, + 0.008771616034209728, + 0.012989237904548645, + 0.009284097701311111, + 0.010525994934141636, + 0.008702362887561321, + 0.013473986648023129, + 0.019032422453165054, + 0.013288683257997036, + 0.008571207523345947, + 0.013388783670961857, + 0.012313980609178543, + 0.00823463685810566, + 0.015264860354363918, + 0.0055016023106873035, + 0.007291674148291349, + 0.00554979732260108, + 0.00477885315194726, + 0.005946872290223837, + 0.0048919375985860825, + 0.004497097805142403, + 0.005230591166764498, + 0.004539509303867817, + 0.003947167657315731, + 0.004792951047420502, + 0.0024682607036083937, + 0.006715331692248583, + 0.004308303352445364, + 0.001914426451548934, + 0.0032131734769791365, + 0.005521410144865513, + 0.011424546130001545, + 0.005192864220589399, + 0.00575770391151309, + 0.006609681528061628, + 0.005782718770205975, + 0.002462478820234537, + 0.004519189707934856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011085351929068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.02059982344508171 + ], + "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": [ + 0.1636352390050888, + 0.31943321228027344, + 0.08804918080568314, + 0.05957746505737305, + 0.04776693135499954, + 0.04937171936035156, + 0.0824173241853714, + 0.06534742563962936, + 0.09296870976686478, + 0.03143277391791344 + ], + "credit_map": [ + 0.025182008743286133, + 0.018288064748048782, + 0.018824569880962372, + 0.012060540728271008, + 0.025500530377030373, + 0.021926462650299072, + 0.019638177007436752, + 0.02221488580107689, + 0.04795606806874275, + 0.036966923624277115, + 0.02715306729078293, + 0.026316536590456963, + 0.06599816679954529, + 0.05248076841235161, + 0.03170220926403999, + 0.030859462916851044, + 0.009608849883079529, + 0.010910351760685444, + 0.007702762261033058, + 0.010334250517189503, + 0.013649356551468372, + 0.012628939002752304, + 0.007975108921527863, + 0.015239553526043892, + 0.0055457595735788345, + 0.006076935213059187, + 0.00801404844969511, + 0.005935242865234613, + 0.009211650118231773, + 0.008538905531167984, + 0.005690081976354122, + 0.010564842261373997, + 0.003191883908584714, + 0.009741265326738358, + 0.004009421914815903, + 0.006187978200614452, + 0.0071965293027460575, + 0.007630121428519487, + 0.004710743669420481, + 0.0050989906303584576, + 0.0030433463398367167, + 0.005650070495903492, + 0.003948931582272053, + 0.006538779474794865, + 0.005833420902490616, + 0.004685087129473686, + 0.011638120748102665, + 0.008033962920308113, + 0.008185535669326782, + 0.010439097881317139, + 0.01044478453695774, + 0.007557680830359459, + 0.014644324779510498, + 0.010472122579813004, + 0.007166641764342785, + 0.013507131487131119, + 0.006299696862697601, + 0.0091354139149189, + 0.008191203698515892, + 0.008682246319949627, + 0.008087264373898506, + 0.008171457797288895, + 0.009050583466887474, + 0.0077295610681176186, + 0.009919906035065651, + 0.012796681374311447, + 0.009381444193422794, + 0.006804872304201126, + 0.019453374668955803, + 0.015758326277136803, + 0.009292307309806347, + 0.009561797603964806, + 0.003733379067853093, + 0.006046697963029146, + 0.003104129806160927, + 0.004249580204486847, + 0.003989274613559246, + 0.0030683595687150955, + 0.0022067849058657885, + 0.005034570582211018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02059982344508171 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.020575417205691338 + ], + "opposite_logprob_tokens": [ + -3.895575523376465 + ], + "logprob_reward": -0.020575417205691338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575417205691338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17939817905426025, + 0.121589295566082, + 0.18994003534317017, + 0.03570878505706787, + 0.1272144615650177, + 0.19584491848945618, + 0.03754831850528717, + 0.041303105652332306, + 0.050750844180583954, + 0.020702002570033073 + ], + "credit_map": [ + 0.026557782664895058, + 0.015071805566549301, + 0.027231231331825256, + 0.013228953815996647, + 0.027934394776821136, + 0.022664109244942665, + 0.019864970818161964, + 0.026844924315810204, + 0.017056141048669815, + 0.009078394621610641, + 0.015711845830082893, + 0.006525967735797167, + 0.02456105500459671, + 0.02378510870039463, + 0.00937130767852068, + 0.015499474480748177, + 0.02555188722908497, + 0.016249869018793106, + 0.014933930709958076, + 0.013876385055482388, + 0.04750094190239906, + 0.02921915426850319, + 0.017803465947508812, + 0.024804404005408287, + 0.003836707677692175, + 0.0057290419936180115, + 0.005003657657653093, + 0.003931253682821989, + 0.0052795703522861, + 0.004437871277332306, + 0.0030966910999268293, + 0.00439399154856801, + 0.01087640319019556, + 0.013629496097564697, + 0.013811880722641945, + 0.008949738927185535, + 0.028941575437784195, + 0.0291867945343256, + 0.009854829870164394, + 0.011963737197220325, + 0.023274337872862816, + 0.02007356286048889, + 0.020352749153971672, + 0.010403493419289589, + 0.052907656878232956, + 0.027442917227745056, + 0.0171369519084692, + 0.0242532379925251, + 0.0042574056424200535, + 0.008334281854331493, + 0.0044205267913639545, + 0.005614323075860739, + 0.0038817620370537043, + 0.00449603796005249, + 0.003077285597100854, + 0.0034666964784264565, + 0.0056917183101177216, + 0.005216046702116728, + 0.00378586957231164, + 0.002312499564141035, + 0.009528890252113342, + 0.00572460051625967, + 0.0029826820828020573, + 0.006060799118131399, + 0.005008333828300238, + 0.006204935256391764, + 0.006530925631523132, + 0.004672829061746597, + 0.010533233173191547, + 0.006202902644872665, + 0.004103721585124731, + 0.007493960205465555, + 0.002947421744465828, + 0.0034590603318065405, + 0.002637701341882348, + 0.0020766614470630884, + 0.003810709808021784, + 0.0026455887127667665, + 0.001473804353736341, + 0.001651054946705699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.020575417205691338 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion panning shot, shot from a slightly upward angle, slowly approaches Zhuge Liang, standing at the bow of a massive ancient warship. He wears a wide-sleeved, dark blue silk robe, its hem intricately embroidered with auspicious cloud patterns in silver thread, which flutters gently in the river breeze. In one hand, he calmly holds a feather fan made of pure white goose feathers, its handle made of warm, dark wood. The other hand rests naturally behind his back, his posture erect and steady as a pine tree, perched on the wet wooden deck, gently heaving with the waves. His face is lean, with a neatly trimmed long beard. His deep, piercing eyes gaze at the distant spot where the water meets the sky, their expression filled with the wisdom and tranquility of strategic planning. In the background, the vast, turbulent river, a deep gray-green hue, is visible. The massive hull of the warship plows through the waves with unwavering determination, stirring up white spray. The sky is filled with thick, layered clouds, with only a faint, divinely inspired golden light piercing the horizon, framing his silhouette with a bright rim of light. The entire painting is characterized by a cold, solemn and majestic tone. The combination of the figure's composed posture and the vast natural landscape creates an epic, historically rich atmosphere.", + "seed": 28154, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_11/index_4780_seed_28154_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": -0.025936549343168736, + "vlm_logprob": -0.025936549343168736, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.025936549343168736, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.08564934879541397 + ], + "opposite_logprob_tokens": [ + -2.585649251937866 + ], + "logprob_reward": -0.08564934879541397, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08564934879541397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.101981520652771, + 0.07827326655387878, + 0.05946595221757889, + 0.06112901493906975, + 0.08197437226772308, + 0.09332429617643356, + 0.26289102435112, + 0.20659132301807404, + 0.020650766789913177, + 0.03371850773692131 + ], + "credit_map": [ + 0.017292480915784836, + 0.01320261787623167, + 0.01153748482465744, + 0.010977338068187237, + 0.014967688359320164, + 0.010792347602546215, + 0.01378652360290289, + 0.009425031952559948, + 0.013114729896187782, + 0.013697026297450066, + 0.007688986137509346, + 0.010531838051974773, + 0.008985236287117004, + 0.007318753283470869, + 0.010649088770151138, + 0.006287601310759783, + 0.006577628199011087, + 0.0063484846614301205, + 0.005904956720769405, + 0.009335189126431942, + 0.008882395923137665, + 0.005843671038746834, + 0.009672489948570728, + 0.006901132874190807, + 0.006242380011826754, + 0.005496680736541748, + 0.00968632660806179, + 0.010504268109798431, + 0.009142488241195679, + 0.005473257042467594, + 0.00869292113929987, + 0.005890686996281147, + 0.012467307969927788, + 0.01299348846077919, + 0.00838136300444603, + 0.014443540945649147, + 0.008299256674945354, + 0.009225545451045036, + 0.010806289501488209, + 0.005357575602829456, + 0.007175215054303408, + 0.009646447375416756, + 0.009663435630500317, + 0.011588353663682938, + 0.016337109729647636, + 0.010185759514570236, + 0.017091384157538414, + 0.011636583134531975, + 0.027798842638731003, + 0.032656025141477585, + 0.03121400997042656, + 0.02864505536854267, + 0.05556362122297287, + 0.034510549157857895, + 0.03146367520093918, + 0.021039223298430443, + 0.018753880634903908, + 0.019129592925310135, + 0.03978731110692024, + 0.031311944127082825, + 0.02867252007126808, + 0.025464922189712524, + 0.024965887889266014, + 0.01850523054599762, + 0.0029739250894635916, + 0.002357733901590109, + 0.003019214142113924, + 0.004033032339066267, + 0.0022028281819075346, + 0.0020205695182085037, + 0.0024051270447671413, + 0.0016383344773203135, + 0.0034231001045554876, + 0.00625659991055727, + 0.006452593486756086, + 0.0050171250477433205, + 0.002890547038987279, + 0.0026707653887569904, + 0.0039606960490345955, + 0.003047078149393201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.08564934879541397 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.018205566331744194 + ], + "opposite_logprob_tokens": [ + -4.018205642700195 + ], + "logprob_reward": -0.018205566331744194, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018205566331744194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1525164097547531, + 0.08160101622343063, + 0.0580897219479084, + 0.12612846493721008, + 0.2323404848575592, + 0.0871732085943222, + 0.05764681100845337, + 0.06576433032751083, + 0.09144619852304459, + 0.04729337990283966 + ], + "credit_map": [ + 0.014686132781207561, + 0.01932716555893421, + 0.01992720365524292, + 0.02209923230111599, + 0.017318839207291603, + 0.020113768056035042, + 0.022028572857379913, + 0.017015492543578148, + 0.011722707189619541, + 0.009301871992647648, + 0.010976937599480152, + 0.012627691961824894, + 0.010049154050648212, + 0.009206628426909447, + 0.011467361822724342, + 0.006248656660318375, + 0.00643058679997921, + 0.005498866550624371, + 0.008170188404619694, + 0.009156089276075363, + 0.009651487693190575, + 0.005518382415175438, + 0.00805314164608717, + 0.005610981024801731, + 0.01509407814592123, + 0.01110066007822752, + 0.020890777930617332, + 0.01743967831134796, + 0.014550242573022842, + 0.016871944069862366, + 0.017092928290367126, + 0.013088157400488853, + 0.037435393780469894, + 0.041193388402462006, + 0.014119425788521767, + 0.026649044826626778, + 0.03441740944981575, + 0.029046140611171722, + 0.031568415462970734, + 0.017911260947585106, + 0.009024844504892826, + 0.01185466255992651, + 0.009761474095284939, + 0.008285865187644958, + 0.013318110257387161, + 0.009676912799477577, + 0.014117036014795303, + 0.01113430131226778, + 0.004810340236872435, + 0.005439733155071735, + 0.007272274699062109, + 0.007526292931288481, + 0.012395189143717289, + 0.010856810957193375, + 0.005558748263865709, + 0.0037874202243983746, + 0.007294725626707077, + 0.006282496266067028, + 0.006655924022197723, + 0.007959499955177307, + 0.01138672698289156, + 0.011963230557739735, + 0.008051244542002678, + 0.006170476321130991, + 0.015079468488693237, + 0.01289124134927988, + 0.012732114642858505, + 0.008916009217500687, + 0.011106349527835846, + 0.010724887251853943, + 0.010974619537591934, + 0.00902150571346283, + 0.0056551177985966206, + 0.004905784968286753, + 0.005831842310726643, + 0.007334710098803043, + 0.005116191226989031, + 0.006417853757739067, + 0.005991871003061533, + 0.006040009204298258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.018205566331744194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.012545830570161343 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012545830570161343, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545830570161343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16325265169143677, + 0.06608358770608902, + 0.11674784868955612, + 0.14701956510543823, + 0.17456230521202087, + 0.06184275448322296, + 0.0813620388507843, + 0.0711781233549118, + 0.06858208030462265, + 0.04936906695365906 + ], + "credit_map": [ + 0.023627031594514847, + 0.019257158041000366, + 0.020388279110193253, + 0.01836046390235424, + 0.018160246312618256, + 0.022091872990131378, + 0.021998388692736626, + 0.019369205459952354, + 0.005248763132840395, + 0.008425138890743256, + 0.007726569660007954, + 0.009291051886975765, + 0.010601419024169445, + 0.007156250532716513, + 0.012165699154138565, + 0.00546869495883584, + 0.009886110201478004, + 0.012316268868744373, + 0.019015850499272346, + 0.014686382375657558, + 0.014504308812320232, + 0.008197550661861897, + 0.02568797394633293, + 0.012453403323888779, + 0.013940724544227123, + 0.012177160009741783, + 0.018596963956952095, + 0.025553716346621513, + 0.020705873146653175, + 0.013175113126635551, + 0.02807915024459362, + 0.014790862798690796, + 0.028283366933465004, + 0.027941545471549034, + 0.014074988663196564, + 0.01899545080959797, + 0.01815216615796089, + 0.019435370340943336, + 0.029306786134839058, + 0.018372615799307823, + 0.007560542784631252, + 0.0077324314042925835, + 0.004331101663410664, + 0.004275256767868996, + 0.013732640072703362, + 0.007886631414294243, + 0.009236706420779228, + 0.007087443023920059, + 0.009352938272058964, + 0.00974724069237709, + 0.015370819717645645, + 0.012579817324876785, + 0.008049314841628075, + 0.009009006433188915, + 0.010785233229398727, + 0.0064676678739488125, + 0.00790359452366829, + 0.0068861981853842735, + 0.0068910070694983006, + 0.008303876966238022, + 0.009788073599338531, + 0.009721014648675919, + 0.013236994855105877, + 0.008447360247373581, + 0.008272401988506317, + 0.007606059312820435, + 0.006531798746436834, + 0.00662787351757288, + 0.01498340256512165, + 0.010630931705236435, + 0.008040175773203373, + 0.0058894408866763115, + 0.005519325844943523, + 0.005338470451533794, + 0.005397892091423273, + 0.006682838778942823, + 0.004835858941078186, + 0.007088625803589821, + 0.00641251727938652, + 0.00809353869408369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.012545830570161343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.011084291152656078 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011084291152656078, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084291152656078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18475671112537384, + 0.09566759318113327, + 0.05900868773460388, + 0.06622984260320663, + 0.21613331139087677, + 0.043872542679309845, + 0.04402667656540871, + 0.07489998638629913, + 0.1298176646232605, + 0.08558693528175354 + ], + "credit_map": [ + 0.023141592741012573, + 0.026049690321087837, + 0.022082701325416565, + 0.02006988227367401, + 0.020990487188100815, + 0.030430378392338753, + 0.022861042991280556, + 0.019130952656269073, + 0.01262228935956955, + 0.009575962089002132, + 0.009423541836440563, + 0.009997555986046791, + 0.022659797221422195, + 0.009453951381146908, + 0.013951081782579422, + 0.007983422838151455, + 0.006700288970023394, + 0.006698623765259981, + 0.008540621027350426, + 0.007200534455478191, + 0.009198539890348911, + 0.005646948702633381, + 0.009828767739236355, + 0.005194371100515127, + 0.008475369773805141, + 0.006044668611139059, + 0.00849567074328661, + 0.00912142638117075, + 0.009111608378589153, + 0.007974212057888508, + 0.011486240662634373, + 0.0055206469260156155, + 0.03876817598938942, + 0.04007259011268616, + 0.01767984963953495, + 0.025227004662156105, + 0.023599490523338318, + 0.021167829632759094, + 0.02941780351102352, + 0.020200582221150398, + 0.0058073499239981174, + 0.004164479207247496, + 0.003726380877196789, + 0.0053089275024831295, + 0.00920805986970663, + 0.005104939453303814, + 0.0066760266199707985, + 0.003876381553709507, + 0.004430826287716627, + 0.0062416791915893555, + 0.005007410421967506, + 0.0036656970623880625, + 0.008776504546403885, + 0.008273906074464321, + 0.004112686030566692, + 0.0035179704427719116, + 0.00580027187243104, + 0.006727112922817469, + 0.009813166223466396, + 0.006552799604833126, + 0.014992302283644676, + 0.016458233818411827, + 0.008251575753092766, + 0.0063045332208275795, + 0.01338612474501133, + 0.015752529725432396, + 0.010674568824470043, + 0.009858256205916405, + 0.040667109191417694, + 0.021127885207533836, + 0.01103136409074068, + 0.007319836877286434, + 0.009051159024238586, + 0.0149369603022933, + 0.013043949380517006, + 0.007798566482961178, + 0.009147160686552525, + 0.009147806093096733, + 0.011192098259925842, + 0.011269246228039265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.011084291152656078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.0021977098658680916 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021977098658680916, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021977098658680916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1876266747713089, + 0.08435624092817307, + 0.06655770540237427, + 0.08445090055465698, + 0.08354010432958603, + 0.16579905152320862, + 0.08212246000766754, + 0.05576279014348984, + 0.108608178794384, + 0.08117592334747314 + ], + "credit_map": [ + 0.02803226187825203, + 0.01943434402346611, + 0.021370962262153625, + 0.026252802461385727, + 0.02472633123397827, + 0.029951581731438637, + 0.02386096492409706, + 0.013997415080666542, + 0.009166790172457695, + 0.007590623572468758, + 0.008975595235824585, + 0.01462068036198616, + 0.009071794338524342, + 0.012133000418543816, + 0.01502762921154499, + 0.00777012761682272, + 0.016091609373688698, + 0.01196394581347704, + 0.006644959095865488, + 0.005830078851431608, + 0.006291513331234455, + 0.00654252665117383, + 0.00903249066323042, + 0.004160585813224316, + 0.006512071006000042, + 0.006617321632802486, + 0.012853535823523998, + 0.013673246838152409, + 0.013280894607305527, + 0.006844877731055021, + 0.016552258282899857, + 0.008116691373288631, + 0.011033732444047928, + 0.011521881446242332, + 0.009841063059866428, + 0.00955086387693882, + 0.010321037843823433, + 0.010194658301770687, + 0.01274656131863594, + 0.008330305106937885, + 0.01582198590040207, + 0.017282485961914062, + 0.039204444736242294, + 0.024034051224589348, + 0.021100614219903946, + 0.01490477379411459, + 0.020388394594192505, + 0.013062299229204655, + 0.012651143595576286, + 0.011941896751523018, + 0.015125660225749016, + 0.008774359710514545, + 0.010541805997490883, + 0.007953990250825882, + 0.008235824294388294, + 0.006897781975567341, + 0.006194722838699818, + 0.006113279145210981, + 0.005738063249737024, + 0.007324701640754938, + 0.00897589884698391, + 0.007987767457962036, + 0.008099078200757504, + 0.00532927643507719, + 0.013496248051524162, + 0.015746233984827995, + 0.008766341023147106, + 0.010564824566245079, + 0.029190242290496826, + 0.013493279926478863, + 0.009249397553503513, + 0.00810161791741848, + 0.00967203639447689, + 0.007513110525906086, + 0.012162123806774616, + 0.014157659374177456, + 0.007850836031138897, + 0.008349323645234108, + 0.009656580165028572, + 0.011814253404736519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0021977098658680916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow-motion panning shot, shot from a slightly upward angle, slowly approaches Zhuge Liang, standing at the bow of a massive ancient warship. He wears a wide-sleeved, dark blue silk robe, its hem intricately embroidered with auspicious cloud patterns in silver thread, which flutters gently in the river breeze. In one hand, he calmly holds a feather fan made of pure white goose feathers, its handle made of warm, dark wood. The other hand rests naturally behind his back, his posture erect and steady as a pine tree, perched on the wet wooden deck, gently heaving with the waves. His face is lean, with a neatly trimmed long beard. His deep, piercing eyes gaze at the distant spot where the water meets the sky, their expression filled with the wisdom and tranquility of strategic planning. In the background, the vast, turbulent river, a deep gray-green hue, is visible. The massive hull of the warship plows through the waves with unwavering determination, stirring up white spray. The sky is filled with thick, layered clouds, with only a faint, divinely inspired golden light piercing the horizon, framing his silhouette with a bright rim of light. The entire painting is characterized by a cold, solemn and majestic tone. The combination of the figure's composed posture and the vast natural landscape creates an epic, historically rich atmosphere.", + "seed": 28155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_11/index_4780_seed_28155_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": -0.11028278274461627, + "vlm_logprob": -0.11028278274461627, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.11028278274461627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.4892428517341614 + ], + "opposite_logprob_tokens": [ + -0.9892428517341614 + ], + "logprob_reward": -0.4892428517341614, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4892428517341614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1657145619392395, + 0.06281749159097672, + 0.06551145017147064, + 0.032584235072135925, + 0.09946726262569427, + 0.04103866219520569, + 0.16814647614955902, + 0.13733099400997162, + 0.02994922176003456, + 0.19743958115577698 + ], + "credit_map": [ + 0.019924867898225784, + 0.02467036060988903, + 0.013435419648885727, + 0.021581217646598816, + 0.027728348970413208, + 0.014645302668213844, + 0.02455664984881878, + 0.019172415137290955, + 0.007312095258384943, + 0.008565033785998821, + 0.008917754516005516, + 0.00822538323700428, + 0.00715709337964654, + 0.00827490258961916, + 0.006587527226656675, + 0.007777705788612366, + 0.008335202932357788, + 0.006954913027584553, + 0.014059805311262608, + 0.007624578196555376, + 0.005432248115539551, + 0.010277613066136837, + 0.004814113490283489, + 0.008012985810637474, + 0.0035299728624522686, + 0.0038152921479195356, + 0.004148426000028849, + 0.0068282149732112885, + 0.00344068487174809, + 0.0033103767782449722, + 0.0037198038771748543, + 0.0037914677523076534, + 0.014049077406525612, + 0.010837928391993046, + 0.013173968531191349, + 0.01742134802043438, + 0.007315699942409992, + 0.01036609522998333, + 0.0058398121036589146, + 0.02046334184706211, + 0.005510108545422554, + 0.00555462995544076, + 0.004985696170479059, + 0.00653287535533309, + 0.005760561674833298, + 0.003818254452198744, + 0.004193098284304142, + 0.004683443810790777, + 0.03577079996466637, + 0.0321471132338047, + 0.02037283033132553, + 0.01595360040664673, + 0.020580697804689407, + 0.02198217064142227, + 0.011196318082511425, + 0.010142975486814976, + 0.0274212583899498, + 0.022879738360643387, + 0.020889490842819214, + 0.013183080591261387, + 0.014643050730228424, + 0.01783090829849243, + 0.011864349246025085, + 0.008619128726422787, + 0.004553245380520821, + 0.004966972395777702, + 0.005224816035479307, + 0.0035829723346978426, + 0.003776869038119912, + 0.0038483119569718838, + 0.001875856309197843, + 0.0021201791241765022, + 0.0742742121219635, + 0.026085257530212402, + 0.01029518898576498, + 0.0052436417900025845, + 0.008307083509862423, + 0.06460631638765335, + 0.0030079390853643417, + 0.005619953852146864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.4892428517341614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.029821116477251053 + ], + "opposite_logprob_tokens": [ + -3.5298211574554443 + ], + "logprob_reward": -0.029821116477251053, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821116477251053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3505324721336365, + 0.1416572630405426, + 0.06901586800813675, + 0.05835501104593277, + 0.05438859388232231, + 0.04227719083428383, + 0.055560290813446045, + 0.04274657741189003, + 0.0355616919696331, + 0.1499050408601761 + ], + "credit_map": [ + 0.05373648554086685, + 0.0547875352203846, + 0.042784322053194046, + 0.042754191905260086, + 0.030090181156992912, + 0.040999364107847214, + 0.03354506939649582, + 0.051835302263498306, + 0.023583365604281425, + 0.020894987508654594, + 0.017540033906698227, + 0.014905991964042187, + 0.014986693859100342, + 0.018265625461935997, + 0.011871268041431904, + 0.019609279930591583, + 0.007620581425726414, + 0.011613010428845882, + 0.008425483480095863, + 0.010300927795469761, + 0.00803754199296236, + 0.008135335519909859, + 0.0059583536349236965, + 0.008924633264541626, + 0.007814410142600536, + 0.006586998235434294, + 0.008343596011400223, + 0.010385329835116863, + 0.005621136631816626, + 0.006437959615141153, + 0.006718890741467476, + 0.006446685176342726, + 0.006695544347167015, + 0.006494732107967138, + 0.008420286700129509, + 0.010211283341050148, + 0.005917826201766729, + 0.006821564864367247, + 0.0036885093431919813, + 0.006138842552900314, + 0.007558529730886221, + 0.005860103759914637, + 0.005293951835483313, + 0.006528784055262804, + 0.0038052531890571117, + 0.004316599573940039, + 0.004326697438955307, + 0.004587269388139248, + 0.012602278962731361, + 0.009510744363069534, + 0.006117792800068855, + 0.005798675119876862, + 0.0063460818491876125, + 0.007565310690551996, + 0.003555776085704565, + 0.004063629079610109, + 0.0075461370870471, + 0.005901149939745665, + 0.00609674071893096, + 0.005377230234444141, + 0.003658630419522524, + 0.005070434883236885, + 0.0041236416436731815, + 0.004972615744918585, + 0.005714562255889177, + 0.0031880331225693226, + 0.0041306582279503345, + 0.0033597496803849936, + 0.007128128316253424, + 0.004782740958034992, + 0.003569307504221797, + 0.003688510973006487, + 0.042012233287096024, + 0.017044352367520332, + 0.00852127280086279, + 0.0060066282749176025, + 0.006507239304482937, + 0.05918555334210396, + 0.004358012694865465, + 0.006269748322665691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.029821116477251053 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.008646904490888119 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008646904490888119, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646904490888119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19543081521987915, + 0.09142397344112396, + 0.03217683732509613, + 0.042892757803201675, + 0.0615592785179615, + 0.05650798976421356, + 0.03440318629145622, + 0.04389428719878197, + 0.042326927185058594, + 0.3993840217590332 + ], + "credit_map": [ + 0.0190763957798481, + 0.02212531305849552, + 0.017979737371206284, + 0.027775608003139496, + 0.02482515014708042, + 0.02734428644180298, + 0.01901218295097351, + 0.03729213401675224, + 0.010447538457810879, + 0.009620629251003265, + 0.009763798676431179, + 0.011371023952960968, + 0.012645996175706387, + 0.010065797716379166, + 0.01445444393903017, + 0.013054735958576202, + 0.005017763935029507, + 0.0036574541591107845, + 0.004243416711688042, + 0.004305334761738777, + 0.004340713378041983, + 0.003941792994737625, + 0.0027732597663998604, + 0.0038970974273979664, + 0.007234600838273764, + 0.00438774935901165, + 0.004663912579417229, + 0.007288618944585323, + 0.0047167339362204075, + 0.004844942130148411, + 0.004776532296091318, + 0.004979660268872976, + 0.0054184263572096825, + 0.007064045872539282, + 0.009471657685935497, + 0.013252784498035908, + 0.006857699248939753, + 0.007419072091579437, + 0.004601866472512484, + 0.0074737221002578735, + 0.006838731002062559, + 0.007841592654585838, + 0.006035330705344677, + 0.006556034088134766, + 0.007017034105956554, + 0.006003408692777157, + 0.009788825176656246, + 0.006427029613405466, + 0.005683195777237415, + 0.005750963930040598, + 0.00633168313652277, + 0.00356148649007082, + 0.003749303286895156, + 0.004005069844424725, + 0.002808366436511278, + 0.0025131122674793005, + 0.00564158521592617, + 0.007519999053329229, + 0.00378049467690289, + 0.005435311701148748, + 0.007165974471718073, + 0.007052718661725521, + 0.0041155447252094746, + 0.0031826524063944817, + 0.007830368354916573, + 0.005827778019011021, + 0.006721354555338621, + 0.003002926241606474, + 0.005891226232051849, + 0.006674729287624359, + 0.0031468388624489307, + 0.003231701208278537, + 0.09422624111175537, + 0.050892651081085205, + 0.02433628961443901, + 0.015400716103613377, + 0.024536559358239174, + 0.15472756326198578, + 0.01597270369529724, + 0.019291244447231293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.008646904490888119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.01607704907655716 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.01607704907655716, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607704907655716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2821991741657257, + 0.11958572268486023, + 0.05733394995331764, + 0.05717088654637337, + 0.059753015637397766, + 0.06476147472858429, + 0.10517874360084534, + 0.10739132016897202, + 0.0547008290886879, + 0.09192495048046112 + ], + "credit_map": [ + 0.02545260824263096, + 0.025488028302788734, + 0.0268191359937191, + 0.04361111670732498, + 0.03560415282845497, + 0.03979038447141647, + 0.029486211016774178, + 0.05594753846526146, + 0.016437314450740814, + 0.01452295109629631, + 0.01786971651017666, + 0.018805209547281265, + 0.011436511762440205, + 0.015867531299591064, + 0.009576193057000637, + 0.015070280060172081, + 0.00739636505022645, + 0.00769537128508091, + 0.00784976314753294, + 0.007097108289599419, + 0.007383113261312246, + 0.008454042486846447, + 0.0053715151734650135, + 0.006086661480367184, + 0.006712072994560003, + 0.0073340474627912045, + 0.005626012571156025, + 0.01050951611250639, + 0.007338480092585087, + 0.0075736078433692455, + 0.005656391382217407, + 0.006420751102268696, + 0.0063729011453688145, + 0.008486802689731121, + 0.008952013216912746, + 0.010567648336291313, + 0.004801361821591854, + 0.006872883066534996, + 0.004736044444143772, + 0.008963353931903839, + 0.00824397150427103, + 0.009306836873292923, + 0.007747141178697348, + 0.010798385366797447, + 0.006691102869808674, + 0.00794209260493517, + 0.006998909171670675, + 0.007033024914562702, + 0.019691910594701767, + 0.01831074059009552, + 0.014946524985134602, + 0.010860634967684746, + 0.012641357257962227, + 0.013549779541790485, + 0.00812588445842266, + 0.007051902357488871, + 0.022343169897794724, + 0.01706632599234581, + 0.01714552752673626, + 0.011695696040987968, + 0.011296449229121208, + 0.012230133637785912, + 0.009367148391902447, + 0.006246862001717091, + 0.009472757577896118, + 0.007780631072819233, + 0.008603246882557869, + 0.004475680645555258, + 0.005739196203649044, + 0.008021503686904907, + 0.004717391915619373, + 0.005890418775379658, + 0.028200270608067513, + 0.020148757845163345, + 0.007014307659119368, + 0.006018502172082663, + 0.009275047108530998, + 0.010392860509455204, + 0.004862457979470491, + 0.006012733094394207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01607704907655716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow-motion pan that slowly approaches the character?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Zhuge Liang standing at the bow of a massive ancient warship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character wear a wide-sleeved dark blue silk robe with silver embroidered cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a white feather fan with a dark wooden handle in one hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a turbulent deep gray-green river under a sky with thick clouds and faint golden 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": [ + -0.007625991944223642 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625991944223642, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625991944223642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21450667083263397, + 0.0593174584209919, + 0.03246011584997177, + 0.03717821091413498, + 0.13428060710430145, + 0.04961538314819336, + 0.1657768189907074, + 0.04663810878992081, + 0.022253938019275665, + 0.23797260224819183 + ], + "credit_map": [ + 0.02694835513830185, + 0.028072362765669823, + 0.023248162120580673, + 0.03898320347070694, + 0.02628135308623314, + 0.02347428724169731, + 0.020252881571650505, + 0.02724607102572918, + 0.008868630975484848, + 0.007385823875665665, + 0.006894888821989298, + 0.0098396185785532, + 0.006616957951337099, + 0.007580925244837999, + 0.004936524201184511, + 0.007194093428552151, + 0.0038369472604244947, + 0.003843439044430852, + 0.005487021990120411, + 0.004555285908281803, + 0.003928510006517172, + 0.004375866614282131, + 0.0028264224529266357, + 0.0036066260654479265, + 0.004814562853425741, + 0.0034504858776926994, + 0.005337253678590059, + 0.008692112751305103, + 0.0031628869473934174, + 0.004127091728150845, + 0.003256065770983696, + 0.004337753169238567, + 0.017477972432971, + 0.014991577714681625, + 0.02058827318251133, + 0.03155219927430153, + 0.010728730820119381, + 0.016199281439185143, + 0.008382734842598438, + 0.014359856955707073, + 0.007460485678166151, + 0.008020195178687572, + 0.00665710587054491, + 0.007384358439594507, + 0.0037146825343370438, + 0.005977616645395756, + 0.003195089055225253, + 0.007205855567008257, + 0.03663318604230881, + 0.031254664063453674, + 0.022057577967643738, + 0.021371548995375633, + 0.013020056299865246, + 0.020281631499528885, + 0.008000423200428486, + 0.013157741166651249, + 0.008502248674631119, + 0.005750627722591162, + 0.007561877369880676, + 0.006670728791505098, + 0.004183921962976456, + 0.005224668886512518, + 0.0034622997045516968, + 0.0052817389369010925, + 0.0031066392548382282, + 0.0029467542190104723, + 0.0035421925131231546, + 0.0023585271555930376, + 0.0036680339835584164, + 0.003216315293684602, + 0.0019464638317003846, + 0.0014690146781504154, + 0.08311676979064941, + 0.05303236469626427, + 0.018271973356604576, + 0.014265929348766804, + 0.013021839782595634, + 0.037803515791893005, + 0.009554500691592693, + 0.008905739523470402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.007625991944223642 + } + ], + "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/0000128/group_3/index_tensor([8940, 8940])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_tensor([8940, 8940])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..23cacc422457a1afcb03b4542a5fed8a35060b3a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_tensor([8940, 8940])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, close-up, static shot, rendered in classic 16-bit game art, focuses on a powerful hand, trembling slightly with suppressed emotion, its knuckles white with exertion. The hand tightly grips a simple wooden brush, veins visible against the pixelated skin. To one side, a candle's pixelated flame flickers, casting a shaky, halftone-like shadow across the rough wooden table in a cycling of warm yellow and orange pixels, shrouding much of the scene in gloom. A brush tip, saturated with rich black ink, strokes across the rough, yellowed paper with resolute, heavy force, leaving behind a deep, powerful stroke. With an uncontrollable tremor, a heavy drop of ink falls from the tip, rapidly spreading across the paper in an ever-expanding mass of black pixels, like a silent oath. Although the camera mainly focuses on his hands, we can still catch a glimpse of the face of this leopard-headed, round-eyed strong man distorted by grief and anger. His furrowed brows and wide-open angry eyes, under the limited tones and clear pixel blocks, convey a solemn, resolute and tragic atmosphere full of retro gaming feel.", + "seed": 1795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_8940_seed_1795_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": -0.08420547135174275, + "vlm_logprob": -0.08420547135174275, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.08420547135174275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2852879464626312 + ], + "opposite_logprob_tokens": [ + -1.4102879762649536 + ], + "logprob_reward": -0.2852879464626312, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2852879464626312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18081225454807281, + 0.09176267683506012, + 0.03816487640142441, + 0.16113032400608063, + 0.04878094047307968, + 0.054161474108695984, + 0.09578924626111984, + 0.22793729603290558, + 0.06613627076148987, + 0.035324614495038986 + ], + "credit_map": [ + 0.025706855580210686, + 0.01887349784374237, + 0.019518684595823288, + 0.017720146104693413, + 0.03227633237838745, + 0.02150443010032177, + 0.020409565418958664, + 0.024802768602967262, + 0.0075821951031684875, + 0.010056051425635815, + 0.007937969639897346, + 0.015906017273664474, + 0.014062401838600636, + 0.011136321350932121, + 0.013997932896018028, + 0.011083785444498062, + 0.005812192801386118, + 0.0037774990778416395, + 0.005363768897950649, + 0.004073320422321558, + 0.003457740880548954, + 0.004336784593760967, + 0.006007236894220114, + 0.0053363339975476265, + 0.010428076609969139, + 0.024199288338422775, + 0.008327111601829529, + 0.008428862318396568, + 0.0424543097615242, + 0.03234127536416054, + 0.009214400313794613, + 0.025737004354596138, + 0.004557142965495586, + 0.004288341850042343, + 0.004842420108616352, + 0.006264895666390657, + 0.005755310878157616, + 0.009792432188987732, + 0.006041538901627064, + 0.007238863501697779, + 0.004189143888652325, + 0.00827016867697239, + 0.004575123079121113, + 0.003987470641732216, + 0.012238947674632072, + 0.0077848793007433414, + 0.0052619329653680325, + 0.007853814400732517, + 0.008161487057805061, + 0.01265175174921751, + 0.008034002967178822, + 0.004982166923582554, + 0.026122689247131348, + 0.012131277471780777, + 0.009120958857238293, + 0.014584919437766075, + 0.014154086820781231, + 0.02639967016875744, + 0.017326660454273224, + 0.019363168627023697, + 0.03135496750473976, + 0.031156709417700768, + 0.052916597574949265, + 0.035265445709228516, + 0.009948821738362312, + 0.0060538495890796185, + 0.006512819789350033, + 0.00794968381524086, + 0.009781572967767715, + 0.006829648744314909, + 0.008616299368441105, + 0.01044357568025589, + 0.0024431191850453615, + 0.0045255632139742374, + 0.002239297842606902, + 0.0019483466167002916, + 0.008633160963654518, + 0.006272547412663698, + 0.003646198194473982, + 0.00561638455837965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.2852879464626312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306282848119736 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306282848119736, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306282848119736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15010188519954681, + 0.16321595013141632, + 0.04958668723702431, + 0.06723081320524216, + 0.06480613350868225, + 0.030676523223519325, + 0.19630734622478485, + 0.08338771015405655, + 0.07032875716686249, + 0.12435819953680038 + ], + "credit_map": [ + 0.010047866962850094, + 0.015738772228360176, + 0.013810113072395325, + 0.01482099387794733, + 0.026383979246020317, + 0.028964273631572723, + 0.0165551844984293, + 0.02378068119287491, + 0.011210543103516102, + 0.01856204681098461, + 0.014185568317770958, + 0.030329136177897453, + 0.02360566519200802, + 0.02046298049390316, + 0.02147725038230419, + 0.023382743820548058, + 0.005070524290204048, + 0.007316658739000559, + 0.005448942072689533, + 0.004497510846704245, + 0.008707137778401375, + 0.006344064604490995, + 0.005551116541028023, + 0.006650727242231369, + 0.006463498342782259, + 0.009628850035369396, + 0.0052490439265966415, + 0.004380513448268175, + 0.014540493488311768, + 0.012074136175215244, + 0.005266337189823389, + 0.009627936407923698, + 0.005290354136377573, + 0.006636500358581543, + 0.007940830662846565, + 0.007908389903604984, + 0.010100947692990303, + 0.008569281548261642, + 0.009277203120291233, + 0.009082615375518799, + 0.0023498423397541046, + 0.0040198336355388165, + 0.0025297440588474274, + 0.0025723855942487717, + 0.005834505427628756, + 0.006291004829108715, + 0.0023588668555021286, + 0.004720337688922882, + 0.017151353880763054, + 0.025036973878741264, + 0.019577063620090485, + 0.010568016208708286, + 0.051695216447114944, + 0.027433037757873535, + 0.015493905171751976, + 0.02935177832841873, + 0.005240784492343664, + 0.010374922305345535, + 0.006733907386660576, + 0.007249833084642887, + 0.011843506246805191, + 0.012119116261601448, + 0.01796700805425644, + 0.011858627200126648, + 0.0050112842582166195, + 0.008921331726014614, + 0.007381908595561981, + 0.007153703831136227, + 0.012285049073398113, + 0.010143829509615898, + 0.007849171757698059, + 0.011582468636333942, + 0.016056643798947334, + 0.015011824667453766, + 0.008525090292096138, + 0.007818816229701042, + 0.024794412776827812, + 0.01810893602669239, + 0.013168994337320328, + 0.020873472094535828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.04306282848119736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2697552442550659, + 0.18264807760715485, + 0.11745868623256683, + 0.06312792003154755, + 0.10297329723834991, + 0.036569640040397644, + 0.017511053010821342, + 0.12583409249782562, + 0.048004619777202606, + 0.03611743077635765 + ], + "credit_map": [ + 0.028300251811742783, + 0.031135467812418938, + 0.02462097257375717, + 0.027710188180208206, + 0.04655900225043297, + 0.0426667146384716, + 0.02796911634504795, + 0.0407935306429863, + 0.025587785989046097, + 0.0143790477886796, + 0.017543580383062363, + 0.01824191026389599, + 0.02208477444946766, + 0.04145064949989319, + 0.01698911190032959, + 0.026371212676167488, + 0.021019166335463524, + 0.013890527188777924, + 0.016153497621417046, + 0.010995897464454174, + 0.011859736405313015, + 0.010979615151882172, + 0.015524494461715221, + 0.017035748809576035, + 0.005940738134086132, + 0.009518599137663841, + 0.004981483798474073, + 0.0033603892661631107, + 0.014841822907328606, + 0.011319860816001892, + 0.003551374189555645, + 0.00961365271359682, + 0.014976161532104015, + 0.009990083053708076, + 0.01076715998351574, + 0.014962568879127502, + 0.008756422437727451, + 0.01733538880944252, + 0.011473177000880241, + 0.01471233181655407, + 0.0038561439141631126, + 0.004335121251642704, + 0.004127271939069033, + 0.0027135012205690145, + 0.004452003166079521, + 0.008277280256152153, + 0.00323217804543674, + 0.005576140712946653, + 0.0017625426407903433, + 0.001972100930288434, + 0.00173323811031878, + 0.002044825116172433, + 0.0032447383273392916, + 0.0021782529074698687, + 0.0019908559042960405, + 0.0025844986084848642, + 0.009853966534137726, + 0.014546086080372334, + 0.01108995359390974, + 0.009642395190894604, + 0.015442360192537308, + 0.021164800971746445, + 0.025427481159567833, + 0.01866704784333706, + 0.00462001608684659, + 0.005260018166154623, + 0.005051059648394585, + 0.005324789322912693, + 0.008068526163697243, + 0.005379634443670511, + 0.006912426557391882, + 0.007388152647763491, + 0.005952707026153803, + 0.003804711392149329, + 0.003974908962845802, + 0.0038730145897716284, + 0.003770706709474325, + 0.0030816702637821436, + 0.005780678242444992, + 0.005879031959921122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.005947155877947807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809179365634918 + ], + "opposite_logprob_tokens": [ + -3.2880918979644775 + ], + "logprob_reward": -0.03809179365634918, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809179365634918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.138654887676239, + 0.03964567556977272, + 0.01982300728559494, + 0.025982286781072617, + 0.035559870302677155, + 0.031220434233546257, + 0.15400224924087524, + 0.024052338674664497, + 0.15556403994560242, + 0.37549519538879395 + ], + "credit_map": [ + 0.009520310908555984, + 0.017296766862273216, + 0.010063666850328445, + 0.014394951052963734, + 0.027671728283166885, + 0.02387371100485325, + 0.013917578384280205, + 0.02191617153584957, + 0.0031951989512890577, + 0.004262829665094614, + 0.0039508468471467495, + 0.006377215962857008, + 0.0056815240532159805, + 0.005880949553102255, + 0.004878973588347435, + 0.005418137647211552, + 0.002914167707785964, + 0.002233991166576743, + 0.0024030019994825125, + 0.0017983869183808565, + 0.0018501868471503258, + 0.0031249106395989656, + 0.0023880929220467806, + 0.0031102688517421484, + 0.002726868260651827, + 0.003162794979289174, + 0.0024373356718569994, + 0.0021585775539278984, + 0.005314442329108715, + 0.004014320205897093, + 0.0023744162172079086, + 0.0037935313303023577, + 0.0023767140228301287, + 0.0037322198040783405, + 0.0020857157651335, + 0.003561490448191762, + 0.005254917778074741, + 0.008865597657859325, + 0.003721656510606408, + 0.005961558315902948, + 0.002453773282468319, + 0.003428549040108919, + 0.002660976257175207, + 0.0020755173172801733, + 0.004333857912570238, + 0.008410238660871983, + 0.0026595438830554485, + 0.005197975318878889, + 0.013180266134440899, + 0.018361497670412064, + 0.010249095968902111, + 0.015568850561976433, + 0.029369892552495003, + 0.018649814650416374, + 0.025191910564899445, + 0.023430921137332916, + 0.001408449956215918, + 0.0029127229936420918, + 0.002074133139103651, + 0.0018929785583168268, + 0.0038906836416572332, + 0.003691951045766473, + 0.004447043873369694, + 0.00373437418602407, + 0.017056193202733994, + 0.016588514670729637, + 0.015320234932005405, + 0.01868024654686451, + 0.022232314571738243, + 0.01776561327278614, + 0.02246854268014431, + 0.025452397763729095, + 0.04452986642718315, + 0.04369792342185974, + 0.02825065515935421, + 0.028475815430283546, + 0.07460499554872513, + 0.05300471559166908, + 0.04190608859062195, + 0.06102513521909714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.03809179365634918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863763228058815 + ], + "opposite_logprob_tokens": [ + -3.048637628555298 + ], + "logprob_reward": -0.04863763228058815, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863763228058815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19654488563537598, + 0.11414121836423874, + 0.03428039327263832, + 0.10445249825716019, + 0.04319891333580017, + 0.02971954457461834, + 0.10946875810623169, + 0.06383389979600906, + 0.05970176309347153, + 0.24465805292129517 + ], + "credit_map": [ + 0.013179796747863293, + 0.02546200342476368, + 0.015092620626091957, + 0.019084002822637558, + 0.04023154079914093, + 0.034112464636564255, + 0.018792741000652313, + 0.030589722096920013, + 0.006958777084946632, + 0.009215597994625568, + 0.007451800163835287, + 0.016011588275432587, + 0.018483726307749748, + 0.027410166338086128, + 0.012143386527895927, + 0.016466174274683, + 0.0034663304686546326, + 0.004643992520868778, + 0.004150372929871082, + 0.002436792477965355, + 0.005206590052694082, + 0.0056117745116353035, + 0.003780221799388528, + 0.004984320607036352, + 0.006927190814167261, + 0.01312814000993967, + 0.007796711288392544, + 0.006970221642404795, + 0.026158379390835762, + 0.01905987784266472, + 0.009644793346524239, + 0.014767185784876347, + 0.004801118280738592, + 0.004351775627583265, + 0.0039000047836452723, + 0.0036062048748135567, + 0.004917757119983435, + 0.010426911525428295, + 0.004111850634217262, + 0.007083291187882423, + 0.004481453914195299, + 0.0042076027020812035, + 0.0033858006354421377, + 0.0027688113041222095, + 0.002600176027044654, + 0.004326494410634041, + 0.0032811248674988747, + 0.004668079782277346, + 0.00945332832634449, + 0.00974206905812025, + 0.0067512160167098045, + 0.010763040743768215, + 0.02161434106528759, + 0.015096423216164112, + 0.019229458644986153, + 0.016818873584270477, + 0.006480811163783073, + 0.00861691776663065, + 0.0061848652549088, + 0.004797331523150206, + 0.009601619094610214, + 0.01056336984038353, + 0.00857779011130333, + 0.00901119876652956, + 0.006679000798612833, + 0.006153774913400412, + 0.005112589802592993, + 0.005151010118424892, + 0.009294962510466576, + 0.011396246030926704, + 0.006098742596805096, + 0.009815431199967861, + 0.02849428355693817, + 0.027496637776494026, + 0.01865798979997635, + 0.02033417299389839, + 0.04682855308055878, + 0.03376810625195503, + 0.029385993257164955, + 0.03969232738018036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04863763228058815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up, static shot, rendered in classic 16-bit game art, focuses on a powerful hand, trembling slightly with suppressed emotion, its knuckles white with exertion. The hand tightly grips a simple wooden brush, veins visible against the pixelated skin. To one side, a candle's pixelated flame flickers, casting a shaky, halftone-like shadow across the rough wooden table in a cycling of warm yellow and orange pixels, shrouding much of the scene in gloom. A brush tip, saturated with rich black ink, strokes across the rough, yellowed paper with resolute, heavy force, leaving behind a deep, powerful stroke. With an uncontrollable tremor, a heavy drop of ink falls from the tip, rapidly spreading across the paper in an ever-expanding mass of black pixels, like a silent oath. Although the camera mainly focuses on his hands, we can still catch a glimpse of the face of this leopard-headed, round-eyed strong man distorted by grief and anger. His furrowed brows and wide-open angry eyes, under the limited tones and clear pixel blocks, convey a solemn, resolute and tragic atmosphere full of retro gaming feel.", + "seed": 1796, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_8940_seed_1796_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": -0.12869127690792084, + "vlm_logprob": -0.12869127690792084, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.12869127690792084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3199833929538727 + ], + "opposite_logprob_tokens": [ + -1.3199833631515503 + ], + "logprob_reward": -0.3199833929538727, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3199833929538727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22006167471408844, + 0.0512232705950737, + 0.048545461148023605, + 0.07699380815029144, + 0.08188529312610626, + 0.029880104586482048, + 0.09953642636537552, + 0.198574036359787, + 0.10063407570123672, + 0.092665895819664 + ], + "credit_map": [ + 0.015310531482100487, + 0.0265982486307621, + 0.021724920719861984, + 0.018205728381872177, + 0.03295071795582771, + 0.034868136048316956, + 0.04231980815529823, + 0.028083588927984238, + 0.004641890991479158, + 0.0036086025647819042, + 0.005455855280160904, + 0.0027370380703359842, + 0.0055695753544569016, + 0.010712084360420704, + 0.011980142444372177, + 0.006518082693219185, + 0.005302209407091141, + 0.003919022157788277, + 0.004575394093990326, + 0.002388749271631241, + 0.006977371405810118, + 0.007442225702106953, + 0.010093292221426964, + 0.007847197353839874, + 0.005019560921937227, + 0.005284082610160112, + 0.005312491673976183, + 0.003172906581312418, + 0.006968958768993616, + 0.006428595166653395, + 0.024650229141116142, + 0.0201569851487875, + 0.009267541579902172, + 0.006545664742588997, + 0.010158327408134937, + 0.005681728944182396, + 0.01009896956384182, + 0.016443727537989616, + 0.010844483971595764, + 0.012844845652580261, + 0.004196816589683294, + 0.0023280992172658443, + 0.004511625971645117, + 0.0027310785371810198, + 0.005816903430968523, + 0.0037820737343281507, + 0.0025752682704478502, + 0.003938239999115467, + 0.008587597869336605, + 0.010547208599746227, + 0.017997248098254204, + 0.008176625706255436, + 0.012081573717296124, + 0.01586570404469967, + 0.012960991822183132, + 0.013319476507604122, + 0.011506523005664349, + 0.013024147599935532, + 0.028052395209670067, + 0.00862203724682331, + 0.0090841855853796, + 0.06694282591342926, + 0.009876581840217113, + 0.05146534740924835, + 0.012360937893390656, + 0.007198060397058725, + 0.022356409579515457, + 0.005996570456773043, + 0.009469515644013882, + 0.021048223599791527, + 0.006888930220156908, + 0.015315431170165539, + 0.007482187822461128, + 0.011515828780829906, + 0.008995125070214272, + 0.00975748524069786, + 0.015611504204571247, + 0.014576408080756664, + 0.013989792205393314, + 0.010737568140029907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.3199833929538727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496106669306755 + ], + "opposite_logprob_tokens": [ + -2.9299609661102295 + ], + "logprob_reward": -0.05496106669306755, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496106669306755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18851439654827118, + 0.12215929478406906, + 0.10813459753990173, + 0.09877055138349533, + 0.13217051327228546, + 0.035187993198633194, + 0.036727938801050186, + 0.11447841674089432, + 0.059208258986473083, + 0.10464794933795929 + ], + "credit_map": [ + 0.01363946869969368, + 0.023208796977996826, + 0.024460334330797195, + 0.01629847101867199, + 0.030388828366994858, + 0.03160596266388893, + 0.031124483793973923, + 0.01778806559741497, + 0.010070267133414745, + 0.014417585916817188, + 0.015198882669210434, + 0.009816461242735386, + 0.021226268261671066, + 0.02394123375415802, + 0.012104349210858345, + 0.015384255908429623, + 0.014016353525221348, + 0.010925968177616596, + 0.013564426451921463, + 0.00669155502691865, + 0.015435670502483845, + 0.020912081003189087, + 0.015188407152891159, + 0.011400147341191769, + 0.006934708915650845, + 0.011046632193028927, + 0.009900952689349651, + 0.005557116121053696, + 0.017162322998046875, + 0.0192312840372324, + 0.016878409311175346, + 0.012059135362505913, + 0.01133432425558567, + 0.015331911854445934, + 0.012738002464175224, + 0.0176997147500515, + 0.01630535162985325, + 0.017597394064068794, + 0.02327774651348591, + 0.01788608729839325, + 0.004925142973661423, + 0.0025897419545799494, + 0.003808951238170266, + 0.002369978930801153, + 0.005617079325020313, + 0.004669586196541786, + 0.004840613342821598, + 0.0063669029623270035, + 0.0040379734709858894, + 0.004510711412876844, + 0.0035270382650196552, + 0.0028422335162758827, + 0.00581013597548008, + 0.0052212136797606945, + 0.005344047676771879, + 0.005434588063508272, + 0.006342001259326935, + 0.008999812416732311, + 0.018036389723420143, + 0.006461845710873604, + 0.009751974605023861, + 0.03221232444047928, + 0.0086415009573102, + 0.024032585322856903, + 0.006778533570468426, + 0.004265506751835346, + 0.01142814103513956, + 0.004100866615772247, + 0.00636334577575326, + 0.013879416510462761, + 0.003672634484246373, + 0.008719819597899914, + 0.008997584693133831, + 0.014338875189423561, + 0.007691365201026201, + 0.011899834498763084, + 0.015624551102519035, + 0.015943603590130806, + 0.014862616546452045, + 0.015289531089365482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.05496106669306755 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370121121406555 + ], + "opposite_logprob_tokens": [ + -3.408701181411743 + ], + "logprob_reward": -0.03370121121406555, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370121121406555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1876850128173828, + 0.04658675193786621, + 0.2208779901266098, + 0.050670139491558075, + 0.05072107911109924, + 0.03041038103401661, + 0.0746804028749466, + 0.2443496286869049, + 0.03524015471339226, + 0.058778442442417145 + ], + "credit_map": [ + 0.011039265431463718, + 0.02612478844821453, + 0.01716010831296444, + 0.012061575427651405, + 0.04560486227273941, + 0.03344468027353287, + 0.02311975695192814, + 0.01912998966872692, + 0.0033653893042355776, + 0.006374228745698929, + 0.0053784726187586784, + 0.003739164210855961, + 0.008710848167538643, + 0.007921353913843632, + 0.005428567063063383, + 0.005668732803314924, + 0.02224651537835598, + 0.018130609765648842, + 0.031612008810043335, + 0.015964290127158165, + 0.01898367516696453, + 0.02390945702791214, + 0.05479499325156212, + 0.035236455500125885, + 0.010274741798639297, + 0.005968163255602121, + 0.00754691194742918, + 0.004210037179291248, + 0.004497496411204338, + 0.006393996998667717, + 0.005233149975538254, + 0.006545646581798792, + 0.007025496568530798, + 0.005123928189277649, + 0.006418562959879637, + 0.0033055474050343037, + 0.00588923878967762, + 0.007959661073982716, + 0.0066869622096419334, + 0.008311689831316471, + 0.003104468109086156, + 0.0019627742003649473, + 0.0026717965956777334, + 0.002004543086513877, + 0.0043482170440256596, + 0.0051518091931939125, + 0.004421746823936701, + 0.006745029706507921, + 0.008205248974263668, + 0.0070566036738455296, + 0.011151004582643509, + 0.005410416983067989, + 0.00950249657034874, + 0.011647844687104225, + 0.010743961669504642, + 0.010962831787765026, + 0.01451144553720951, + 0.016872741281986237, + 0.04343239590525627, + 0.009181937202811241, + 0.0109345568343997, + 0.07526625692844391, + 0.008571522310376167, + 0.06557876616716385, + 0.004239632282406092, + 0.0030290300492197275, + 0.00816289521753788, + 0.002265307819470763, + 0.0038274861872196198, + 0.005966200027614832, + 0.0031208479776978493, + 0.004628758877515793, + 0.005133810453116894, + 0.006963601801544428, + 0.004529098514467478, + 0.005700874608010054, + 0.011799956671893597, + 0.008848508819937706, + 0.007736139930784702, + 0.008066457696259022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.03370121121406555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494854226708412 + ], + "opposite_logprob_tokens": [ + -2.929948568344116 + ], + "logprob_reward": -0.05494854226708412, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494854226708412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2017461359500885, + 0.05072624608874321, + 0.11143960058689117, + 0.04158232733607292, + 0.08348692953586578, + 0.03684236854314804, + 0.05344332382082939, + 0.18996597826480865, + 0.09102858603000641, + 0.13973857462406158 + ], + "credit_map": [ + 0.014784499071538448, + 0.022491877898573875, + 0.020605068653821945, + 0.017483603209257126, + 0.04595961794257164, + 0.039181679487228394, + 0.023317161947488785, + 0.01792263612151146, + 0.0033180115278810263, + 0.007296557072550058, + 0.004065025597810745, + 0.003539386671036482, + 0.007847379893064499, + 0.00948471948504448, + 0.007312029600143433, + 0.007863135077059269, + 0.010132834315299988, + 0.007551294285804033, + 0.010938622057437897, + 0.005555336829274893, + 0.0126434825360775, + 0.01279193814843893, + 0.0325034037232399, + 0.01932268589735031, + 0.005344593431800604, + 0.0035873304586857557, + 0.007591845467686653, + 0.0034623700194060802, + 0.005607339553534985, + 0.006116047967225313, + 0.005051748361438513, + 0.004821052309125662, + 0.010218670591711998, + 0.010000655427575111, + 0.00870890449732542, + 0.012542251497507095, + 0.008561435155570507, + 0.008830060251057148, + 0.014871245250105858, + 0.009753705002367496, + 0.00360082252882421, + 0.0038670070935040712, + 0.002957172691822052, + 0.0030959362629801035, + 0.005516196135431528, + 0.006478691007941961, + 0.0058006108738482, + 0.005525933112949133, + 0.004073733929544687, + 0.0063208299688994884, + 0.007685400079935789, + 0.005575367249548435, + 0.007859324105083942, + 0.007621556520462036, + 0.0073630004189908504, + 0.006944108288735151, + 0.006308197043836117, + 0.00756469601765275, + 0.016995634883642197, + 0.005725666880607605, + 0.016870703548192978, + 0.06957503408193588, + 0.011393360793590546, + 0.05553267151117325, + 0.0053651658818125725, + 0.006488381419330835, + 0.007917446084320545, + 0.0033191931433975697, + 0.006935948505997658, + 0.029544666409492493, + 0.005399032030254602, + 0.026058748364448547, + 0.011046315543353558, + 0.023012865334749222, + 0.014031454920768738, + 0.014149674214422703, + 0.020633820444345474, + 0.022414004430174828, + 0.016750972718000412, + 0.017699459567666054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.05494854226708412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986217141151428 + ], + "opposite_logprob_tokens": [ + -1.804862141609192 + ], + "logprob_reward": -0.17986217141151428, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986217141151428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08122868835926056, + 0.05551103129982948, + 0.09763478487730026, + 0.05316317453980446, + 0.21993477642536163, + 0.06529176235198975, + 0.052346087992191315, + 0.2649965286254883, + 0.030293520539999008, + 0.07959960401058197 + ], + "credit_map": [ + 0.0064012170769274235, + 0.009707561694085598, + 0.013511182740330696, + 0.007204854860901833, + 0.013643452897667885, + 0.012212323024868965, + 0.010222422890365124, + 0.008325671777129173, + 0.003155843820422888, + 0.007223749067634344, + 0.003985635936260223, + 0.002949211047962308, + 0.009938565082848072, + 0.011229395866394043, + 0.01010236144065857, + 0.006926269736140966, + 0.00653233053162694, + 0.006323796231299639, + 0.006219936069101095, + 0.004599435720592737, + 0.014251619577407837, + 0.013115506619215012, + 0.027469824999570847, + 0.019122334197163582, + 0.005257442593574524, + 0.005955932196229696, + 0.0048996517434716225, + 0.003309240797534585, + 0.007431806996464729, + 0.008036408573389053, + 0.011561792343854904, + 0.006710898596793413, + 0.013186509720981121, + 0.02210022509098053, + 0.01925707422196865, + 0.022265281528234482, + 0.034713443368673325, + 0.045027848333120346, + 0.03467552736401558, + 0.028708890080451965, + 0.004197490867227316, + 0.004341613966971636, + 0.00423905486240983, + 0.0038472756277769804, + 0.01201849989593029, + 0.010472098365426064, + 0.010381188243627548, + 0.0157945416867733, + 0.005334616173058748, + 0.005343940109014511, + 0.0063874670304358006, + 0.003373937914147973, + 0.007481282576918602, + 0.008027471601963043, + 0.008550948463380337, + 0.00784642156213522, + 0.011132637970149517, + 0.011937214992940426, + 0.022127635776996613, + 0.009182875975966454, + 0.012703104875981808, + 0.09653834998607635, + 0.013147376477718353, + 0.08822733908891678, + 0.0018680752255022526, + 0.001604480086825788, + 0.0028776649851351976, + 0.0010593782644718885, + 0.002476524794474244, + 0.01059280801564455, + 0.0013585977721959352, + 0.008455989882349968, + 0.006758766248822212, + 0.01137487217783928, + 0.00652562128379941, + 0.008257363922894001, + 0.014390276744961739, + 0.013054260984063148, + 0.009800268337130547, + 0.009438179433345795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.17986217141151428 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up, static shot, rendered in classic 16-bit game art, focuses on a powerful hand, trembling slightly with suppressed emotion, its knuckles white with exertion. The hand tightly grips a simple wooden brush, veins visible against the pixelated skin. To one side, a candle's pixelated flame flickers, casting a shaky, halftone-like shadow across the rough wooden table in a cycling of warm yellow and orange pixels, shrouding much of the scene in gloom. A brush tip, saturated with rich black ink, strokes across the rough, yellowed paper with resolute, heavy force, leaving behind a deep, powerful stroke. With an uncontrollable tremor, a heavy drop of ink falls from the tip, rapidly spreading across the paper in an ever-expanding mass of black pixels, like a silent oath. Although the camera mainly focuses on his hands, we can still catch a glimpse of the face of this leopard-headed, round-eyed strong man distorted by grief and anger. His furrowed brows and wide-open angry eyes, under the limited tones and clear pixel blocks, convey a solemn, resolute and tragic atmosphere full of retro gaming feel.", + "seed": 1797, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_8940_seed_1797_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": -0.10056143822148442, + "vlm_logprob": -0.10056143822148442, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.10056143822148442, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2070099264383316 + ], + "opposite_logprob_tokens": [ + -1.7070099115371704 + ], + "logprob_reward": -0.2070099264383316, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2070099264383316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17697174847126007, + 0.09407836943864822, + 0.05743573606014252, + 0.09156008809804916, + 0.14475710690021515, + 0.055940672755241394, + 0.25005489587783813, + 0.03865142911672592, + 0.06045828387141228, + 0.030091658234596252 + ], + "credit_map": [ + 0.024689335376024246, + 0.03213024139404297, + 0.03184111416339874, + 0.019956817850470543, + 0.013468077406287193, + 0.01560931745916605, + 0.015355418436229229, + 0.023921431973576546, + 0.008674819022417068, + 0.02597779780626297, + 0.008076947182416916, + 0.004566284827888012, + 0.013535607606172562, + 0.009534014388918877, + 0.006094456650316715, + 0.017618447542190552, + 0.008164994418621063, + 0.008140886202454567, + 0.006291908212006092, + 0.004259358625859022, + 0.008405148051679134, + 0.00949412677437067, + 0.0047901226207613945, + 0.007889190688729286, + 0.009043293073773384, + 0.010838363319635391, + 0.010134114883840084, + 0.009182693436741829, + 0.012726990506052971, + 0.013753410428762436, + 0.012999048456549644, + 0.012882179580628872, + 0.017973706126213074, + 0.0316082239151001, + 0.014028959907591343, + 0.010070692747831345, + 0.017092352733016014, + 0.018956726416945457, + 0.012105393223464489, + 0.022921057417988777, + 0.007058270275592804, + 0.0072523909620940685, + 0.006872037425637245, + 0.0054613216780126095, + 0.00565376179292798, + 0.007938181050121784, + 0.008684025146067142, + 0.007020686287432909, + 0.028549799695611, + 0.034584496170282364, + 0.0415504053235054, + 0.024697771295905113, + 0.025611773133277893, + 0.031661469489336014, + 0.0318235419690609, + 0.031575653702020645, + 0.007145746145397425, + 0.004928478505462408, + 0.005260784178972244, + 0.0034086096566170454, + 0.00406227633357048, + 0.0036782072857022285, + 0.004410954657942057, + 0.005756369326263666, + 0.004616981837898493, + 0.008012320846319199, + 0.00695473188534379, + 0.005668368190526962, + 0.00746874837204814, + 0.015691056847572327, + 0.005594543181359768, + 0.006451532244682312, + 0.0033364053815603256, + 0.003755028359591961, + 0.002959418809041381, + 0.0026960421819239855, + 0.005960599984973669, + 0.004264663904905319, + 0.0029268746729940176, + 0.004192626103758812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.2070099264383316 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062058594077825546 + ], + "opposite_logprob_tokens": [ + -2.812058687210083 + ], + "logprob_reward": -0.062058594077825546, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062058594077825546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33085817098617554, + 0.11584365367889404, + 0.037573304027318954, + 0.09642001241445541, + 0.09763237833976746, + 0.04105215147137642, + 0.11211880296468735, + 0.056210774928331375, + 0.07742144167423248, + 0.03486931696534157 + ], + "credit_map": [ + 0.05540485307574272, + 0.05377814173698425, + 0.058865927159786224, + 0.04187322407960892, + 0.023194342851638794, + 0.027060870081186295, + 0.029806086793541908, + 0.04087473079562187, + 0.017507892102003098, + 0.021511636674404144, + 0.019250378012657166, + 0.008188629522919655, + 0.010456598363816738, + 0.010566161945462227, + 0.010944576933979988, + 0.0174177847802639, + 0.0050235316157341, + 0.004922318737953901, + 0.006563861388713121, + 0.0028500561602413654, + 0.004962176084518433, + 0.004220517817884684, + 0.003147645154967904, + 0.005883194971829653, + 0.010283676907420158, + 0.010954484343528748, + 0.00961865670979023, + 0.006600752007216215, + 0.019431455060839653, + 0.0192966740578413, + 0.007961229421198368, + 0.012273087166249752, + 0.010567856021225452, + 0.012012568302452564, + 0.013745644129812717, + 0.006968529894948006, + 0.019242726266384125, + 0.016005264595150948, + 0.010669327341020107, + 0.008420457132160664, + 0.0051119644194841385, + 0.004812250845134258, + 0.005833533126860857, + 0.0031221795361489058, + 0.0058895982801914215, + 0.008032583631575108, + 0.004145777318626642, + 0.004104264546185732, + 0.009369983337819576, + 0.014324305579066277, + 0.014238269999623299, + 0.009422583505511284, + 0.011622152291238308, + 0.020729752257466316, + 0.015197120606899261, + 0.017214635387063026, + 0.007203505840152502, + 0.006538055837154388, + 0.009976151399314404, + 0.006203570403158665, + 0.008066880516707897, + 0.006883528549224138, + 0.005879448261111975, + 0.005459635052829981, + 0.00930223148316145, + 0.009708407334983349, + 0.011618257500231266, + 0.006843667943030596, + 0.00910399854183197, + 0.015950480476021767, + 0.0069263833574950695, + 0.00796801783144474, + 0.005404279101639986, + 0.0040063499473035336, + 0.006237307097762823, + 0.0024127152282744646, + 0.00516530079767108, + 0.004720661323517561, + 0.0027949295472353697, + 0.004127772059291601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.062058594077825546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252490285784006 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252490285784006, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252490285784006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2482277899980545, + 0.09842856228351593, + 0.07464924454689026, + 0.08745645731687546, + 0.11055897176265717, + 0.042223233729600906, + 0.20881496369838715, + 0.05056456848978996, + 0.04239951819181442, + 0.03667670115828514 + ], + "credit_map": [ + 0.0329069122672081, + 0.04021152853965759, + 0.04039061814546585, + 0.03055923618376255, + 0.0200358759611845, + 0.020134633406996727, + 0.02413414977490902, + 0.039854854345321655, + 0.013987026177346706, + 0.01592858135700226, + 0.01978524774312973, + 0.005216066725552082, + 0.011713451705873013, + 0.010467272251844406, + 0.00831065233796835, + 0.013020259328186512, + 0.008854323998093605, + 0.009158339351415634, + 0.009245053865015507, + 0.006031431723386049, + 0.0114930160343647, + 0.012132697738707066, + 0.007305544801056385, + 0.010428834706544876, + 0.008653851225972176, + 0.011376386508345604, + 0.010702191852033138, + 0.007805601228028536, + 0.010295971296727657, + 0.01243201456964016, + 0.013009457848966122, + 0.01318097673356533, + 0.02272975631058216, + 0.014836667105555534, + 0.023747004568576813, + 0.00869110506027937, + 0.011447915807366371, + 0.011456229723989964, + 0.00896601751446724, + 0.008684273809194565, + 0.006444499362260103, + 0.005054084118455648, + 0.006978250574320555, + 0.0035316862631589174, + 0.004927164409309626, + 0.0051124305464327335, + 0.005388200748711824, + 0.004786916542798281, + 0.029565081000328064, + 0.03133124113082886, + 0.03546139970421791, + 0.016262196004390717, + 0.02076309733092785, + 0.022698169574141502, + 0.024908754974603653, + 0.027825042605400085, + 0.007608484476804733, + 0.006329753436148167, + 0.007854755036532879, + 0.006929904222488403, + 0.0043833451345562935, + 0.004796264693140984, + 0.006356008816510439, + 0.006306050345301628, + 0.007033127825707197, + 0.005413273815065622, + 0.010888653807342052, + 0.0033249524421989918, + 0.0027449994813650846, + 0.005310169421136379, + 0.00343914027325809, + 0.004245202988386154, + 0.00506999809294939, + 0.004635221790522337, + 0.006133159156888723, + 0.0034180248621851206, + 0.004466566257178783, + 0.003909496124833822, + 0.003683796152472496, + 0.005360436160117388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.005252490285784006 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864104092121124 + ], + "opposite_logprob_tokens": [ + -3.0486409664154053 + ], + "logprob_reward": -0.04864104092121124, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864104092121124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2712893784046173, + 0.15400095283985138, + 0.04414701834321022, + 0.11621151119470596, + 0.22730690240859985, + 0.042552292346954346, + 0.04822627454996109, + 0.024483084678649902, + 0.0334818996489048, + 0.03830072656273842 + ], + "credit_map": [ + 0.03185349330306053, + 0.04143082723021507, + 0.030811628326773643, + 0.023756491020321846, + 0.03613750636577606, + 0.04841084033250809, + 0.023015040904283524, + 0.03587355092167854, + 0.013649219647049904, + 0.055351629853248596, + 0.008671198971569538, + 0.008146461099386215, + 0.012427803128957748, + 0.01233354490250349, + 0.009175065904855728, + 0.034246016293764114, + 0.005152353551238775, + 0.009649327024817467, + 0.0047426773235201836, + 0.003391413716599345, + 0.004433991387486458, + 0.0048041269183158875, + 0.004656936042010784, + 0.007316193543374538, + 0.011916499584913254, + 0.013172395527362823, + 0.011035907082259655, + 0.011849906295537949, + 0.01921040378510952, + 0.0212669987231493, + 0.013017091900110245, + 0.014742307364940643, + 0.027724266052246094, + 0.05982141196727753, + 0.018689142540097237, + 0.010933419689536095, + 0.02916891500353813, + 0.023939508944749832, + 0.015380390919744968, + 0.04164984077215195, + 0.00669514387845993, + 0.004851780366152525, + 0.008051429875195026, + 0.004509247839450836, + 0.005341795738786459, + 0.004773654509335756, + 0.004554850049316883, + 0.003774387761950493, + 0.007808800786733627, + 0.006202970165759325, + 0.007119323126971722, + 0.0033738419879227877, + 0.005392234772443771, + 0.007174286525696516, + 0.004143653903156519, + 0.007011158857494593, + 0.0040844641625881195, + 0.0031755196396261454, + 0.0034021090250462294, + 0.0022336721885949373, + 0.0023886223789304495, + 0.0026426417753100395, + 0.003143226495012641, + 0.0034128304105252028, + 0.004807347431778908, + 0.003505622036755085, + 0.003789691487327218, + 0.003541900310665369, + 0.004342662636190653, + 0.005471888463944197, + 0.004196071531623602, + 0.0038267162162810564, + 0.003272070549428463, + 0.005191996693611145, + 0.004608611576259136, + 0.002626356203109026, + 0.007857228629291058, + 0.007268862333148718, + 0.0031045181676745415, + 0.004371083341538906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04864104092121124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17984513938426971 + ], + "opposite_logprob_tokens": [ + -1.8048450946807861 + ], + "logprob_reward": -0.17984513938426971, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17984513938426971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23919086158275604, + 0.19765233993530273, + 0.06607610732316971, + 0.11244460195302963, + 0.11463294923305511, + 0.08730903267860413, + 0.04580177739262581, + 0.024805402383208275, + 0.07903465628623962, + 0.033052265644073486 + ], + "credit_map": [ + 0.03599529713392258, + 0.03102315030992031, + 0.0470384918153286, + 0.027178553864359856, + 0.01989973895251751, + 0.027679771184921265, + 0.017543505877256393, + 0.03283235803246498, + 0.03317207470536232, + 0.030193563550710678, + 0.060733940452337265, + 0.013178409077227116, + 0.016041269525885582, + 0.015376454219222069, + 0.011026215739548206, + 0.017930420115590096, + 0.010038010776042938, + 0.00933857075870037, + 0.015014096163213253, + 0.0039895945228636265, + 0.007536901161074638, + 0.007471330463886261, + 0.00360583933070302, + 0.009081760421395302, + 0.015614908188581467, + 0.012946554459631443, + 0.01670999452471733, + 0.007906973361968994, + 0.017744451761245728, + 0.01778346300125122, + 0.010572515428066254, + 0.013165743090212345, + 0.019907059147953987, + 0.016151323914527893, + 0.025214625522494316, + 0.007655836641788483, + 0.011669382452964783, + 0.016450347378849983, + 0.007924046367406845, + 0.009660325944423676, + 0.009554515592753887, + 0.009628423489630222, + 0.011027174070477486, + 0.008189987391233444, + 0.011897865682840347, + 0.016467520967125893, + 0.011444366537034512, + 0.00909917801618576, + 0.005624820478260517, + 0.006127931177616119, + 0.0059101153165102005, + 0.00336252199485898, + 0.004320543725043535, + 0.007762093562632799, + 0.0049384161829948425, + 0.007755335420370102, + 0.00448373518884182, + 0.002796251093968749, + 0.0030065265018492937, + 0.0022446291986852884, + 0.003769573522731662, + 0.0028271330520510674, + 0.0023777515161782503, + 0.00329980137757957, + 0.017391769215464592, + 0.011125639081001282, + 0.023169206455349922, + 0.005205627530813217, + 0.004264083690941334, + 0.006105354055762291, + 0.004842972848564386, + 0.006930011790245771, + 0.005995443556457758, + 0.004303080029785633, + 0.00734702916815877, + 0.0018972865073010325, + 0.004069884307682514, + 0.003229953348636627, + 0.002163570374250412, + 0.0040460205636918545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.17984513938426971 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, close-up, static shot, rendered in classic 16-bit game art, focuses on a powerful hand, trembling slightly with suppressed emotion, its knuckles white with exertion. The hand tightly grips a simple wooden brush, veins visible against the pixelated skin. To one side, a candle's pixelated flame flickers, casting a shaky, halftone-like shadow across the rough wooden table in a cycling of warm yellow and orange pixels, shrouding much of the scene in gloom. A brush tip, saturated with rich black ink, strokes across the rough, yellowed paper with resolute, heavy force, leaving behind a deep, powerful stroke. With an uncontrollable tremor, a heavy drop of ink falls from the tip, rapidly spreading across the paper in an ever-expanding mass of black pixels, like a silent oath. Although the camera mainly focuses on his hands, we can still catch a glimpse of the face of this leopard-headed, round-eyed strong man distorted by grief and anger. His furrowed brows and wide-open angry eyes, under the limited tones and clear pixel blocks, convey a solemn, resolute and tragic atmosphere full of retro gaming feel.", + "seed": 1798, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_3/index_8940_seed_1798_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": -0.07486038934439421, + "vlm_logprob": -0.07486038934439421, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.07486038934439421, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14919237792491913 + ], + "opposite_logprob_tokens": [ + -2.0241923332214355 + ], + "logprob_reward": -0.14919237792491913, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14919237792491913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17556804418563843, + 0.032590556889772415, + 0.08565279096364975, + 0.050630293786525726, + 0.19717511534690857, + 0.03528456762433052, + 0.0582834854722023, + 0.29632407426834106, + 0.029890872538089752, + 0.03860026225447655 + ], + "credit_map": [ + 0.014810873195528984, + 0.016210803762078285, + 0.025232963263988495, + 0.04048412665724754, + 0.014307106845080853, + 0.015561673790216446, + 0.03230918571352959, + 0.016651295125484467, + 0.0034534165170043707, + 0.002831683261319995, + 0.00484838429838419, + 0.005900598596781492, + 0.0032464361283928156, + 0.0031417442951351404, + 0.0043188100680708885, + 0.004849481396377087, + 0.009749622084200382, + 0.008671198040246964, + 0.01116905827075243, + 0.006711920723319054, + 0.011215855367481709, + 0.015269280411303043, + 0.008516456000506878, + 0.014349398203194141, + 0.004943105857819319, + 0.008178208954632282, + 0.006701949518173933, + 0.006102593149989843, + 0.0047006430104374886, + 0.00630529411137104, + 0.006725683342665434, + 0.0069728088565170765, + 0.032043296843767166, + 0.05648394301533699, + 0.03440625220537186, + 0.008501127362251282, + 0.0158381424844265, + 0.019585201516747475, + 0.009195826016366482, + 0.02112128585577011, + 0.004509693942964077, + 0.00326594989746809, + 0.00644186744466424, + 0.0033359008375555277, + 0.005908783990889788, + 0.0025219745002686977, + 0.005994901992380619, + 0.003305492689833045, + 0.009288564324378967, + 0.008569817990064621, + 0.009640865959227085, + 0.007239806465804577, + 0.004602963104844093, + 0.005464636255055666, + 0.007178823463618755, + 0.006298005115240812, + 0.04404627904295921, + 0.10664226859807968, + 0.0858367532491684, + 0.010944304056465626, + 0.016436778008937836, + 0.010652665048837662, + 0.007816573604941368, + 0.01394842378795147, + 0.0037032528780400753, + 0.002715489361435175, + 0.0054665240459144115, + 0.0030943870078772306, + 0.003497065743431449, + 0.003699222346767783, + 0.004021070897579193, + 0.003693858627229929, + 0.005154397804290056, + 0.005503046792000532, + 0.003984347451478243, + 0.002786512952297926, + 0.005017563235014677, + 0.00668359687551856, + 0.003038988681510091, + 0.006431804969906807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.14919237792491913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002414762973785 + ], + "opposite_logprob_tokens": [ + -2.695024251937866 + ], + "logprob_reward": -0.07002414762973785, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002414762973785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1574622094631195, + 0.05837445706129074, + 0.12031954526901245, + 0.05965573340654373, + 0.1785299926996231, + 0.0282236747443676, + 0.09462296962738037, + 0.18174982070922852, + 0.022712068632245064, + 0.09834951907396317 + ], + "credit_map": [ + 0.018119171261787415, + 0.01805248111486435, + 0.014735422097146511, + 0.020797260105609894, + 0.014147026464343071, + 0.018221117556095123, + 0.015059988014400005, + 0.03832974657416344, + 0.007260459009557962, + 0.004587088245898485, + 0.008777167648077011, + 0.006620013155043125, + 0.004536017309874296, + 0.007550479378551245, + 0.0070449733175337315, + 0.011998258531093597, + 0.011815441772341728, + 0.01899687573313713, + 0.018651600927114487, + 0.009852682240307331, + 0.018765049055218697, + 0.012536576949059963, + 0.012160742655396461, + 0.017540574073791504, + 0.005959669128060341, + 0.007118326611816883, + 0.00837209913879633, + 0.007794660981744528, + 0.008450505323708057, + 0.007081417832523584, + 0.007835354655981064, + 0.00704369880259037, + 0.042417701333761215, + 0.021109512075781822, + 0.03328365832567215, + 0.009856211952865124, + 0.01814875937998295, + 0.0212202537804842, + 0.008128348737955093, + 0.02436555176973343, + 0.002859731437638402, + 0.0036210177931934595, + 0.004798962734639645, + 0.002447072183713317, + 0.0031547907274216413, + 0.003267331514507532, + 0.004535661544650793, + 0.0035391084384173155, + 0.012336361221969128, + 0.015550869517028332, + 0.010205216705799103, + 0.01664307713508606, + 0.008993408642709255, + 0.006612483412027359, + 0.013369028456509113, + 0.010912531986832619, + 0.036544572561979294, + 0.06746309995651245, + 0.038804471492767334, + 0.005611801519989967, + 0.010755921714007854, + 0.006883830297738314, + 0.005615961272269487, + 0.010070165619254112, + 0.003937048371881247, + 0.0015916329575702548, + 0.006438600365072489, + 0.0019132021116092801, + 0.001991481985896826, + 0.0018326088320463896, + 0.0025559552013874054, + 0.0024515397381037474, + 0.019184088334441185, + 0.00928126648068428, + 0.025311369448900223, + 0.006346759386360645, + 0.009768866002559662, + 0.012248543091118336, + 0.005344065837562084, + 0.01086456049233675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.07002414762973785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409455131739378 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.00409455131739378, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409455131739378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11246048659086227, + 0.0768032819032669, + 0.09946683794260025, + 0.0996408760547638, + 0.1490364521741867, + 0.03758369758725166, + 0.0780731663107872, + 0.25847187638282776, + 0.05501134321093559, + 0.03345195949077606 + ], + "credit_map": [ + 0.012003275565803051, + 0.015075079165399075, + 0.01796802319586277, + 0.014864728786051273, + 0.010211487300693989, + 0.01416188944131136, + 0.013886362314224243, + 0.01428964827209711, + 0.009595044888556004, + 0.007565178442746401, + 0.009612694382667542, + 0.004996407777070999, + 0.006150138098746538, + 0.014296120032668114, + 0.008185994811356068, + 0.016401702538132668, + 0.00997830368578434, + 0.018024802207946777, + 0.011849869042634964, + 0.008716040290892124, + 0.01811334490776062, + 0.014033488929271698, + 0.00953645445406437, + 0.009214531630277634, + 0.011345300823450089, + 0.009549831971526146, + 0.017102615907788277, + 0.007284790743142366, + 0.01663338765501976, + 0.014743945561349392, + 0.010169153101742268, + 0.012811851687729359, + 0.017943019047379494, + 0.051270343363285065, + 0.044517114758491516, + 0.00621236814185977, + 0.006737957708537579, + 0.006570491474121809, + 0.005795697681605816, + 0.009989453479647636, + 0.004577723331749439, + 0.004015841521322727, + 0.01027908455580473, + 0.0024122060276567936, + 0.0046365647576749325, + 0.003922467585653067, + 0.003231960581615567, + 0.004507849458605051, + 0.011921349912881851, + 0.008641115389764309, + 0.010940058156847954, + 0.0077236006036400795, + 0.01322654727846384, + 0.009386359713971615, + 0.007104016374796629, + 0.009130110964179039, + 0.023330017924308777, + 0.08543816953897476, + 0.09768815338611603, + 0.012516195885837078, + 0.011732272803783417, + 0.007809416390955448, + 0.007576480507850647, + 0.012381166219711304, + 0.011568695306777954, + 0.004449875093996525, + 0.015425003133714199, + 0.00340640707872808, + 0.005298313684761524, + 0.006015096791088581, + 0.0035794752184301615, + 0.005268477369099855, + 0.006582612171769142, + 0.003377305343747139, + 0.007666755933314562, + 0.002019789069890976, + 0.0035398313775658607, + 0.00424480065703392, + 0.001941492548212409, + 0.0040793754160404205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.00409455131739378 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809110447764397 + ], + "opposite_logprob_tokens": [ + -3.2880911827087402 + ], + "logprob_reward": -0.03809110447764397, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809110447764397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23628662526607513, + 0.06752333045005798, + 0.058313921093940735, + 0.1360027939081192, + 0.2116246372461319, + 0.03490183874964714, + 0.09492580592632294, + 0.06750401109457016, + 0.05044461414217949, + 0.04247239604592323 + ], + "credit_map": [ + 0.024745019152760506, + 0.028118886053562164, + 0.023981884121894836, + 0.05988861247897148, + 0.021566351875662804, + 0.016505355015397072, + 0.039014458656311035, + 0.02246607281267643, + 0.008182770572602749, + 0.006198454182595015, + 0.012909187003970146, + 0.009941540658473969, + 0.007471153046935797, + 0.0061223795637488365, + 0.009172454476356506, + 0.007525391411036253, + 0.0063554453663527966, + 0.007556192576885223, + 0.007504256907850504, + 0.0046505252830684185, + 0.008498304523527622, + 0.009122450836002827, + 0.006007219664752483, + 0.008619526401162148, + 0.014019995927810669, + 0.01969011127948761, + 0.023734217509627342, + 0.015595408156514168, + 0.012485587038099766, + 0.017013292759656906, + 0.015668461099267006, + 0.017795734107494354, + 0.02887248806655407, + 0.05309014022350311, + 0.034595392644405365, + 0.008176787756383419, + 0.023680957034230232, + 0.02458282932639122, + 0.009691285900771618, + 0.028934750705957413, + 0.004813954699784517, + 0.004034822806715965, + 0.007412953767925501, + 0.002940716687589884, + 0.004359210841357708, + 0.003279192140325904, + 0.0036463181022554636, + 0.0044146692380309105, + 0.013081992045044899, + 0.00975396204739809, + 0.01441067922860384, + 0.009312165901064873, + 0.017182381823658943, + 0.01133897714316845, + 0.007368362508714199, + 0.012477289885282516, + 0.007502485066652298, + 0.018340086564421654, + 0.019557757303118706, + 0.004297612700611353, + 0.003545177634805441, + 0.0044158948585391045, + 0.00488830404356122, + 0.004956696182489395, + 0.010340883396565914, + 0.00353802228346467, + 0.014716538600623608, + 0.0053812251426279545, + 0.0049466160126030445, + 0.0024446090683341026, + 0.004189596977084875, + 0.004887130111455917, + 0.006352672819048166, + 0.004122154787182808, + 0.010011089965701103, + 0.002308549592271447, + 0.0046078977175056934, + 0.006338467355817556, + 0.0022298998665064573, + 0.0065016676671803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.03809110447764397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 rendered in classic 16-bit game art with pixelated textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a powerful hand with visible white knuckles and trembling motion gripping a wooden brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a candle with a flickering pixelated flame visible on the side, casting halftone-like shadows on a rough wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy drop of black ink fall from the brush tip and spread rapidly across yellowed paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a leopard-headed man with round eyes, furrowed brows, and angry expression visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1128997653722763 + ], + "opposite_logprob_tokens": [ + -2.2378997802734375 + ], + "logprob_reward": -0.1128997653722763, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1128997653722763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19512152671813965, + 0.06101580709218979, + 0.07734104990959167, + 0.28423476219177246, + 0.06499334424734116, + 0.04870140179991722, + 0.07611238956451416, + 0.04110544174909592, + 0.040679432451725006, + 0.11069485545158386 + ], + "credit_map": [ + 0.02037332020699978, + 0.014893715269863605, + 0.019332656636834145, + 0.025246478617191315, + 0.02199108898639679, + 0.02625042200088501, + 0.019107401371002197, + 0.04792643338441849, + 0.00746800797060132, + 0.00416075624525547, + 0.008174771443009377, + 0.00452479999512434, + 0.007331641390919685, + 0.008851725608110428, + 0.006759785581380129, + 0.01374431885778904, + 0.007868586108088493, + 0.01305973157286644, + 0.012501513585448265, + 0.006496200803667307, + 0.01067813578993082, + 0.0065407478250563145, + 0.01094484981149435, + 0.009251278825104237, + 0.0302498247474432, + 0.03601711988449097, + 0.0351383350789547, + 0.018262149766087532, + 0.052307091653347015, + 0.04420297220349312, + 0.027124710381031036, + 0.0409325547516346, + 0.00972549058496952, + 0.008688502945005894, + 0.02209147810935974, + 0.0030356349889189005, + 0.005298458971083164, + 0.00516494270414114, + 0.0033303131349384785, + 0.007658517919480801, + 0.00502465246245265, + 0.004692541901022196, + 0.008832942694425583, + 0.002981978002935648, + 0.008011192083358765, + 0.006821400951594114, + 0.005120920017361641, + 0.007215774618089199, + 0.009283044375479221, + 0.008028467185795307, + 0.008776409551501274, + 0.010076750069856644, + 0.007042245008051395, + 0.010224963538348675, + 0.006724070757627487, + 0.01595643348991871, + 0.004918395075947046, + 0.006973154842853546, + 0.008631071075797081, + 0.0019170410232618451, + 0.0057013751938939095, + 0.004700093064457178, + 0.003354797838255763, + 0.004909511189907789, + 0.006854756735265255, + 0.003115815808996558, + 0.009038440883159637, + 0.0034081877674907446, + 0.005614764057099819, + 0.00419223215430975, + 0.003830400062724948, + 0.004624835681170225, + 0.023423178121447563, + 0.010308787226676941, + 0.029141630977392197, + 0.008112968876957893, + 0.011891365982592106, + 0.012326568365097046, + 0.005879921838641167, + 0.009610435925424099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.1128997653722763 + } + ], + "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/0000128/group_4/index_tensor([6389, 6389])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_4/index_tensor([6389, 6389])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..50243cf8cdf7f02c80e1da70d58aaf7acc4deec8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_4/index_tensor([6389, 6389])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Dizzying first-person perspective footage places the viewer directly inside a perpetually spinning hamster wheel, creating a dizzying and hypnotic sensory experience. Locked in a rapid, continuous 360-degree roll, the camera captures the rhythmic flashes of thin, polished metal bars passing in front of the lens, their cool, reflective surfaces capturing stroboscopic light from the outside world. Through these fleeting gaps, the exterior environment becomes a chaotic yet mesmerizing spectacle. Still objects and defined shapes distort and stretch into elongated, flowing streaks of extreme motion blur, their distinctive forms dissolving into swirls of abstract color. The world outside the cage becomes a seamless, kaleidoscopic tapestry, filling the frame, vibrant hues of red, blue, and yellow blending and smearing with each rotation. The bars' relentless rotation and stroboscopic effects create a disorienting, almost psychedelic atmosphere, immersing the viewer in a looping, visually overwhelming, and beautiful abstract reality.", + "seed": 1884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_4/index_6389_seed_1884_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": -0.03982549230568111, + "vlm_logprob": -0.03982549230568111, + "vlm_yes_no_margin": 4.199999976158142, + "vlm_token_credit": -0.03982549230568111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.13364659249782562 + ], + "opposite_logprob_tokens": [ + -2.1336464881896973 + ], + "logprob_reward": -0.13364659249782562, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13364659249782562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3033328652381897, + 0.12477295100688934, + 0.0907074436545372, + 0.11734306812286377, + 0.04503636062145233, + 0.05958811938762665, + 0.09698303043842316, + 0.04888032376766205, + 0.060094911605119705, + 0.053260836750268936 + ], + "credit_map": [ + 0.020087383687496185, + 0.04016765207052231, + 0.04215247184038162, + 0.0232606939971447, + 0.06088276579976082, + 0.03985050320625305, + 0.044942766427993774, + 0.03198864310979843, + 0.006809172686189413, + 0.020366817712783813, + 0.014032370410859585, + 0.01587575301527977, + 0.014286103658378124, + 0.013212542980909348, + 0.02887938730418682, + 0.011310804635286331, + 0.007040238007903099, + 0.014831000939011574, + 0.007559378631412983, + 0.00820905715227127, + 0.015882330015301704, + 0.014887982048094273, + 0.008736793883144855, + 0.013560663908720016, + 0.026404589414596558, + 0.010251467116177082, + 0.006738041527569294, + 0.03440900146961212, + 0.010304631665349007, + 0.01616712287068367, + 0.006569496355950832, + 0.006498716305941343, + 0.0040587252005934715, + 0.006345842033624649, + 0.0034026424400508404, + 0.006655620411038399, + 0.006648516282439232, + 0.004371441900730133, + 0.005690994206815958, + 0.00786257442086935, + 0.0056416685692965984, + 0.0064743198454380035, + 0.006038778927177191, + 0.00597855681553483, + 0.011273948475718498, + 0.008099483326077461, + 0.00828044954687357, + 0.007800910621881485, + 0.009685840457677841, + 0.014132712967693806, + 0.010192451998591423, + 0.011073492467403412, + 0.013146322220563889, + 0.012613406404852867, + 0.012596380896866322, + 0.013542420230805874, + 0.004483871161937714, + 0.006796122062951326, + 0.005683466792106628, + 0.00606796657666564, + 0.005717962048947811, + 0.005158015992492437, + 0.007584770675748587, + 0.007388146128505468, + 0.0030447079334408045, + 0.0050047277472913265, + 0.0066408393904566765, + 0.021732108667492867, + 0.012063145637512207, + 0.004947942215949297, + 0.004107585176825523, + 0.002553856000304222, + 0.005138591397553682, + 0.005037188995629549, + 0.0060369838029146194, + 0.00788022205233574, + 0.011078045703470707, + 0.007439808454364538, + 0.00582916522398591, + 0.004820835776627064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.13364659249782562 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.002227446297183633 + ], + "opposite_logprob_tokens": [ + -6.127227306365967 + ], + "logprob_reward": -0.002227446297183633, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002227446297183633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09507576376199722, + 0.40141046047210693, + 0.1126483827829361, + 0.09555546939373016, + 0.03863733261823654, + 0.04632788524031639, + 0.07162196189165115, + 0.05739227309823036, + 0.043835267424583435, + 0.03749515861272812 + ], + "credit_map": [ + 0.007071512285619974, + 0.010282386094331741, + 0.017956245690584183, + 0.008772894740104675, + 0.013478439301252365, + 0.012013398110866547, + 0.012227347120642662, + 0.013273552060127258, + 0.03144208714365959, + 0.10118988901376724, + 0.025691891089081764, + 0.032208096235990524, + 0.04926075041294098, + 0.05767720937728882, + 0.06430551409721375, + 0.03963504731655121, + 0.012627868913114071, + 0.015414158813655376, + 0.008029282093048096, + 0.00942362193018198, + 0.016008101403713226, + 0.021137790754437447, + 0.015101074241101742, + 0.014906493946909904, + 0.02685210481286049, + 0.007832543924450874, + 0.006918591447174549, + 0.028238143771886826, + 0.00658057676628232, + 0.005042932461947203, + 0.006514810491353273, + 0.007575768977403641, + 0.0032384698279201984, + 0.007129693403840065, + 0.0032319549936801195, + 0.005576689727604389, + 0.0057511418126523495, + 0.0036442463751882315, + 0.004011054988950491, + 0.006054086610674858, + 0.005155529826879501, + 0.006070380564779043, + 0.004534577019512653, + 0.005591752473264933, + 0.008705735206604004, + 0.005786166526377201, + 0.006203362252563238, + 0.004280386492609978, + 0.005688505247235298, + 0.008446785621345043, + 0.008428801782429218, + 0.014524825848639011, + 0.009284228086471558, + 0.007208277937024832, + 0.011103649623692036, + 0.006936897058039904, + 0.006303907372057438, + 0.006906991358846426, + 0.0065680439583957195, + 0.008861140348017216, + 0.005422691814601421, + 0.0058990162797272205, + 0.011165698058903217, + 0.006264784839004278, + 0.0031044038478285074, + 0.0036677485331892967, + 0.004563020076602697, + 0.010767748579382896, + 0.011927221901714802, + 0.0037312335334718227, + 0.003423680318519473, + 0.0026502127293497324, + 0.003325272351503372, + 0.005853825714439154, + 0.004324665293097496, + 0.00799441710114479, + 0.004029633942991495, + 0.0043478659354150295, + 0.004947145935148001, + 0.002672334434464574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.002227446297183633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.04865863919258118 + ], + "opposite_logprob_tokens": [ + -3.048658609390259 + ], + "logprob_reward": -0.04865863919258118, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865863919258118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16446974873542786, + 0.1563078910112381, + 0.04494527727365494, + 0.11245306581258774, + 0.02843540534377098, + 0.22888442873954773, + 0.04966558888554573, + 0.0954681858420372, + 0.08834326267242432, + 0.03102712333202362 + ], + "credit_map": [ + 0.01081494428217411, + 0.026227539405226707, + 0.020576003938913345, + 0.014714638702571392, + 0.023501507937908173, + 0.028295472264289856, + 0.019155602902173996, + 0.021184027194976807, + 0.012659779749810696, + 0.04285148158669472, + 0.013270122930407524, + 0.012707418762147427, + 0.01959512010216713, + 0.02058638632297516, + 0.018352515995502472, + 0.016285067424178123, + 0.005987401120364666, + 0.005013804417103529, + 0.004606662783771753, + 0.006240829825401306, + 0.007294026669114828, + 0.006707950960844755, + 0.005396549124270678, + 0.003698052139952779, + 0.013066170737147331, + 0.015286031179130077, + 0.011602574028074741, + 0.017283892259001732, + 0.014163062907755375, + 0.02274823747575283, + 0.007023803889751434, + 0.01127929799258709, + 0.0024384851567447186, + 0.005102616734802723, + 0.0023785645607858896, + 0.005280400160700083, + 0.0038036333862692118, + 0.0024345621932297945, + 0.003507375018671155, + 0.0034897667355835438, + 0.01665269397199154, + 0.02954678051173687, + 0.021557500585913658, + 0.025235414505004883, + 0.05087607726454735, + 0.03290894627571106, + 0.02707209251821041, + 0.02503492496907711, + 0.00672242371365428, + 0.005675220862030983, + 0.008111946284770966, + 0.007865700870752335, + 0.006838591769337654, + 0.0045051416382193565, + 0.005942191928625107, + 0.00400437181815505, + 0.009209875017404556, + 0.013557560741901398, + 0.011517422273755074, + 0.014756008051335812, + 0.012793483212590218, + 0.012825622223317623, + 0.0093013821169734, + 0.01150683593004942, + 0.005709153134375811, + 0.00467925937846303, + 0.011703955940902233, + 0.021214589476585388, + 0.030349304899573326, + 0.005461379885673523, + 0.006253902800381184, + 0.0029717173893004656, + 0.0028901712503284216, + 0.003950847778469324, + 0.004129356704652309, + 0.0036774466279894114, + 0.0060673546977341175, + 0.003579052397981286, + 0.0037123828660696745, + 0.0030205138027668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.04865863919258118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.008642295375466347 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642295375466347, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642295375466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12004909664392471, + 0.29585516452789307, + 0.08195734769105911, + 0.05436760559678078, + 0.05627567321062088, + 0.09146226942539215, + 0.10750762373209, + 0.042265232652425766, + 0.06834428012371063, + 0.08191575109958649 + ], + "credit_map": [ + 0.010123003274202347, + 0.016727961599826813, + 0.02006441541016102, + 0.014174189418554306, + 0.01526843011379242, + 0.014800004661083221, + 0.0157881248742342, + 0.013102957978844643, + 0.01980346255004406, + 0.05926581472158432, + 0.032569561153650284, + 0.03916635364294052, + 0.032568007707595825, + 0.03805999830365181, + 0.04691845923662186, + 0.02750350721180439, + 0.007944954559206963, + 0.013745175674557686, + 0.005944692064076662, + 0.008417910896241665, + 0.009501809254288673, + 0.022431068122386932, + 0.006977341137826443, + 0.006994393188506365, + 0.008643425069749355, + 0.006065875291824341, + 0.007381567731499672, + 0.007352769840508699, + 0.007699958048760891, + 0.008000961504876614, + 0.005320035852491856, + 0.0039030099287629128, + 0.00431786710396409, + 0.006600476801395416, + 0.004687363747507334, + 0.00694690877571702, + 0.010173204354941845, + 0.005754147190600634, + 0.010593331418931484, + 0.007202375680208206, + 0.008970576338469982, + 0.0077199204824864864, + 0.010850092396140099, + 0.011418422684073448, + 0.016208617016673088, + 0.009808288887143135, + 0.015195818617939949, + 0.0112905353307724, + 0.00823032297194004, + 0.012937307357788086, + 0.012769512832164764, + 0.02265847846865654, + 0.013014261610805988, + 0.010952894575893879, + 0.016986070200800896, + 0.009958776645362377, + 0.004703995771706104, + 0.00527846859768033, + 0.00492702005431056, + 0.0048536332324147224, + 0.004107986111193895, + 0.0051944926381111145, + 0.005994702223688364, + 0.007204936817288399, + 0.003917297348380089, + 0.002769164042547345, + 0.008105326443910599, + 0.018062056973576546, + 0.022688789293169975, + 0.005171481519937515, + 0.005317529197782278, + 0.002312630647793412, + 0.00572848878800869, + 0.008060592226684093, + 0.008786476217210293, + 0.013644947670400143, + 0.013065257109701633, + 0.010561946779489517, + 0.013073348440229893, + 0.0089946947991848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.008642295375466347 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.005952488165348768 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952488165348768, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952488165348768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08284217864274979, + 0.08168338239192963, + 0.08767674118280411, + 0.307584285736084, + 0.0753399133682251, + 0.04113373905420303, + 0.18545441329479218, + 0.059888679534196854, + 0.014086127281188965, + 0.06431051343679428 + ], + "credit_map": [ + 0.0070564029738307, + 0.014606761746108532, + 0.016765695065259933, + 0.007478827610611916, + 0.009966474026441574, + 0.00844850018620491, + 0.00969758816063404, + 0.008821931667625904, + 0.005486128386110067, + 0.014407780021429062, + 0.008379897102713585, + 0.01244199089705944, + 0.008646226488053799, + 0.01147821731865406, + 0.013641230762004852, + 0.007201910950243473, + 0.008701327256858349, + 0.01906086690723896, + 0.004217850975692272, + 0.0069752647541463375, + 0.01013015117496252, + 0.025060953572392464, + 0.006498823408037424, + 0.00703149801120162, + 0.05221528559923172, + 0.02626989595592022, + 0.020524751394987106, + 0.08621062338352203, + 0.035761214792728424, + 0.05419443920254707, + 0.015302272513508797, + 0.017105791717767715, + 0.004954858683049679, + 0.011358910240232944, + 0.005179549567401409, + 0.008875751867890358, + 0.012240663170814514, + 0.005408815573900938, + 0.015905117616057396, + 0.011416254565119743, + 0.002712707268074155, + 0.00535623449832201, + 0.0034106357488781214, + 0.00462071318179369, + 0.01023021712899208, + 0.007009317632764578, + 0.004334701225161552, + 0.0034592142328619957, + 0.02521098218858242, + 0.023337675258517265, + 0.018348874524235725, + 0.0219953041523695, + 0.020889846608042717, + 0.03168472275137901, + 0.02391073666512966, + 0.020076284185051918, + 0.0061426665633916855, + 0.00732545368373394, + 0.006074040196835995, + 0.011749902740120888, + 0.008072598837316036, + 0.004907544236630201, + 0.009231068193912506, + 0.006385405547916889, + 0.001261899364180863, + 0.0018838540418073535, + 0.0019521176582202315, + 0.002712559886276722, + 0.0018370840698480606, + 0.00134182907640934, + 0.0019772849045693874, + 0.0011194981634616852, + 0.0052562979981303215, + 0.006746964529156685, + 0.008194251917302608, + 0.007761199492961168, + 0.01230965182185173, + 0.009148238226771355, + 0.008859870955348015, + 0.006034037098288536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.005952488165348768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Dizzying first-person perspective footage places the viewer directly inside a perpetually spinning hamster wheel, creating a dizzying and hypnotic sensory experience. Locked in a rapid, continuous 360-degree roll, the camera captures the rhythmic flashes of thin, polished metal bars passing in front of the lens, their cool, reflective surfaces capturing stroboscopic light from the outside world. Through these fleeting gaps, the exterior environment becomes a chaotic yet mesmerizing spectacle. Still objects and defined shapes distort and stretch into elongated, flowing streaks of extreme motion blur, their distinctive forms dissolving into swirls of abstract color. The world outside the cage becomes a seamless, kaleidoscopic tapestry, filling the frame, vibrant hues of red, blue, and yellow blending and smearing with each rotation. The bars' relentless rotation and stroboscopic effects create a disorienting, almost psychedelic atmosphere, immersing the viewer in a looping, visually overwhelming, and beautiful abstract reality.", + "seed": 1885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_4/index_6389_seed_1885_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": -0.022293895343318583, + "vlm_logprob": -0.022293895343318583, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.022293895343318583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.0956585556268692 + ], + "opposite_logprob_tokens": [ + -2.470658540725708 + ], + "logprob_reward": -0.0956585556268692, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0956585556268692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16316549479961395, + 0.2466503381729126, + 0.021547256037592888, + 0.016196468845009804, + 0.03790965676307678, + 0.07805236428976059, + 0.06396390497684479, + 0.026107192039489746, + 0.27769777178764343, + 0.06870946288108826 + ], + "credit_map": [ + 0.01980496756732464, + 0.011631589382886887, + 0.05899186059832573, + 0.014515395276248455, + 0.01147244032472372, + 0.010792129673063755, + 0.010363941080868244, + 0.02559315785765648, + 0.07954921573400497, + 0.024516304954886436, + 0.017738284543156624, + 0.010550511069595814, + 0.01513509638607502, + 0.013595745898783207, + 0.017828714102506638, + 0.06773647665977478, + 0.0023719831369817257, + 0.002526855794712901, + 0.0031853013206273317, + 0.001919110189191997, + 0.0029269508086144924, + 0.003166784532368183, + 0.0027657721657305956, + 0.0026844977401196957, + 0.0025996239855885506, + 0.0026831133291125298, + 0.0010516350157558918, + 0.0012799263931810856, + 0.002454453380778432, + 0.002653669100254774, + 0.0015554045094177127, + 0.00191864266525954, + 0.0035998879466205835, + 0.0034533219877630472, + 0.0038844530936330557, + 0.005256458651274443, + 0.007391492370516062, + 0.003899948438629508, + 0.006114061456173658, + 0.004310034215450287, + 0.007899665273725986, + 0.004004513379186392, + 0.025119638070464134, + 0.005818938370794058, + 0.004917627200484276, + 0.020918192341923714, + 0.004236853215843439, + 0.005136941093951464, + 0.005462423898279667, + 0.013685515150427818, + 0.010288617573678493, + 0.005209523718804121, + 0.009281359612941742, + 0.007071679458022118, + 0.007267094682902098, + 0.00569769274443388, + 0.004032769240438938, + 0.0020143776200711727, + 0.0029801365453749895, + 0.002777560381218791, + 0.005514075048267841, + 0.0036497688852250576, + 0.002124516060575843, + 0.003013986861333251, + 0.02332010306417942, + 0.015940867364406586, + 0.04281975328922272, + 0.011982669122517109, + 0.10333048552274704, + 0.013779105618596077, + 0.016503192484378815, + 0.05002162232995033, + 0.004713910631835461, + 0.006954784970730543, + 0.0032074893824756145, + 0.0027955060359090567, + 0.0265275277197361, + 0.014829680323600769, + 0.0061182123608887196, + 0.0035623519215732813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0956585556268692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.002840890549123287 + ], + "opposite_logprob_tokens": [ + -5.877840995788574 + ], + "logprob_reward": -0.002840890549123287, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002840890549123287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26460716128349304, + 0.13051241636276245, + 0.09985005855560303, + 0.0629391074180603, + 0.05724172666668892, + 0.14534258842468262, + 0.06685781478881836, + 0.04355767369270325, + 0.07451128959655762, + 0.05458014830946922 + ], + "credit_map": [ + 0.027691621333360672, + 0.01926538534462452, + 0.0704403966665268, + 0.02787129394710064, + 0.025837894529104233, + 0.03355972468852997, + 0.023954205214977264, + 0.03598663583397865, + 0.019697951152920723, + 0.01702229119837284, + 0.019335510209202766, + 0.0188699159771204, + 0.011629227548837662, + 0.013362204656004906, + 0.016258805990219116, + 0.01433650404214859, + 0.010940203443169594, + 0.007347906474024057, + 0.018260590732097626, + 0.008042515255510807, + 0.010316422209143639, + 0.017811909317970276, + 0.012661151587963104, + 0.014469360932707787, + 0.007764542009681463, + 0.004632644355297089, + 0.00788143090903759, + 0.007273537106812, + 0.011686036363244057, + 0.009076789952814579, + 0.006437123753130436, + 0.00818700809031725, + 0.005718734115362167, + 0.005602685268968344, + 0.007359987124800682, + 0.006482721772044897, + 0.010514930821955204, + 0.006830340716987848, + 0.008108319714665413, + 0.006624005734920502, + 0.018616583198308945, + 0.013879615813493729, + 0.05008609592914581, + 0.011010470800101757, + 0.008519426919519901, + 0.02135775424540043, + 0.0138706024736166, + 0.008002030663192272, + 0.0063355001620948315, + 0.01128372922539711, + 0.009920994751155376, + 0.004443260841071606, + 0.008844885975122452, + 0.013895335607230663, + 0.006510675884783268, + 0.005623430013656616, + 0.007659881841391325, + 0.0037908400408923626, + 0.005088833626359701, + 0.005495950113981962, + 0.006292181555181742, + 0.005928113125264645, + 0.004190297331660986, + 0.005111577454954386, + 0.008161005564033985, + 0.011683794669806957, + 0.007926132529973984, + 0.008326278999447823, + 0.011085192672908306, + 0.007265234366059303, + 0.007997138425707817, + 0.012066510505974293, + 0.006976400967687368, + 0.008645404130220413, + 0.005515595432370901, + 0.005415854509919882, + 0.013257262296974659, + 0.005753319710493088, + 0.005950744729489088, + 0.0030655639711767435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.002840890549123287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.005255692172795534 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255692172795534, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255692172795534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0834907814860344, + 0.11849842965602875, + 0.06225774064660072, + 0.02840593084692955, + 0.033327292650938034, + 0.23071561753749847, + 0.05292602255940437, + 0.04782954975962639, + 0.20172782242298126, + 0.14082077145576477 + ], + "credit_map": [ + 0.007532850839197636, + 0.013550257310271263, + 0.012173164635896683, + 0.014735754579305649, + 0.011743965558707714, + 0.008941682055592537, + 0.0069205970503389835, + 0.007892503403127193, + 0.034103430807590485, + 0.011768783442676067, + 0.00646604411303997, + 0.011472828686237335, + 0.006535932421684265, + 0.007142259739339352, + 0.012194273993372917, + 0.02881486713886261, + 0.006395174656063318, + 0.0055528501980006695, + 0.005194423720240593, + 0.005106144119054079, + 0.018185541033744812, + 0.00684385746717453, + 0.007753367535769939, + 0.007226384244859219, + 0.00405037822201848, + 0.003695943858474493, + 0.002478779060766101, + 0.0028839597944170237, + 0.00424343952909112, + 0.0052879247814416885, + 0.002772887237370014, + 0.002992618829011917, + 0.003978407941758633, + 0.00400252640247345, + 0.0033971259836107492, + 0.004951979964971542, + 0.004054813180118799, + 0.0037179244682192802, + 0.005373901687562466, + 0.0038506134878844023, + 0.028953662142157555, + 0.02326486073434353, + 0.0803188607096672, + 0.01738136261701584, + 0.01250607892870903, + 0.035561155527830124, + 0.020806068554520607, + 0.0119235934689641, + 0.006099354475736618, + 0.009213992394506931, + 0.0053567346185445786, + 0.004924925044178963, + 0.009550313465297222, + 0.00905906967818737, + 0.00447858264669776, + 0.00424305209890008, + 0.006307428702712059, + 0.00429924251511693, + 0.004435297567397356, + 0.00465558422729373, + 0.010322589427232742, + 0.005807406269013882, + 0.006939058657735586, + 0.005062936805188656, + 0.015376443043351173, + 0.011494426056742668, + 0.027361251413822174, + 0.007383210584521294, + 0.07771017402410507, + 0.01008452009409666, + 0.012002727016806602, + 0.04031507298350334, + 0.009077288210391998, + 0.012651961296796799, + 0.004535341635346413, + 0.005629234481602907, + 0.05060034245252609, + 0.039532821625471115, + 0.009904061444103718, + 0.008889717049896717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.005255692172795534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.004095026291906834 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095026291906834, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095026291906834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15522357821464539, + 0.11289265751838684, + 0.11774519085884094, + 0.020715801045298576, + 0.09920001775026321, + 0.05692669376730919, + 0.15349802374839783, + 0.03260480985045433, + 0.18005050718784332, + 0.0711427628993988 + ], + "credit_map": [ + 0.013464546762406826, + 0.01630173623561859, + 0.04449484869837761, + 0.016666769981384277, + 0.014970715157687664, + 0.012098654173314571, + 0.013380428776144981, + 0.023845873773097992, + 0.03165031224489212, + 0.010467370972037315, + 0.008987173438072205, + 0.00846512895077467, + 0.007089740596711636, + 0.008998882956802845, + 0.010000952519476414, + 0.02723308838903904, + 0.008432542905211449, + 0.00980877224355936, + 0.01670410856604576, + 0.01013597659766674, + 0.019842788577079773, + 0.020605310797691345, + 0.01496109552681446, + 0.01725458912551403, + 0.002629683818668127, + 0.003296811133623123, + 0.0012287823483347893, + 0.0014479509554803371, + 0.0035470149014145136, + 0.004252156242728233, + 0.002584070898592472, + 0.0017293293494731188, + 0.012289442121982574, + 0.009845214895904064, + 0.014025943353772163, + 0.011662951670587063, + 0.016042765229940414, + 0.011514874175190926, + 0.011352285742759705, + 0.01246653776615858, + 0.012307384982705116, + 0.005771400406956673, + 0.005603130906820297, + 0.0034748504403978586, + 0.0049914028495550156, + 0.014471123926341534, + 0.004542828071862459, + 0.005764570087194443, + 0.013546951115131378, + 0.042255185544490814, + 0.026859605684876442, + 0.011021801270544529, + 0.017944617196917534, + 0.017678774893283844, + 0.010574914515018463, + 0.013616170734167099, + 0.0047195348888635635, + 0.0028862820472568274, + 0.0033613238483667374, + 0.00294488831423223, + 0.004733341280370951, + 0.00398465059697628, + 0.005644000601023436, + 0.004330787807703018, + 0.014699310064315796, + 0.013321791775524616, + 0.014180752448737621, + 0.008511045016348362, + 0.06507544219493866, + 0.011378446593880653, + 0.01139166858047247, + 0.0414920337498188, + 0.004355917684733868, + 0.006186371669173241, + 0.003592101391404867, + 0.003747150069102645, + 0.02432999014854431, + 0.019489696249365807, + 0.004838983062654734, + 0.004602550528943539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.004095026291906834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.0036193120758980513 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036193120758980513, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036193120758980513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25961360335350037, + 0.19043412804603577, + 0.08670360594987869, + 0.03938819468021393, + 0.08145362138748169, + 0.09509535133838654, + 0.03456832841038704, + 0.03384820371866226, + 0.11559956520795822, + 0.0632954016327858 + ], + "credit_map": [ + 0.027739936485886574, + 0.01932063139975071, + 0.08182917535305023, + 0.023461775854229927, + 0.01741788350045681, + 0.018189501017332077, + 0.018799399957060814, + 0.05285530909895897, + 0.06113114207983017, + 0.017500195652246475, + 0.01340574212372303, + 0.010833621956408024, + 0.010361338965594769, + 0.011972523294389248, + 0.014865819364786148, + 0.05036373808979988, + 0.007176010869443417, + 0.00640789745375514, + 0.009855235926806927, + 0.006832826882600784, + 0.011416763067245483, + 0.015250463970005512, + 0.016702570021152496, + 0.013061843812465668, + 0.0034757480025291443, + 0.006026019807904959, + 0.0032790943514555693, + 0.0057071419432759285, + 0.006904966197907925, + 0.006825616583228111, + 0.0036980127915740013, + 0.003471596632152796, + 0.009423146955668926, + 0.011274296790361404, + 0.0112624391913414, + 0.010755791328847408, + 0.011049945838749409, + 0.006698921788483858, + 0.009676317684352398, + 0.011312766931951046, + 0.013819283805787563, + 0.008337117731571198, + 0.034603677690029144, + 0.00696165207773447, + 0.004755698144435883, + 0.013044399209320545, + 0.009218910709023476, + 0.004354612901806831, + 0.003139100968837738, + 0.007894097827374935, + 0.00794188678264618, + 0.002452016342431307, + 0.00415217038244009, + 0.0027045863680541515, + 0.0032889775466173887, + 0.0029954954516142607, + 0.005366598255932331, + 0.0037145311944186687, + 0.003731055883690715, + 0.0035942434333264828, + 0.007151193916797638, + 0.003925795201212168, + 0.0031169585417956114, + 0.0032478277571499348, + 0.010275779291987419, + 0.00754590705037117, + 0.01291430089622736, + 0.005208844784647226, + 0.04194408655166626, + 0.00710646016523242, + 0.006353762932121754, + 0.02425043284893036, + 0.0024557020515203476, + 0.005138241685926914, + 0.003132397774606943, + 0.0029698885045945644, + 0.02411937527358532, + 0.01685987412929535, + 0.004786909092217684, + 0.0038330156821757555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0036193120758980513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Dizzying first-person perspective footage places the viewer directly inside a perpetually spinning hamster wheel, creating a dizzying and hypnotic sensory experience. Locked in a rapid, continuous 360-degree roll, the camera captures the rhythmic flashes of thin, polished metal bars passing in front of the lens, their cool, reflective surfaces capturing stroboscopic light from the outside world. Through these fleeting gaps, the exterior environment becomes a chaotic yet mesmerizing spectacle. Still objects and defined shapes distort and stretch into elongated, flowing streaks of extreme motion blur, their distinctive forms dissolving into swirls of abstract color. The world outside the cage becomes a seamless, kaleidoscopic tapestry, filling the frame, vibrant hues of red, blue, and yellow blending and smearing with each rotation. The bars' relentless rotation and stroboscopic effects create a disorienting, almost psychedelic atmosphere, immersing the viewer in a looping, visually overwhelming, and beautiful abstract reality.", + "seed": 1886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_4/index_6389_seed_1886_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": -0.02219948945567012, + "vlm_logprob": -0.02219948945567012, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02219948945567012, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.061623238027095795 + ], + "opposite_logprob_tokens": [ + -2.9366233348846436 + ], + "logprob_reward": -0.061623238027095795, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.061623238027095795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10292480885982513, + 0.22807621955871582, + 0.04988865926861763, + 0.08149222284555435, + 0.10079639405012131, + 0.1450461894273758, + 0.17788921296596527, + 0.03677991405129433, + 0.035963039845228195, + 0.04114332050085068 + ], + "credit_map": [ + 0.01632041111588478, + 0.018042253330349922, + 0.008573970757424831, + 0.017284760251641273, + 0.008659495040774345, + 0.01106940396130085, + 0.011419229209423065, + 0.011555285193026066, + 0.017856145277619362, + 0.006971259601414204, + 0.01640293002128601, + 0.01195928268134594, + 0.016010137274861336, + 0.08618170022964478, + 0.01359573658555746, + 0.05909903720021248, + 0.008845854550600052, + 0.006116338539868593, + 0.009788474068045616, + 0.005939236376434565, + 0.004170716740190983, + 0.005660582333803177, + 0.0038320375606417656, + 0.005535420961678028, + 0.008531944826245308, + 0.018517455086112022, + 0.008235977962613106, + 0.006608526222407818, + 0.010534067638218403, + 0.008903016336262226, + 0.010506539605557919, + 0.009654697962105274, + 0.016635814681649208, + 0.012646137736737728, + 0.007716754917055368, + 0.008621683344244957, + 0.015337706543505192, + 0.014457211829721928, + 0.012413124553859234, + 0.012967955321073532, + 0.030140580609440804, + 0.018625430762767792, + 0.01717369444668293, + 0.01633484847843647, + 0.01607603020966053, + 0.02481936104595661, + 0.010672694072127342, + 0.01120354887098074, + 0.018944527953863144, + 0.029585661366581917, + 0.01725323311984539, + 0.02615043893456459, + 0.0264638289809227, + 0.018206115812063217, + 0.015520316548645496, + 0.025765080004930496, + 0.0029935920611023903, + 0.006246602162718773, + 0.0026420406065881252, + 0.0035986779257655144, + 0.004867361392825842, + 0.005067312624305487, + 0.003411856945604086, + 0.00795246846973896, + 0.005030024331063032, + 0.006205820012837648, + 0.0030160285532474518, + 0.0038219664711505175, + 0.005043615587055683, + 0.006088425405323505, + 0.0030702874064445496, + 0.0036868685856461525, + 0.0049873460084199905, + 0.004807830322533846, + 0.005880387499928474, + 0.0052653709426522255, + 0.0035895961336791515, + 0.005853899754583836, + 0.005090057849884033, + 0.005668828263878822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.061623238027095795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.007649060804396868 + ], + "opposite_logprob_tokens": [ + -4.882648944854736 + ], + "logprob_reward": -0.007649060804396868, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649060804396868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08894751220941544, + 0.23582656681537628, + 0.02775774523615837, + 0.05350103601813316, + 0.10114266723394394, + 0.17268550395965576, + 0.10465516149997711, + 0.11801527440547943, + 0.05766984447836876, + 0.03979865834116936 + ], + "credit_map": [ + 0.015325285494327545, + 0.013500525616109371, + 0.010793778114020824, + 0.009054784663021564, + 0.007001119200140238, + 0.01174442283809185, + 0.009470580145716667, + 0.012057016603648663, + 0.007365243975073099, + 0.012098119594156742, + 0.013882448896765709, + 0.016736244782805443, + 0.018831266090273857, + 0.06852266192436218, + 0.024544931948184967, + 0.07384566217660904, + 0.004022687207907438, + 0.004150559194386005, + 0.003702213754877448, + 0.003215367440134287, + 0.003999315667897463, + 0.0035094276536256075, + 0.0023903187830001116, + 0.002767854603007436, + 0.005705842282623053, + 0.011714184656739235, + 0.005493205972015858, + 0.004327095113694668, + 0.007089966908097267, + 0.006593857891857624, + 0.0065979305654764175, + 0.005978952627629042, + 0.012453710660338402, + 0.008756700903177261, + 0.014034321531653404, + 0.012562469579279423, + 0.015197700820863247, + 0.013413017615675926, + 0.012854378670454025, + 0.011870366521179676, + 0.04280266538262367, + 0.010628169402480125, + 0.011940657161176205, + 0.014566722325980663, + 0.009228996001183987, + 0.06093549728393555, + 0.006401189602911472, + 0.016181595623493195, + 0.01091004628688097, + 0.017781469970941544, + 0.00845395028591156, + 0.013619057834148407, + 0.015420475043356419, + 0.011997842229902744, + 0.008659442886710167, + 0.017812877893447876, + 0.008347999304533005, + 0.031700119376182556, + 0.00360693852417171, + 0.008849588222801685, + 0.007516734767705202, + 0.013537893071770668, + 0.010383260436356068, + 0.034072741866111755, + 0.006958390586078167, + 0.009135279804468155, + 0.0035695608239620924, + 0.005544070154428482, + 0.011115701869130135, + 0.008368127979338169, + 0.005577953066676855, + 0.0074007585644721985, + 0.004569710697978735, + 0.003875915426760912, + 0.0051157064735889435, + 0.0066034020856022835, + 0.0039323074743151665, + 0.006264101713895798, + 0.005316822789609432, + 0.004120693076401949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.007649060804396868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.009785539470613003 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785539470613003, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785539470613003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0922243595123291, + 0.16257739067077637, + 0.09566276520490646, + 0.04326940327882767, + 0.0847407653927803, + 0.13443998992443085, + 0.11180181056261063, + 0.07928537577390671, + 0.14906704425811768, + 0.04693106561899185 + ], + "credit_map": [ + 0.01783161424100399, + 0.015274809673428535, + 0.013035674579441547, + 0.008234860375523567, + 0.006079749669879675, + 0.012241964228451252, + 0.00744225550442934, + 0.012083431705832481, + 0.021578777581453323, + 0.0069075520150363445, + 0.012289046309888363, + 0.007829551585018635, + 0.016557877883315086, + 0.06013986095786095, + 0.012547587975859642, + 0.02472713030874729, + 0.011227416805922985, + 0.012390184216201305, + 0.016979403793811798, + 0.00960095040500164, + 0.010234231129288673, + 0.014644810929894447, + 0.008081641048192978, + 0.01250412780791521, + 0.004997635260224342, + 0.00960566196590662, + 0.004963007755577564, + 0.003435213351622224, + 0.004567017778754234, + 0.006856377702206373, + 0.003930691629648209, + 0.004913798533380032, + 0.009334815666079521, + 0.012992093339562416, + 0.006810093764215708, + 0.008176366798579693, + 0.01646455004811287, + 0.011213700287044048, + 0.010178976692259312, + 0.009570170193910599, + 0.02417057938873768, + 0.008472292684018612, + 0.019222816452383995, + 0.010156150907278061, + 0.008790862746536732, + 0.04567912220954895, + 0.00591972516849637, + 0.012028436176478863, + 0.010493326932191849, + 0.015380579978227615, + 0.014446069486439228, + 0.01700788550078869, + 0.018393924459815025, + 0.012136436067521572, + 0.009858491830527782, + 0.014085090719163418, + 0.009442439302802086, + 0.012647811323404312, + 0.00482791056856513, + 0.0053422218188643456, + 0.006772862747311592, + 0.009422712959349155, + 0.007282540202140808, + 0.023546874523162842, + 0.017984788864850998, + 0.03189786896109581, + 0.007517535239458084, + 0.008386155590415001, + 0.01490387599915266, + 0.03944593667984009, + 0.011685596778988838, + 0.017245274037122726, + 0.006088205613195896, + 0.005281594581902027, + 0.005438975989818573, + 0.008179091848433018, + 0.004834709223359823, + 0.0056046792306005955, + 0.005370922386646271, + 0.006132885813713074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 5 + ], + "selected_reward": -0.009785539470613003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.023294596001505852 + ], + "opposite_logprob_tokens": [ + -3.773294687271118 + ], + "logprob_reward": -0.023294596001505852, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294596001505852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07923603802919388, + 0.12784285843372345, + 0.05541983246803284, + 0.06143154576420784, + 0.06838230788707733, + 0.2179410755634308, + 0.048606209456920624, + 0.09425588697195053, + 0.17845860123634338, + 0.06842563301324844 + ], + "credit_map": [ + 0.014545805752277374, + 0.012629535049200058, + 0.010921933688223362, + 0.007593363057821989, + 0.005335192661732435, + 0.010422901250422001, + 0.006009363569319248, + 0.011777942068874836, + 0.004499104339629412, + 0.004618316423147917, + 0.006294977385550737, + 0.004130294546484947, + 0.009506524540483952, + 0.050203122198581696, + 0.007012185174971819, + 0.04157833755016327, + 0.011778438463807106, + 0.007284062914550304, + 0.009501887485384941, + 0.0062671806663274765, + 0.005025289487093687, + 0.006608225870877504, + 0.004628397990018129, + 0.004326347727328539, + 0.006269979290664196, + 0.009736225008964539, + 0.00723186694085598, + 0.004520113579928875, + 0.0058478438295423985, + 0.013215713202953339, + 0.007525504566729069, + 0.007084297481924295, + 0.00782567448914051, + 0.006511115934699774, + 0.005384413991123438, + 0.0065665896981954575, + 0.013374135829508305, + 0.011235064826905727, + 0.008673038333654404, + 0.008812274783849716, + 0.054614558815956116, + 0.013602719642221928, + 0.0338178351521492, + 0.0631447583436966, + 0.012945136986672878, + 0.02026519738137722, + 0.009725503623485565, + 0.009825369343161583, + 0.0042360760271549225, + 0.0053998976945877075, + 0.004235798958688974, + 0.007921973243355751, + 0.009553823620080948, + 0.004786896985024214, + 0.004937706980854273, + 0.007534034550189972, + 0.007949956692755222, + 0.019243353977799416, + 0.004150694236159325, + 0.007391270715743303, + 0.008274544961750507, + 0.01311054639518261, + 0.008639165200293064, + 0.025496356189250946, + 0.019095951691269875, + 0.05348972976207733, + 0.005091545637696981, + 0.005754017736762762, + 0.0093983244150877, + 0.056129612028598785, + 0.010927965864539146, + 0.01857144758105278, + 0.007349127437919378, + 0.006213020998984575, + 0.009574965573847294, + 0.008338399231433868, + 0.006216760259121656, + 0.012237357906997204, + 0.007011218462139368, + 0.011484777554869652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.023294596001505852 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.008645012974739075 + ], + "opposite_logprob_tokens": [ + -4.758645057678223 + ], + "logprob_reward": -0.008645012974739075, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645012974739075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12330559641122818, + 0.08093589544296265, + 0.0474223829805851, + 0.10942074656486511, + 0.0598670169711113, + 0.11554111540317535, + 0.0649879053235054, + 0.10856331139802933, + 0.18892055749893188, + 0.10103549063205719 + ], + "credit_map": [ + 0.01876183971762657, + 0.016481585800647736, + 0.011803229339420795, + 0.015392176806926727, + 0.009629740379750729, + 0.0194956474006176, + 0.014106366783380508, + 0.01763499528169632, + 0.008248535916209221, + 0.003786802990362048, + 0.005911295302212238, + 0.004058247897773981, + 0.007234168238937855, + 0.031073739752173424, + 0.005577872972935438, + 0.015045227482914925, + 0.006149432621896267, + 0.005603765603154898, + 0.009692054241895676, + 0.004634479992091656, + 0.00480037834495306, + 0.0072416625916957855, + 0.004019866697490215, + 0.005280738230794668, + 0.010204744525253773, + 0.01429364737123251, + 0.012274393811821938, + 0.009411598555743694, + 0.01946650817990303, + 0.011164749041199684, + 0.01917574368417263, + 0.013429353013634682, + 0.00645034434273839, + 0.010021788999438286, + 0.0073324223048985004, + 0.006359840277582407, + 0.008798010647296906, + 0.007289371453225613, + 0.007546361070126295, + 0.006068875547498465, + 0.016519643366336823, + 0.006610917393118143, + 0.010524268262088299, + 0.019473696127533913, + 0.006281865295022726, + 0.04148547723889351, + 0.0041704424656927586, + 0.010474796406924725, + 0.006363058928400278, + 0.00811956636607647, + 0.006187788210809231, + 0.006633838173002005, + 0.008009226992726326, + 0.006916707381606102, + 0.00906949583441019, + 0.013688217848539352, + 0.015858806669712067, + 0.011075672693550587, + 0.008495324291288853, + 0.007277454249560833, + 0.010271085426211357, + 0.01969061978161335, + 0.01131852250546217, + 0.024575822055339813, + 0.020401567220687866, + 0.056141287088394165, + 0.005471719894558191, + 0.006124621257185936, + 0.00778612494468689, + 0.061706215143203735, + 0.011810087598860264, + 0.019478917121887207, + 0.009592747315764427, + 0.011243872344493866, + 0.012451019138097763, + 0.01563938893377781, + 0.007182967849075794, + 0.01669287495315075, + 0.01068489532917738, + 0.017547713592648506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.008645012974739075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Dizzying first-person perspective footage places the viewer directly inside a perpetually spinning hamster wheel, creating a dizzying and hypnotic sensory experience. Locked in a rapid, continuous 360-degree roll, the camera captures the rhythmic flashes of thin, polished metal bars passing in front of the lens, their cool, reflective surfaces capturing stroboscopic light from the outside world. Through these fleeting gaps, the exterior environment becomes a chaotic yet mesmerizing spectacle. Still objects and defined shapes distort and stretch into elongated, flowing streaks of extreme motion blur, their distinctive forms dissolving into swirls of abstract color. The world outside the cage becomes a seamless, kaleidoscopic tapestry, filling the frame, vibrant hues of red, blue, and yellow blending and smearing with each rotation. The bars' relentless rotation and stroboscopic effects create a disorienting, almost psychedelic atmosphere, immersing the viewer in a looping, visually overwhelming, and beautiful abstract reality.", + "seed": 1887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000128/group_4/index_6389_seed_1887_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": -0.014623863762244583, + "vlm_logprob": -0.014623863762244583, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.014623863762244583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.050654854625463486 + ], + "opposite_logprob_tokens": [ + -3.175654888153076 + ], + "logprob_reward": -0.050654854625463486, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.050654854625463486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1690095067024231, + 0.09082124382257462, + 0.046237844973802567, + 0.09856635332107544, + 0.07765822112560272, + 0.2688986659049988, + 0.12263207882642746, + 0.04653223231434822, + 0.028474584221839905, + 0.05116930976510048 + ], + "credit_map": [ + 0.032135553658008575, + 0.013174938969314098, + 0.018146321177482605, + 0.017894737422466278, + 0.01840629056096077, + 0.01971353404223919, + 0.028557417914271355, + 0.020980700850486755, + 0.00791985634714365, + 0.012260437943041325, + 0.008801009505987167, + 0.01670008897781372, + 0.008651378564536572, + 0.009983797557651997, + 0.008206799626350403, + 0.018297875300049782, + 0.003152536926791072, + 0.004811766091734171, + 0.0043226853013038635, + 0.004349047783762217, + 0.005632524844259024, + 0.013891772367060184, + 0.004365734290331602, + 0.005711773410439491, + 0.012019707821309566, + 0.007303962484002113, + 0.018125535920262337, + 0.008666766807436943, + 0.009866916574537754, + 0.018467949703335762, + 0.009652488864958286, + 0.014463025145232677, + 0.010515947826206684, + 0.010890170931816101, + 0.011917590163648129, + 0.01638142764568329, + 0.007367576938122511, + 0.007385632023215294, + 0.0070152259431779385, + 0.006184649188071489, + 0.03457937389612198, + 0.036926448345184326, + 0.05593656003475189, + 0.051561493426561356, + 0.02852722443640232, + 0.017496470361948013, + 0.02812594547867775, + 0.015745140612125397, + 0.00956184882670641, + 0.01743847317993641, + 0.016524283215403557, + 0.015202239155769348, + 0.027646027505397797, + 0.0114144803956151, + 0.010111769661307335, + 0.014732961542904377, + 0.003947481978684664, + 0.004640757106244564, + 0.0049728392623364925, + 0.004739442374557257, + 0.00961345061659813, + 0.0060850041918456554, + 0.0072144255973398685, + 0.005318832583725452, + 0.002205903409048915, + 0.002526462310925126, + 0.004230512771755457, + 0.0032994779758155346, + 0.002075376221910119, + 0.004272593650966883, + 0.005979281384497881, + 0.003884978126734495, + 0.005817270837724209, + 0.005403495859354734, + 0.008405344560742378, + 0.005441220477223396, + 0.004718732088804245, + 0.006695023272186518, + 0.00662454217672348, + 0.00806367676705122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.050654854625463486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.007649888750165701 + ], + "opposite_logprob_tokens": [ + -4.882649898529053 + ], + "logprob_reward": -0.007649888750165701, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649888750165701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.256714403629303, + 0.08520986884832382, + 0.06134960055351257, + 0.10272731631994247, + 0.120515376329422, + 0.11487177759408951, + 0.09449988603591919, + 0.061459362506866455, + 0.06759992241859436, + 0.03505241498351097 + ], + "credit_map": [ + 0.022396394982933998, + 0.016408752650022507, + 0.014374868012964725, + 0.017085585743188858, + 0.019773215055465698, + 0.06445858627557755, + 0.027913354337215424, + 0.07430367916822433, + 0.014041909947991371, + 0.009560373611748219, + 0.016958072781562805, + 0.009741846472024918, + 0.007859701290726662, + 0.007696267683058977, + 0.009927182458341122, + 0.009424520656466484, + 0.005262661259621382, + 0.006041763350367546, + 0.005802258849143982, + 0.006742148660123348, + 0.00844431109726429, + 0.007281893398612738, + 0.009979316033422947, + 0.011795257218182087, + 0.010060098022222519, + 0.007526714820414782, + 0.01555559877306223, + 0.009176482446491718, + 0.010875239036977291, + 0.021660590544342995, + 0.007435232400894165, + 0.020437363535165787, + 0.012917404994368553, + 0.01715160347521305, + 0.0156723540276289, + 0.015500634908676147, + 0.017765169963240623, + 0.018064241856336594, + 0.012811065651476383, + 0.010632909834384918, + 0.015396558679640293, + 0.01831413432955742, + 0.017942257225513458, + 0.01662766933441162, + 0.015260232612490654, + 0.009782982058823109, + 0.012837161310017109, + 0.008710800670087337, + 0.010224484838545322, + 0.012264581397175789, + 0.01777534931898117, + 0.013464323244988918, + 0.01484993752092123, + 0.01041906513273716, + 0.007543397601693869, + 0.007958759553730488, + 0.0048743183724582195, + 0.005885752849280834, + 0.0074878777377307415, + 0.005185895133763552, + 0.01314530149102211, + 0.008253582753241062, + 0.008918548934161663, + 0.007708093151450157, + 0.004342655185610056, + 0.005575408693403006, + 0.008378046564757824, + 0.0068578096106648445, + 0.005902778822928667, + 0.008622650057077408, + 0.0180897768586874, + 0.009830806404352188, + 0.0034762388095259666, + 0.004166707396507263, + 0.005375100299715996, + 0.003691432299092412, + 0.003324235789477825, + 0.005221580155193806, + 0.004861340392380953, + 0.004935785196721554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.007649888750165701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.00673760985955596 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.00673760985955596, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673760985955596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21243880689144135, + 0.09394653141498566, + 0.1323925405740738, + 0.10873328894376755, + 0.061873842030763626, + 0.12416861206293106, + 0.09131449460983276, + 0.10515260696411133, + 0.02231309935450554, + 0.04766610637307167 + ], + "credit_map": [ + 0.03543844446539879, + 0.018633879721164703, + 0.017212161794304848, + 0.01873713545501232, + 0.02541709505021572, + 0.03198911249637604, + 0.023728739470243454, + 0.041282281279563904, + 0.010882087051868439, + 0.01445315033197403, + 0.01501526404172182, + 0.008744402788579464, + 0.009216656908392906, + 0.010556373745203018, + 0.01158041413873434, + 0.013498196378350258, + 0.010751727968454361, + 0.010635138489305973, + 0.009526110254228115, + 0.01147555373609066, + 0.013908454217016697, + 0.04630988463759422, + 0.010755909606814384, + 0.019029781222343445, + 0.013338570483028889, + 0.010699357837438583, + 0.018199382349848747, + 0.010494686663150787, + 0.009049182757735252, + 0.022502582520246506, + 0.010112576186656952, + 0.014336966909468174, + 0.00884249433875084, + 0.00865323469042778, + 0.010250133462250233, + 0.008130199275910854, + 0.0077242739498615265, + 0.006353981792926788, + 0.005672415718436241, + 0.006247120443731546, + 0.01618039794266224, + 0.022395221516489983, + 0.016903597861528397, + 0.017123239114880562, + 0.014108018018305302, + 0.010299700312316418, + 0.015752388164401054, + 0.011406059376895428, + 0.005387828219681978, + 0.014375236816704273, + 0.009862696751952171, + 0.010287987999618053, + 0.026192914694547653, + 0.00915567297488451, + 0.007233079522848129, + 0.008819092065095901, + 0.00784334633499384, + 0.007054986897855997, + 0.01685236021876335, + 0.007595754694193602, + 0.020428184419870377, + 0.016029266640543938, + 0.012586287222802639, + 0.016762439161539078, + 0.0021532534155994654, + 0.002370185684412718, + 0.0037064512725919485, + 0.003143651643767953, + 0.002103496575728059, + 0.00260354345664382, + 0.0037301843985915184, + 0.002502336399629712, + 0.006773838307708502, + 0.0052568684332072735, + 0.0066750613041222095, + 0.004436111077666283, + 0.004715677350759506, + 0.0069090696051716805, + 0.005695758853107691, + 0.007203731220215559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.00673760985955596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.002820087829604745 + ], + "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": [ + 0.19299012422561646, + 0.1741938292980194, + 0.05816887691617012, + 0.036018889397382736, + 0.08834096789360046, + 0.14590899646282196, + 0.09348666667938232, + 0.05016542971134186, + 0.12293356657028198, + 0.03779266029596329 + ], + "credit_map": [ + 0.017388958483934402, + 0.013094711117446423, + 0.00919672753661871, + 0.009081857278943062, + 0.014873160049319267, + 0.05014174431562424, + 0.016230907291173935, + 0.06298204511404037, + 0.013866566121578217, + 0.01819595880806446, + 0.014809682033956051, + 0.03424395993351936, + 0.02127825655043125, + 0.022817669436335564, + 0.019055867567658424, + 0.029925871640443802, + 0.004031042102724314, + 0.004049460869282484, + 0.005347085185348988, + 0.004196121823042631, + 0.005592727568000555, + 0.017491383478045464, + 0.004014573059976101, + 0.013446484692394733, + 0.004455454647541046, + 0.0034954389557242393, + 0.005688496399670839, + 0.004339981824159622, + 0.0042837439104914665, + 0.0052145374938845634, + 0.004179050214588642, + 0.004362184088677168, + 0.01948556862771511, + 0.012484060600399971, + 0.009961088187992573, + 0.01634204015135765, + 0.008223469369113445, + 0.007684765383601189, + 0.006604830734431744, + 0.007555149029940367, + 0.01883167028427124, + 0.02992953173816204, + 0.02279389649629593, + 0.02169662155210972, + 0.012883885763585567, + 0.010486901737749577, + 0.016286615282297134, + 0.012999875470995903, + 0.010424596257507801, + 0.0157974474132061, + 0.016517193987965584, + 0.014716431498527527, + 0.011520488187670708, + 0.009261127561330795, + 0.007952322252094746, + 0.007297060452401638, + 0.004834243096411228, + 0.004294358193874359, + 0.00706953089684248, + 0.0037636668421328068, + 0.010552534833550453, + 0.006851790472865105, + 0.0065569281578063965, + 0.0062423767521977425, + 0.008052093908190727, + 0.00734301283955574, + 0.013112669810652733, + 0.015175853855907917, + 0.009157158434391022, + 0.01601702719926834, + 0.034868646413087845, + 0.019207101315259933, + 0.004007366020232439, + 0.0037062917836010456, + 0.006628995295614004, + 0.004016080405563116, + 0.003870604792609811, + 0.005757818464189768, + 0.004296278581023216, + 0.005509224720299244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.002820087829604745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a first-person perspective inside a hamster wheel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, polished metal bars visible passing in front of the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, continuous 360-degree roll?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do still objects outside the cage appear as elongated streaks with extreme motion blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant hues of red, blue, and yellow visible blending 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": [ + -0.00525687774643302 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.00525687774643302, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525687774643302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1844961792230606, + 0.06871140003204346, + 0.052562907338142395, + 0.05309335142374039, + 0.08247029781341553, + 0.1502857357263565, + 0.16953207552433014, + 0.13201764225959778, + 0.07415670901536942, + 0.0326736755669117 + ], + "credit_map": [ + 0.022483818233013153, + 0.012908834032714367, + 0.010989363305270672, + 0.01343423593789339, + 0.01171707920730114, + 0.049165237694978714, + 0.011108196340501308, + 0.05268941447138786, + 0.00649638008326292, + 0.007255061063915491, + 0.011240056715905666, + 0.0109281986951828, + 0.007663676515221596, + 0.009984414093196392, + 0.008138040080666542, + 0.007005565799772739, + 0.004209067672491074, + 0.009097542613744736, + 0.0075074369087815285, + 0.00706931296736002, + 0.00513231847435236, + 0.006036404520273209, + 0.0060700406320393085, + 0.0074407802894711494, + 0.00703887315467, + 0.004546414595097303, + 0.0100999865680933, + 0.005287958309054375, + 0.0053248219192028046, + 0.00719976332038641, + 0.0039111035875976086, + 0.00968442764133215, + 0.012937058694660664, + 0.013812936842441559, + 0.0095490338280797, + 0.013488673605024815, + 0.008309426717460155, + 0.009441597387194633, + 0.008082066662609577, + 0.006849509198218584, + 0.0158443134278059, + 0.031977418810129166, + 0.01869964972138405, + 0.01820581965148449, + 0.016042103990912437, + 0.01646559312939644, + 0.01472568605095148, + 0.01832515187561512, + 0.016742294654250145, + 0.0240925420075655, + 0.03146185353398323, + 0.023319195955991745, + 0.025865843519568443, + 0.016411850228905678, + 0.014088042080402374, + 0.017550457268953323, + 0.008309224620461464, + 0.011979454196989536, + 0.00969174224883318, + 0.02249995246529579, + 0.021935060620307922, + 0.02065039798617363, + 0.021468572318553925, + 0.015483233146369457, + 0.007197720929980278, + 0.004552760161459446, + 0.008378014899790287, + 0.009379977360367775, + 0.004851899575442076, + 0.015968553721904755, + 0.01226598396897316, + 0.011561800725758076, + 0.0032214426901191473, + 0.0030739286448806524, + 0.004826472606509924, + 0.002932206029072404, + 0.0034158152993768454, + 0.008104262873530388, + 0.0040100193582475185, + 0.0030895282980054617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.00525687774643302 + } + ], + "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/0000155/group_1/index_tensor([6583, 6583])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_tensor([6583, 6583])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..91b544ea3ba2ba79bccb84a0f7b85e65cbd757f6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_tensor([6583, 6583])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a mid-range shot shot at high frame rate slow motion, the camera gently pans sideways with the subject, perfectly capturing the scene in the quintessential Miyazaki aesthetic. At the center of the frame, a young Miao girl is immersed in an ancient ethnic ritual. She wears an intricate, layered silver crown intricately carved with butterflies and floral patterns. Silver chains and silver flakes dangle gently with her movements, reflecting soft, fragmented light in the air. With clear, bright black eyes and a naturally rosy complexion, her expression is solemn and serene. She wears a traditional double-breasted blouse and pleated long skirt made of dark blue cotton and linen. The collar, cuffs, and hem are covered in intricate, hand-stitched embroidery depicting mountains, rivers, birds, animals, and ancient totems. Directly before her stands a weathered Miao wooden stilt house, its dark brown wood bearing the marks of time. In the background, rolling green mountains, gently shrouded in morning mist, rise. Soft morning sunlight penetrated the mist, creating clear, Tyndall-effect beams that illuminated countless dancing golden motes in the air. The girl's steps were graceful and precise, her arms elegantly extended, her wrists deftly twisted, and her heavy skirt gracefully unfurled with her gentle spin, creating a flowing ripple of color. Each movement flowed as smoothly as flowing water, without the slightest pause or blur, creating a sacred atmosphere of tranquility, purity, and vitality.", + "seed": 15396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_6583_seed_15396_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": -0.03859227448701859, + "vlm_logprob": -0.03859227448701859, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.03859227448701859, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02446693927049637 + ], + "opposite_logprob_tokens": [ + -3.7744669914245605 + ], + "logprob_reward": -0.02446693927049637, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02446693927049637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20091743767261505, + 0.047658007591962814, + 0.03726048022508621, + 0.2014015167951584, + 0.04472436383366585, + 0.056652627885341644, + 0.12963427603244781, + 0.10322695970535278, + 0.1342696100473404, + 0.04425470903515816 + ], + "credit_map": [ + 0.016743768006563187, + 0.022986359894275665, + 0.02725899964570999, + 0.020987829193472862, + 0.03671371564269066, + 0.0243624709546566, + 0.02678120695054531, + 0.025083085522055626, + 0.0029307501390576363, + 0.003815323580056429, + 0.004794333595782518, + 0.006304481066763401, + 0.007906707003712654, + 0.007595526985824108, + 0.006808851845562458, + 0.007502033840864897, + 0.002740359865128994, + 0.002882555825635791, + 0.0041674361564219, + 0.005500913597643375, + 0.005528191104531288, + 0.005227858200669289, + 0.003516254248097539, + 0.007696912158280611, + 0.011837332509458065, + 0.014206846244633198, + 0.015946868807077408, + 0.012271414510905743, + 0.02626415342092514, + 0.0395859032869339, + 0.021434392780065536, + 0.05985460802912712, + 0.004012190271168947, + 0.0028483730275183916, + 0.00492192804813385, + 0.00591817032545805, + 0.005463914014399052, + 0.007266720291227102, + 0.005038746166974306, + 0.009254319593310356, + 0.00620295898988843, + 0.003916596062481403, + 0.00931758712977171, + 0.006051986012607813, + 0.007564470637589693, + 0.006988872773945332, + 0.007619617041200399, + 0.00899053830653429, + 0.008975404314696789, + 0.018418265506625175, + 0.015172556042671204, + 0.014459789730608463, + 0.02225634455680847, + 0.011255011893808842, + 0.022470945492386818, + 0.016625959426164627, + 0.006901513785123825, + 0.008196987211704254, + 0.013169295154511929, + 0.008545923978090286, + 0.019886555150151253, + 0.013378314673900604, + 0.014629511162638664, + 0.018518853932619095, + 0.011779585853219032, + 0.017369959503412247, + 0.01683155819773674, + 0.025055361911654472, + 0.01288189459592104, + 0.021428203210234642, + 0.015837183222174644, + 0.0130858663469553, + 0.0032225260511040688, + 0.004767400212585926, + 0.007228496950119734, + 0.00598697317764163, + 0.00757902767509222, + 0.004625926725566387, + 0.005588544066995382, + 0.0052558137103915215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.02446693927049637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091247200965881 + ], + "opposite_logprob_tokens": [ + -4.511091232299805 + ], + "logprob_reward": -0.011091247200965881, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091247200965881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26893219351768494, + 0.042823754251003265, + 0.11263407766819, + 0.11790717393159866, + 0.10286631435155869, + 0.08901840448379517, + 0.03397819772362709, + 0.08634520322084427, + 0.1019500344991684, + 0.04354464262723923 + ], + "credit_map": [ + 0.01753496378660202, + 0.02096956968307495, + 0.02726578526198864, + 0.035442207008600235, + 0.040051743388175964, + 0.03458835184574127, + 0.04943304881453514, + 0.043646544218063354, + 0.0031793885864317417, + 0.0034259348176419735, + 0.004819596651941538, + 0.004932743031531572, + 0.008404277265071869, + 0.00595158850774169, + 0.00625415425747633, + 0.005856070201843977, + 0.012100063264369965, + 0.008420979604125023, + 0.013440723530948162, + 0.012500295415520668, + 0.017222728580236435, + 0.016963792964816093, + 0.01433839462697506, + 0.017647109925746918, + 0.009146568365395069, + 0.005235753953456879, + 0.009640523232519627, + 0.0058531085960567, + 0.009417918510735035, + 0.036537930369377136, + 0.008517789654433727, + 0.033557575196027756, + 0.004185350611805916, + 0.004396296571940184, + 0.007921237498521805, + 0.0062090628780424595, + 0.007688296493142843, + 0.03182096406817436, + 0.005140660796314478, + 0.03550444543361664, + 0.008587663061916828, + 0.008900749497115612, + 0.011545414105057716, + 0.009108876809477806, + 0.01399044506251812, + 0.014554386027157307, + 0.012887117452919483, + 0.009443750604987144, + 0.0029940507374703884, + 0.0035928981378674507, + 0.003848418826237321, + 0.0029318586457520723, + 0.005700194742530584, + 0.003877022536471486, + 0.006219753064215183, + 0.004814001265913248, + 0.006944139022380114, + 0.009467968717217445, + 0.009957804344594479, + 0.01005296129733324, + 0.015813477337360382, + 0.012088875286281109, + 0.009585956111550331, + 0.012434026226401329, + 0.01416800171136856, + 0.008884777314960957, + 0.017812397330999374, + 0.015467195771634579, + 0.010464846156537533, + 0.012087244540452957, + 0.012718947604298592, + 0.01034662127494812, + 0.0040889922529459, + 0.005318982060998678, + 0.006330170668661594, + 0.004854686092585325, + 0.005391761660575867, + 0.007500733248889446, + 0.004071277566254139, + 0.005988043267279863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.011091247200965881 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815908521413803 + ], + "opposite_logprob_tokens": [ + -3.529815912246704 + ], + "logprob_reward": -0.029815908521413803, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815908521413803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20552611351013184, + 0.05086278170347214, + 0.04685871675610542, + 0.09465885162353516, + 0.17373979091644287, + 0.07204998284578323, + 0.07659300416707993, + 0.06817254424095154, + 0.1363464891910553, + 0.07519172877073288 + ], + "credit_map": [ + 0.01863113045692444, + 0.020618155598640442, + 0.022512268275022507, + 0.021990137174725533, + 0.0353570394217968, + 0.025411855429410934, + 0.03412104770541191, + 0.026884479448199272, + 0.004293651785701513, + 0.005656133871525526, + 0.005015380680561066, + 0.005872790701687336, + 0.008489670231938362, + 0.008396714925765991, + 0.0067185028456151485, + 0.00641994271427393, + 0.0033541845623403788, + 0.004324533045291901, + 0.006245817057788372, + 0.005960519891232252, + 0.008348368108272552, + 0.004812201950699091, + 0.009186281822621822, + 0.0046268124133348465, + 0.00683547742664814, + 0.0037067746743559837, + 0.009839773178100586, + 0.008864258415997028, + 0.00764048146083951, + 0.026661580428481102, + 0.005773020442575216, + 0.025337494909763336, + 0.006929117254912853, + 0.005391617305576801, + 0.014200967736542225, + 0.0102060716599226, + 0.012257298454642296, + 0.054618772119283676, + 0.008443569764494896, + 0.061692386865615845, + 0.006656661629676819, + 0.0062038106843829155, + 0.008181562647223473, + 0.007667898200452328, + 0.007031720131635666, + 0.009833586402237415, + 0.012562907300889492, + 0.013911842368543148, + 0.005394576583057642, + 0.006992696318775415, + 0.007593621965497732, + 0.00849369540810585, + 0.01343273464590311, + 0.01241876557469368, + 0.012612464837729931, + 0.009654452092945576, + 0.0056607541628181934, + 0.004196682944893837, + 0.007599557284265757, + 0.004656750708818436, + 0.012586476281285286, + 0.007810472976416349, + 0.01018359325826168, + 0.015478262677788734, + 0.01463115680962801, + 0.016813497990369797, + 0.022616691887378693, + 0.02774527668952942, + 0.013442667201161385, + 0.014530298300087452, + 0.01525419857352972, + 0.01131270918995142, + 0.0063361735083162785, + 0.007478602230548859, + 0.00966832134872675, + 0.008981370367109776, + 0.008848200552165508, + 0.01504372339695692, + 0.0076985289342701435, + 0.011136814020574093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.029815908521413803 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048636723309755325 + ], + "opposite_logprob_tokens": [ + -3.0486366748809814 + ], + "logprob_reward": -0.048636723309755325, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048636723309755325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28239378333091736, + 0.05969855561852455, + 0.014567408710718155, + 0.018808428198099136, + 0.05009807273745537, + 0.02250114642083645, + 0.031680699437856674, + 0.13668982684612274, + 0.35223183035850525, + 0.03133029118180275 + ], + "credit_map": [ + 0.015287749469280243, + 0.024690162390470505, + 0.024203181266784668, + 0.033314432948827744, + 0.05516902729868889, + 0.03477630019187927, + 0.05215410143136978, + 0.04279881715774536, + 0.003750741947442293, + 0.005292890593409538, + 0.0061952900141477585, + 0.00818599108606577, + 0.011923442594707012, + 0.007416686974465847, + 0.009721718728542328, + 0.007211790885776281, + 0.002308666706085205, + 0.0012656257022172213, + 0.0028852263931185007, + 0.0016454074066132307, + 0.0018890883075073361, + 0.0014238908188417554, + 0.0015683980891481042, + 0.0015811047051101923, + 0.0019543420057743788, + 0.0013259686529636383, + 0.0027176938019692898, + 0.0016566101694479585, + 0.001721565262414515, + 0.0035148290917277336, + 0.0012548393569886684, + 0.004662578459829092, + 0.002213607542216778, + 0.0012860294664278626, + 0.004176792223006487, + 0.0026314719580113888, + 0.0019287860486656427, + 0.01703108660876751, + 0.0024306972045451403, + 0.018399598076939583, + 0.0019387113861739635, + 0.0011650710366666317, + 0.0024762272369116545, + 0.002228607190772891, + 0.002867147559300065, + 0.0031137606129050255, + 0.003128952346742153, + 0.005582668352872133, + 0.003356854896992445, + 0.006393902003765106, + 0.0031321989372372627, + 0.002993797417730093, + 0.004501624032855034, + 0.004386818502098322, + 0.00279495632275939, + 0.004120547324419022, + 0.009408342652022839, + 0.012646950781345367, + 0.017328841611742973, + 0.012643374502658844, + 0.030272914096713066, + 0.016478732228279114, + 0.016722820699214935, + 0.021187832579016685, + 0.03218181058764458, + 0.030622392892837524, + 0.0587833970785141, + 0.052394382655620575, + 0.045827288180589676, + 0.0409093052148819, + 0.03511875122785568, + 0.05639447644352913, + 0.00225232751108706, + 0.0031560391653329134, + 0.003193202428519726, + 0.004489096812903881, + 0.006631822790950537, + 0.0035383731592446566, + 0.003893532557412982, + 0.004175896290689707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.048636723309755325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895055413246155 + ], + "opposite_logprob_tokens": [ + -2.5789506435394287 + ], + "logprob_reward": -0.07895055413246155, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895055413246155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24113643169403076, + 0.04522685334086418, + 0.026732292026281357, + 0.07715263962745667, + 0.3487156331539154, + 0.04450579360127449, + 0.03290754556655884, + 0.06880948692560196, + 0.06856023520231247, + 0.04625310003757477 + ], + "credit_map": [ + 0.011819188483059406, + 0.019525451585650444, + 0.01467388030141592, + 0.021368250250816345, + 0.04332824796438217, + 0.03946612402796745, + 0.03810194507241249, + 0.05285334214568138, + 0.002808487508445978, + 0.0036249039694666862, + 0.004064011853188276, + 0.005218535661697388, + 0.008627597242593765, + 0.0063882553949952126, + 0.005747279152274132, + 0.008747778832912445, + 0.0024840945843607187, + 0.001830870402045548, + 0.003177632112056017, + 0.0028926918748766184, + 0.004497753921896219, + 0.004024295601993799, + 0.0027355188503861427, + 0.005089432001113892, + 0.004797626752406359, + 0.0035519495140761137, + 0.005933354143053293, + 0.0030295713804662228, + 0.004787130281329155, + 0.02480805292725563, + 0.004591566044837236, + 0.025653382763266563, + 0.01265807170420885, + 0.01001362781971693, + 0.026331748813390732, + 0.016139158979058266, + 0.019111214205622673, + 0.1191544234752655, + 0.01462634839117527, + 0.13068100810050964, + 0.005565939005464315, + 0.004078839439898729, + 0.0073357149958610535, + 0.004564451519399881, + 0.00585659546777606, + 0.005540451034903526, + 0.005792554933577776, + 0.005771243013441563, + 0.0031596971675753593, + 0.0022224935237318277, + 0.004630769602954388, + 0.004055443685501814, + 0.004225807264447212, + 0.0031473725102841854, + 0.006703654304146767, + 0.004762303549796343, + 0.004210196901112795, + 0.007851981557905674, + 0.006638698745518923, + 0.007042239420115948, + 0.014924764633178711, + 0.00929954182356596, + 0.00893343985080719, + 0.009908619336783886, + 0.010765607468783855, + 0.006900170352309942, + 0.011154063045978546, + 0.009026980958878994, + 0.00746576813980937, + 0.007195066660642624, + 0.00877124909311533, + 0.007281324826180935, + 0.00409475015476346, + 0.005394202657043934, + 0.00628864299505949, + 0.005567074287682772, + 0.006220167502760887, + 0.007279280573129654, + 0.005184066016227007, + 0.0062249088659882545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.07895055413246155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a mid-range shot shot at high frame rate slow motion, the camera gently pans sideways with the subject, perfectly capturing the scene in the quintessential Miyazaki aesthetic. At the center of the frame, a young Miao girl is immersed in an ancient ethnic ritual. She wears an intricate, layered silver crown intricately carved with butterflies and floral patterns. Silver chains and silver flakes dangle gently with her movements, reflecting soft, fragmented light in the air. With clear, bright black eyes and a naturally rosy complexion, her expression is solemn and serene. She wears a traditional double-breasted blouse and pleated long skirt made of dark blue cotton and linen. The collar, cuffs, and hem are covered in intricate, hand-stitched embroidery depicting mountains, rivers, birds, animals, and ancient totems. Directly before her stands a weathered Miao wooden stilt house, its dark brown wood bearing the marks of time. In the background, rolling green mountains, gently shrouded in morning mist, rise. Soft morning sunlight penetrated the mist, creating clear, Tyndall-effect beams that illuminated countless dancing golden motes in the air. The girl's steps were graceful and precise, her arms elegantly extended, her wrists deftly twisted, and her heavy skirt gracefully unfurled with her gentle spin, creating a flowing ripple of color. Each movement flowed as smoothly as flowing water, without the slightest pause or blur, creating a sacred atmosphere of tranquility, purity, and vitality.", + "seed": 15397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_6583_seed_15397_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": -0.09872606266289949, + "vlm_logprob": -0.09872606266289949, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.09872606266289949, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00914692971855402 + ], + "opposite_logprob_tokens": [ + -4.8841471672058105 + ], + "logprob_reward": -0.00914692971855402, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00914692971855402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1885150969028473, + 0.18401706218719482, + 0.07169574499130249, + 0.11042794585227966, + 0.10146258026361465, + 0.042691946029663086, + 0.028754111379384995, + 0.100705087184906, + 0.11523381620645523, + 0.05649658292531967 + ], + "credit_map": [ + 0.030847758054733276, + 0.01900719664990902, + 0.027050770819187164, + 0.024421462789177895, + 0.02077905647456646, + 0.019757354632019997, + 0.027109025046229362, + 0.019542476162314415, + 0.031214706599712372, + 0.01744792051613331, + 0.0358722023665905, + 0.01802513562142849, + 0.020379122346639633, + 0.01723308488726616, + 0.024210067465901375, + 0.01963481307029724, + 0.009884431958198547, + 0.00913192518055439, + 0.006445142440497875, + 0.008385572582483292, + 0.013984281569719315, + 0.008474010974168777, + 0.009370412677526474, + 0.006019964814186096, + 0.013165434822440147, + 0.014929568395018578, + 0.007564160507172346, + 0.011623654514551163, + 0.025875652208924294, + 0.014449077658355236, + 0.015076030045747757, + 0.007744363974779844, + 0.011022583581507206, + 0.012296724133193493, + 0.006369837559759617, + 0.010042645037174225, + 0.029998943209648132, + 0.015438972972333431, + 0.00785532034933567, + 0.008437548764050007, + 0.00457897549495101, + 0.0049219257198274136, + 0.004945887252688408, + 0.006215990986675024, + 0.007487318944185972, + 0.0051928251050412655, + 0.004923087544739246, + 0.004425935447216034, + 0.003463335335254669, + 0.004042735323309898, + 0.0031071321573108435, + 0.004496414214372635, + 0.004342638421803713, + 0.0031632492318749428, + 0.0034648175351321697, + 0.0026737882290035486, + 0.010063877329230309, + 0.0129357585683465, + 0.02043072134256363, + 0.010216857306659222, + 0.0147422319278121, + 0.012924693524837494, + 0.009214797988533974, + 0.010176151990890503, + 0.01484739687293768, + 0.012208931148052216, + 0.029984235763549805, + 0.013289417140185833, + 0.01683821901679039, + 0.01205603126436472, + 0.007601839490234852, + 0.008407752960920334, + 0.004838918335735798, + 0.011132106184959412, + 0.007274874020367861, + 0.0059150210581719875, + 0.007866894826292992, + 0.009030858054757118, + 0.005873965565115213, + 0.004563942551612854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.00914692971855402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091010645031929 + ], + "opposite_logprob_tokens": [ + -4.511091232299805 + ], + "logprob_reward": -0.011091010645031929, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091010645031929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22389288246631622, + 0.18492738902568817, + 0.12007489800453186, + 0.06449159234762192, + 0.059907592833042145, + 0.04581593722105026, + 0.02515752986073494, + 0.16832762956619263, + 0.043498411774635315, + 0.06390617042779922 + ], + "credit_map": [ + 0.03102334775030613, + 0.03565235435962677, + 0.019036024808883667, + 0.02547229267656803, + 0.04396599158644676, + 0.02635531686246395, + 0.026523923501372337, + 0.015863627195358276, + 0.027621788904070854, + 0.016878541558980942, + 0.02083435095846653, + 0.016772089526057243, + 0.022640300914645195, + 0.01928343065083027, + 0.029564060270786285, + 0.03133281692862511, + 0.01278689969331026, + 0.012790917418897152, + 0.008922700770199299, + 0.01092201191931963, + 0.03499630466103554, + 0.017723260447382927, + 0.013828746043145657, + 0.0081040533259511, + 0.008982354775071144, + 0.01361737959086895, + 0.006107649300247431, + 0.008003036491572857, + 0.007029619067907333, + 0.006437490228563547, + 0.007772861514240503, + 0.006541194394230843, + 0.008232053369283676, + 0.007593988906592131, + 0.00555360596626997, + 0.005753674078732729, + 0.013871883042156696, + 0.007808004505932331, + 0.005206421483308077, + 0.005887961946427822, + 0.005490171257406473, + 0.005560022313147783, + 0.006025379989296198, + 0.006349869538098574, + 0.006463770288974047, + 0.005209258757531643, + 0.005935136694461107, + 0.0047823283821344376, + 0.0024691836442798376, + 0.003626760793849826, + 0.0017887636786326766, + 0.0028336155228316784, + 0.005685006733983755, + 0.003262540092691779, + 0.002790096215903759, + 0.0027015639934688807, + 0.01270431187003851, + 0.01911979727447033, + 0.0223966296762228, + 0.02849395014345646, + 0.028410572558641434, + 0.020323941484093666, + 0.019478246569633484, + 0.01740017533302307, + 0.004243273288011551, + 0.0041282931342720985, + 0.0035535639617592096, + 0.005796941928565502, + 0.008623008616268635, + 0.007095617242157459, + 0.004029988311231136, + 0.006027723662555218, + 0.013242466375231743, + 0.00957778561860323, + 0.005478386301547289, + 0.00412291893735528, + 0.006943881511688232, + 0.015196423046290874, + 0.00489794509485364, + 0.0044463626109063625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.011091010645031929 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199479207396507 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.018199479207396507, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199479207396507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23959308862686157, + 0.11500769108533859, + 0.08749429881572723, + 0.1047988161444664, + 0.04783997684717178, + 0.06139425188302994, + 0.03439083322882652, + 0.10295277833938599, + 0.06422355026006699, + 0.1423046886920929 + ], + "credit_map": [ + 0.03322243317961693, + 0.03434694930911064, + 0.016856824979186058, + 0.01810791715979576, + 0.049156833440065384, + 0.030453288927674294, + 0.03255883976817131, + 0.02489001862704754, + 0.01855987124145031, + 0.010087933391332626, + 0.018177399411797523, + 0.015498160384595394, + 0.012611448764801025, + 0.010229969397187233, + 0.014723103493452072, + 0.01511981151998043, + 0.013214277103543282, + 0.00898749753832817, + 0.008956638164818287, + 0.011502591893076897, + 0.01703845150768757, + 0.01058888342231512, + 0.010574768297374249, + 0.006631188094615936, + 0.010952752083539963, + 0.014665862545371056, + 0.008100176230072975, + 0.013433093205094337, + 0.01412538904696703, + 0.01345957349985838, + 0.023715896531939507, + 0.0063460757955908775, + 0.007164833601564169, + 0.007608415093272924, + 0.004936710000038147, + 0.003870110958814621, + 0.0098135219886899, + 0.005794386379420757, + 0.0032557661179453135, + 0.005396238528192043, + 0.006254231557250023, + 0.008062480948865414, + 0.008423018269240856, + 0.008338753134012222, + 0.008521889336407185, + 0.006651570089161396, + 0.008980345912277699, + 0.00616196496412158, + 0.0037342170253396034, + 0.004318946041166782, + 0.004064961336553097, + 0.006351934280246496, + 0.005009017419070005, + 0.0038319560699164867, + 0.0037411206867545843, + 0.003338680835440755, + 0.008658112958073616, + 0.01423158310353756, + 0.011318058706820011, + 0.014439697377383709, + 0.01845623552799225, + 0.013582532294094563, + 0.012904410250484943, + 0.009362149983644485, + 0.0069956728257238865, + 0.006387907080352306, + 0.007536705117672682, + 0.00576970586553216, + 0.012952095828950405, + 0.0077370875515043736, + 0.00783105380833149, + 0.009013327769935131, + 0.018876736983656883, + 0.02155398763716221, + 0.021404238417744637, + 0.01848413236439228, + 0.018268832936882973, + 0.017721638083457947, + 0.011157335713505745, + 0.014837787486612797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.018199479207396507 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42886051535606384 + ], + "opposite_logprob_tokens": [ + -1.0538605451583862 + ], + "logprob_reward": -0.42886051535606384, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42886051535606384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33918556571006775, + 0.1052188128232956, + 0.045298755168914795, + 0.09888152778148651, + 0.07689960300922394, + 0.04638124257326126, + 0.039718806743621826, + 0.14890779554843903, + 0.06579919904470444, + 0.03370872512459755 + ], + "credit_map": [ + 0.04763304069638252, + 0.03958716616034508, + 0.053216684609651566, + 0.04410432651638985, + 0.05393630266189575, + 0.030858803540468216, + 0.046690814197063446, + 0.023158421739935875, + 0.02320428006350994, + 0.010800330899655819, + 0.017924580723047256, + 0.00810230616480112, + 0.013812458142638206, + 0.011818417347967625, + 0.011405919678509235, + 0.008150518871843815, + 0.006422263104468584, + 0.005771308671683073, + 0.005815217271447182, + 0.0050426991656422615, + 0.009830966591835022, + 0.005370384082198143, + 0.004244059324264526, + 0.002801851835101843, + 0.021777931600809097, + 0.016286160796880722, + 0.009796918369829655, + 0.008883127942681313, + 0.011188899166882038, + 0.011402016505599022, + 0.007636116351932287, + 0.011910364031791687, + 0.00925721600651741, + 0.010712150484323502, + 0.007312940899282694, + 0.00883445329964161, + 0.012956520542502403, + 0.008593723177909851, + 0.006074109580367804, + 0.013158487156033516, + 0.006301363930106163, + 0.006565602496266365, + 0.00635431706905365, + 0.004856014624238014, + 0.007912767119705677, + 0.005094792228192091, + 0.0053491597063839436, + 0.003947223071008921, + 0.0076037500984966755, + 0.007977788336575031, + 0.004291259218007326, + 0.004083102103322744, + 0.004764297511428595, + 0.003920266404747963, + 0.003440435975790024, + 0.0036379096563905478, + 0.011857106350362301, + 0.01850806549191475, + 0.02557132951915264, + 0.023516755551099777, + 0.02352549321949482, + 0.01903250254690647, + 0.01711125299334526, + 0.00978529080748558, + 0.00635948171839118, + 0.010174190625548363, + 0.00784645602107048, + 0.0047236597165465355, + 0.015357245691120625, + 0.007486844435334206, + 0.007578998804092407, + 0.006272324360907078, + 0.007963764481246471, + 0.0036145146004855633, + 0.004218039568513632, + 0.0032798105385154486, + 0.0031037277076393366, + 0.006887351628392935, + 0.0027646266389638186, + 0.0018768886802718043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.42886051535606384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026332378387451172 + ], + "opposite_logprob_tokens": [ + -3.651332378387451 + ], + "logprob_reward": -0.026332378387451172, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332378387451172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.205409973859787, + 0.15218167006969452, + 0.08428479731082916, + 0.17464111745357513, + 0.045652393251657486, + 0.029615595936775208, + 0.031064555048942566, + 0.10965610295534134, + 0.10871272534132004, + 0.05878102779388428 + ], + "credit_map": [ + 0.02645624428987503, + 0.0345500148832798, + 0.022763580083847046, + 0.020342892035841942, + 0.043863292783498764, + 0.022888129577040672, + 0.022284293547272682, + 0.012261539697647095, + 0.021626846864819527, + 0.015795422717928886, + 0.015014293603599072, + 0.012480121105909348, + 0.029533768072724342, + 0.01827942207455635, + 0.02039264142513275, + 0.019059155136346817, + 0.010617625899612904, + 0.009106091223657131, + 0.009081905707716942, + 0.011237611062824726, + 0.01684170961380005, + 0.008525297045707703, + 0.010205877013504505, + 0.008668689988553524, + 0.020275183022022247, + 0.020820748060941696, + 0.017868373543024063, + 0.023213258013129234, + 0.03414159268140793, + 0.020341655239462852, + 0.02490631863474846, + 0.013073998503386974, + 0.006576197687536478, + 0.007092072628438473, + 0.003946845419704914, + 0.003561705583706498, + 0.007303537800908089, + 0.005465204827487469, + 0.0042157541029155254, + 0.007491075899451971, + 0.004042927175760269, + 0.0035828573163598776, + 0.002978217089548707, + 0.004084462765604258, + 0.0056027635000646114, + 0.0037370172794908285, + 0.0026458879001438618, + 0.0029414640739560127, + 0.0035528913140296936, + 0.0032137183006852865, + 0.0030031073838472366, + 0.005339331924915314, + 0.005361310206353664, + 0.0035755110438913107, + 0.0036196457222104073, + 0.0033990414813160896, + 0.009291944094002247, + 0.01309603825211525, + 0.010658117942512035, + 0.013419597409665585, + 0.02360599860548973, + 0.015695229172706604, + 0.012208253145217896, + 0.011680934578180313, + 0.011341297999024391, + 0.012637751176953316, + 0.012378828600049019, + 0.011443047784268856, + 0.026495350524783134, + 0.013455688022077084, + 0.009705496951937675, + 0.01125528197735548, + 0.009528791531920433, + 0.0068972171284258366, + 0.007252698298543692, + 0.00563711766153574, + 0.009969847276806831, + 0.0070464350283145905, + 0.007306377869099379, + 0.005142546724528074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.026332378387451172 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a mid-range shot shot at high frame rate slow motion, the camera gently pans sideways with the subject, perfectly capturing the scene in the quintessential Miyazaki aesthetic. At the center of the frame, a young Miao girl is immersed in an ancient ethnic ritual. She wears an intricate, layered silver crown intricately carved with butterflies and floral patterns. Silver chains and silver flakes dangle gently with her movements, reflecting soft, fragmented light in the air. With clear, bright black eyes and a naturally rosy complexion, her expression is solemn and serene. She wears a traditional double-breasted blouse and pleated long skirt made of dark blue cotton and linen. The collar, cuffs, and hem are covered in intricate, hand-stitched embroidery depicting mountains, rivers, birds, animals, and ancient totems. Directly before her stands a weathered Miao wooden stilt house, its dark brown wood bearing the marks of time. In the background, rolling green mountains, gently shrouded in morning mist, rise. Soft morning sunlight penetrated the mist, creating clear, Tyndall-effect beams that illuminated countless dancing golden motes in the air. The girl's steps were graceful and precise, her arms elegantly extended, her wrists deftly twisted, and her heavy skirt gracefully unfurled with her gentle spin, creating a flowing ripple of color. Each movement flowed as smoothly as flowing water, without the slightest pause or blur, creating a sacred atmosphere of tranquility, purity, and vitality.", + "seed": 15398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_6583_seed_15398_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": -0.024925540760159494, + "vlm_logprob": -0.024925540760159494, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.024925540760159494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008552944287657738 + ], + "opposite_logprob_tokens": [ + -4.8835530281066895 + ], + "logprob_reward": -0.008552944287657738, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008552944287657738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10527987033128738, + 0.050003621727228165, + 0.08195746690034866, + 0.15968897938728333, + 0.32377389073371887, + 0.049661073833703995, + 0.06544803082942963, + 0.05137110501527786, + 0.03707369789481163, + 0.07574227452278137 + ], + "credit_map": [ + 0.01372138224542141, + 0.01320846751332283, + 0.018268318846821785, + 0.0086613604798913, + 0.01913375034928322, + 0.009237157180905342, + 0.013248201459646225, + 0.009801238775253296, + 0.006005957722663879, + 0.004552575759589672, + 0.00590266939252615, + 0.0037162560038268566, + 0.01208852231502533, + 0.006097724661231041, + 0.0068887039087712765, + 0.004751213826239109, + 0.010678824037313461, + 0.013990100473165512, + 0.009334934875369072, + 0.009536273777484894, + 0.011839814484119415, + 0.007989367470145226, + 0.009954003617167473, + 0.008634161204099655, + 0.023339351639151573, + 0.009424274787306786, + 0.04385050758719444, + 0.013733970932662487, + 0.011708320118486881, + 0.030909990891814232, + 0.012440036982297897, + 0.014282523654401302, + 0.07504887878894806, + 0.013179427944123745, + 0.13946843147277832, + 0.021538082510232925, + 0.011792637407779694, + 0.03313278779387474, + 0.010178019292652607, + 0.01943562924861908, + 0.010465673170983791, + 0.006455669179558754, + 0.011941525153815746, + 0.0034046401269733906, + 0.003933309111744165, + 0.0037596486508846283, + 0.0037265203427523375, + 0.0059740906581282616, + 0.0031038066372275352, + 0.004891462158411741, + 0.012134137563407421, + 0.025954267010092735, + 0.005674297921359539, + 0.006445171311497688, + 0.003706119954586029, + 0.0035387754905968904, + 0.0054774209856987, + 0.005290298257023096, + 0.00765776913613081, + 0.004470171872526407, + 0.0052837952971458435, + 0.0051228380762040615, + 0.00697513110935688, + 0.011093683540821075, + 0.00659717433154583, + 0.00758933974429965, + 0.0038025269750505686, + 0.00439003761857748, + 0.004572974517941475, + 0.003136197803542018, + 0.003891601925715804, + 0.003093848703429103, + 0.005046632140874863, + 0.02628173865377903, + 0.004077537450939417, + 0.004353211261332035, + 0.0047165220603346825, + 0.005422352813184261, + 0.003808035282418132, + 0.022036246955394745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.008552944287657738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17459537088871002, + 0.04748398810625076, + 0.12593932449817657, + 0.16130419075489044, + 0.10111670941114426, + 0.07960031181573868, + 0.03782283514738083, + 0.07240116596221924, + 0.07041776925325394, + 0.12931831181049347 + ], + "credit_map": [ + 0.021995168179273605, + 0.024633601307868958, + 0.024726558476686478, + 0.03409944474697113, + 0.023412838578224182, + 0.012667794711887836, + 0.016583625227212906, + 0.016476353630423546, + 0.005112885031849146, + 0.004943403881043196, + 0.009563586674630642, + 0.007051967084407806, + 0.006215013097971678, + 0.004730219952762127, + 0.004804082680493593, + 0.005062830168753862, + 0.014664880931377411, + 0.019357386976480484, + 0.014293583109974861, + 0.019348135218024254, + 0.017844650894403458, + 0.011268073692917824, + 0.01333832461386919, + 0.01582428440451622, + 0.013138440437614918, + 0.012482866644859314, + 0.025858363136649132, + 0.02129547670483589, + 0.029276428744196892, + 0.016271943226456642, + 0.026109322905540466, + 0.016871344298124313, + 0.021427109837532043, + 0.008393041789531708, + 0.03954692929983139, + 0.006057399325072765, + 0.004482494667172432, + 0.010651566088199615, + 0.004098603967577219, + 0.006459565367549658, + 0.014538697898387909, + 0.008461013436317444, + 0.019305896013975143, + 0.008350335992872715, + 0.004550371319055557, + 0.007776852231472731, + 0.005964183714240789, + 0.010652962140738964, + 0.004787640180438757, + 0.004738359246402979, + 0.003712898353114724, + 0.011493174359202385, + 0.003287476720288396, + 0.004105749540030956, + 0.003019147552549839, + 0.0026783915236592293, + 0.009148121811449528, + 0.006044487003237009, + 0.00723289093002677, + 0.0075319549068808556, + 0.006306443363428116, + 0.009840103797614574, + 0.011997932568192482, + 0.014299232512712479, + 0.011700402945280075, + 0.010610228404402733, + 0.011075018905103207, + 0.007611684966832399, + 0.012201225385069847, + 0.006140390411019325, + 0.005966314114630222, + 0.005112512968480587, + 0.019754942506551743, + 0.03391536325216293, + 0.014143584296107292, + 0.008757448755204678, + 0.008354168385267258, + 0.009188453666865826, + 0.009189227595925331, + 0.026015115901827812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160689540207386 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160689540207386, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160689540207386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09775539487600327, + 0.041820671409368515, + 0.16659216582775116, + 0.05056146904826164, + 0.13417983055114746, + 0.038444217294454575, + 0.0553257055580616, + 0.07011260092258453, + 0.1138683333992958, + 0.23133960366249084 + ], + "credit_map": [ + 0.012562958523631096, + 0.01089881919324398, + 0.011566161178052425, + 0.010697985999286175, + 0.021129392087459564, + 0.008451673202216625, + 0.012324591167271137, + 0.01012381725013256, + 0.0050435373559594154, + 0.007066997699439526, + 0.005772938020527363, + 0.004085715394467115, + 0.0076154121197760105, + 0.003459224244579673, + 0.005170871037989855, + 0.0036059755366295576, + 0.02467634528875351, + 0.033193815499544144, + 0.021286720409989357, + 0.02192765660583973, + 0.019271906465291977, + 0.010216976515948772, + 0.02353237196803093, + 0.01248636469244957, + 0.006245406344532967, + 0.008010179735720158, + 0.006444101687520742, + 0.008135017938911915, + 0.006143794395029545, + 0.006209046579897404, + 0.003932955674827099, + 0.005440967623144388, + 0.02676093764603138, + 0.01248551532626152, + 0.04817211627960205, + 0.010714282281696796, + 0.005883833859115839, + 0.012253317050635815, + 0.006836506072431803, + 0.011073309928178787, + 0.004253122955560684, + 0.004471598658710718, + 0.0074498821049928665, + 0.003620575647801161, + 0.0039252350106835365, + 0.004683347884565592, + 0.0032345964573323727, + 0.006805861834436655, + 0.006002300418913364, + 0.0024635393638163805, + 0.006827561184763908, + 0.02447452023625374, + 0.002568603726103902, + 0.00690044229850173, + 0.0030409623868763447, + 0.0030477819964289665, + 0.006783097982406616, + 0.005407922435551882, + 0.007917282171547413, + 0.006418897770345211, + 0.004877174738794565, + 0.0070120966993272305, + 0.014969377778470516, + 0.016726750880479813, + 0.018327215686440468, + 0.024646446108818054, + 0.012246550992131233, + 0.013687539845705032, + 0.015549213625490665, + 0.009296847507357597, + 0.011105778627097607, + 0.009008743800222874, + 0.016037680208683014, + 0.07649724930524826, + 0.01427374966442585, + 0.012664140202105045, + 0.01617414318025112, + 0.016666701063513756, + 0.010545068420469761, + 0.06848087906837463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.0160689540207386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999225169420242 + ], + "opposite_logprob_tokens": [ + -2.6949923038482666 + ], + "logprob_reward": -0.06999225169420242, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999225169420242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15906906127929688, + 0.04903867095708847, + 0.07108305394649506, + 0.032200369983911514, + 0.15451094508171082, + 0.0627056360244751, + 0.1277095377445221, + 0.22623957693576813, + 0.05965336412191391, + 0.05778972804546356 + ], + "credit_map": [ + 0.020989222452044487, + 0.014282765798270702, + 0.043058089911937714, + 0.0234430693089962, + 0.02005797065794468, + 0.011425462551414967, + 0.014148433692753315, + 0.011664047837257385, + 0.006366388406604528, + 0.009756400249898434, + 0.010171012952923775, + 0.004959950689226389, + 0.006441005505621433, + 0.003929695580154657, + 0.004343530163168907, + 0.003070687409490347, + 0.007296107243746519, + 0.0047312513925135136, + 0.016048209741711617, + 0.0144661208614707, + 0.010845186188817024, + 0.005600528791546822, + 0.005977075081318617, + 0.006118574645370245, + 0.00403056014329195, + 0.006050353404134512, + 0.004373992793262005, + 0.0034646845888346434, + 0.004219168797135353, + 0.004543035291135311, + 0.0027445845771580935, + 0.0027739889919757843, + 0.03410233557224274, + 0.010480465367436409, + 0.059665534645318985, + 0.011247091926634312, + 0.0074269697070121765, + 0.015400417149066925, + 0.006819375324994326, + 0.009368766099214554, + 0.005642178002744913, + 0.007727271877229214, + 0.01272400189191103, + 0.005413148086518049, + 0.008812515996396542, + 0.008552764542400837, + 0.004785363562405109, + 0.009048386476933956, + 0.014970234595239162, + 0.007048540282994509, + 0.014468550682067871, + 0.057248011231422424, + 0.0045156292617321014, + 0.016304422169923782, + 0.007231955882161856, + 0.005922190845012665, + 0.02540505677461624, + 0.011390665546059608, + 0.03166745975613594, + 0.017973925918340683, + 0.015561491250991821, + 0.025932686403393745, + 0.042413026094436646, + 0.05589526891708374, + 0.010678980499505997, + 0.014676541090011597, + 0.006370948161929846, + 0.0052855005487799644, + 0.005342164076864719, + 0.005627977661788464, + 0.00474395789206028, + 0.006927290931344032, + 0.004844377748668194, + 0.019409505650401115, + 0.0036914609372615814, + 0.0030180574394762516, + 0.0035681554581969976, + 0.004051234573125839, + 0.002548693912103772, + 0.016658244654536247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.06999225169420242 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023269904777407646 + ], + "opposite_logprob_tokens": [ + -3.7732698917388916 + ], + "logprob_reward": -0.023269904777407646, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023269904777407646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1327824890613556, + 0.055143434554338455, + 0.04493720829486847, + 0.1445079743862152, + 0.2557894289493561, + 0.06481693685054779, + 0.03575264289975166, + 0.10216725617647171, + 0.04020492732524872, + 0.1238977313041687 + ], + "credit_map": [ + 0.01846601627767086, + 0.013131796382367611, + 0.015540266409516335, + 0.011289801448583603, + 0.03253177925944328, + 0.012758873403072357, + 0.013770509511232376, + 0.015293445438146591, + 0.0065359859727323055, + 0.004303827881813049, + 0.007558004464954138, + 0.003959120251238346, + 0.012613855302333832, + 0.007133462466299534, + 0.0077807242050766945, + 0.005258447490632534, + 0.0042611719109117985, + 0.005641052033752203, + 0.00750312302261591, + 0.008999211713671684, + 0.006601804401725531, + 0.0031933810096234083, + 0.00425154622644186, + 0.004485921002924442, + 0.018768789246678352, + 0.013452992774546146, + 0.03528143838047981, + 0.012979882769286633, + 0.010192946530878544, + 0.030501773580908775, + 0.008701667189598083, + 0.014628495089709759, + 0.053070954978466034, + 0.01807212643325329, + 0.10223409533500671, + 0.018182730302214622, + 0.011973865330219269, + 0.02371377870440483, + 0.010748970322310925, + 0.01779290847480297, + 0.01072605699300766, + 0.0075037600472569466, + 0.018756527453660965, + 0.004821158479899168, + 0.006944241002202034, + 0.005265682470053434, + 0.004836915992200375, + 0.005962598603218794, + 0.003933030646294355, + 0.0022741081193089485, + 0.004901152104139328, + 0.013473022729158401, + 0.0028449916280806065, + 0.004054397810250521, + 0.002217982430011034, + 0.0020539569668471813, + 0.009482678957283497, + 0.011247146874666214, + 0.00865578930824995, + 0.013951477594673634, + 0.015921056270599365, + 0.009036309085786343, + 0.014721215702593327, + 0.01915157586336136, + 0.0056131803430616856, + 0.004871842451393604, + 0.005173865705728531, + 0.00577401602640748, + 0.006595284678041935, + 0.0041543301194906235, + 0.004405909683555365, + 0.003616499016061425, + 0.010589689016342163, + 0.04005376622080803, + 0.00860920362174511, + 0.0072592743672430515, + 0.009369252249598503, + 0.009194968268275261, + 0.0061237867921590805, + 0.03269778937101364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.023269904777407646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a mid-range shot shot at high frame rate slow motion, the camera gently pans sideways with the subject, perfectly capturing the scene in the quintessential Miyazaki aesthetic. At the center of the frame, a young Miao girl is immersed in an ancient ethnic ritual. She wears an intricate, layered silver crown intricately carved with butterflies and floral patterns. Silver chains and silver flakes dangle gently with her movements, reflecting soft, fragmented light in the air. With clear, bright black eyes and a naturally rosy complexion, her expression is solemn and serene. She wears a traditional double-breasted blouse and pleated long skirt made of dark blue cotton and linen. The collar, cuffs, and hem are covered in intricate, hand-stitched embroidery depicting mountains, rivers, birds, animals, and ancient totems. Directly before her stands a weathered Miao wooden stilt house, its dark brown wood bearing the marks of time. In the background, rolling green mountains, gently shrouded in morning mist, rise. Soft morning sunlight penetrated the mist, creating clear, Tyndall-effect beams that illuminated countless dancing golden motes in the air. The girl's steps were graceful and precise, her arms elegantly extended, her wrists deftly twisted, and her heavy skirt gracefully unfurled with her gentle spin, creating a flowing ripple of color. Each movement flowed as smoothly as flowing water, without the slightest pause or blur, creating a sacred atmosphere of tranquility, purity, and vitality.", + "seed": 15399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_1/index_6583_seed_15399_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": -0.03834907505661249, + "vlm_logprob": -0.03834907505661249, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03834907505661249, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00767354853451252 + ], + "opposite_logprob_tokens": [ + -5.132673740386963 + ], + "logprob_reward": -0.00767354853451252, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00767354853451252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2012842893600464, + 0.08418542146682739, + 0.08663924783468246, + 0.07909110933542252, + 0.04634486883878708, + 0.043584611266851425, + 0.08154808729887009, + 0.09182126820087433, + 0.11566250771284103, + 0.16983862221240997 + ], + "credit_map": [ + 0.033233899623155594, + 0.02089497074484825, + 0.02454913966357708, + 0.030681956559419632, + 0.02687220275402069, + 0.021388541907072067, + 0.025015324354171753, + 0.018648263067007065, + 0.017811710014939308, + 0.015392757952213287, + 0.012374445796012878, + 0.00802512839436531, + 0.011593134142458439, + 0.006237075664103031, + 0.005091015249490738, + 0.007660153787583113, + 0.007706240750849247, + 0.016917865723371506, + 0.015496782027184963, + 0.010727927088737488, + 0.01711229421198368, + 0.007491062395274639, + 0.005177421495318413, + 0.006009652744978666, + 0.009302058257162571, + 0.012916258536279202, + 0.016132479533553123, + 0.009656984359025955, + 0.011699600145220757, + 0.0062504601664841175, + 0.0071230679750442505, + 0.006010205950587988, + 0.004411264788359404, + 0.006193670444190502, + 0.007784335408359766, + 0.005767204333096743, + 0.007655941881239414, + 0.004010310862213373, + 0.004292763769626617, + 0.006229378283023834, + 0.004965862724930048, + 0.007101662922650576, + 0.005865021143108606, + 0.005685082636773586, + 0.007078020367771387, + 0.004055376164615154, + 0.003991043195128441, + 0.004842543974518776, + 0.013301817700266838, + 0.01133181806653738, + 0.012512613087892532, + 0.008578489534556866, + 0.012557116337120533, + 0.007180973421782255, + 0.009070810861885548, + 0.007014441303908825, + 0.010585110634565353, + 0.007655498571693897, + 0.016975564882159233, + 0.005942200310528278, + 0.020719332620501518, + 0.015100895427167416, + 0.008116309531033039, + 0.0067263515666127205, + 0.010776962153613567, + 0.006507696583867073, + 0.013569638133049011, + 0.00968074519187212, + 0.026514461264014244, + 0.02750426158308983, + 0.009825545363128185, + 0.011283191852271557, + 0.011765391565859318, + 0.010638621635735035, + 0.008515455760061741, + 0.006743662524968386, + 0.05930880457162857, + 0.05750066787004471, + 0.008185271173715591, + 0.007180747110396624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.00767354853451252 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794392623007298 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794392623007298, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794392623007298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11332236230373383, + 0.08182068169116974, + 0.04267452657222748, + 0.03840591385960579, + 0.03827580437064171, + 0.056997817009687424, + 0.09928891807794571, + 0.10453159362077713, + 0.09004820883274078, + 0.3346341848373413 + ], + "credit_map": [ + 0.016027964651584625, + 0.015545640140771866, + 0.01982412301003933, + 0.009336697869002819, + 0.019648153334856033, + 0.010611713863909245, + 0.011368343606591225, + 0.010959727689623833, + 0.008414936251938343, + 0.016707733273506165, + 0.01019712258130312, + 0.00813545472919941, + 0.015653913840651512, + 0.006500150542706251, + 0.007334443740546703, + 0.008876929059624672, + 0.004571075551211834, + 0.00737935584038496, + 0.006735327187925577, + 0.0068939789198338985, + 0.007567716296762228, + 0.0031001020688563585, + 0.0027606242801994085, + 0.003666350618004799, + 0.004104308318346739, + 0.006604984402656555, + 0.0071754599921405315, + 0.007478838320821524, + 0.004178465344011784, + 0.002809775760397315, + 0.002993646077811718, + 0.0030604349449276924, + 0.005597472190856934, + 0.005769768264144659, + 0.004710133653134108, + 0.003665830008685589, + 0.005902035161852837, + 0.0034365621395409107, + 0.004014371428638697, + 0.005179631523787975, + 0.004738619551062584, + 0.007552765775471926, + 0.0066698212176561356, + 0.007921424694359303, + 0.01020792592316866, + 0.00624782033264637, + 0.006995157338678837, + 0.006664282642304897, + 0.008859152905642986, + 0.01619226671755314, + 0.011291076429188251, + 0.010643089190125465, + 0.01976875774562359, + 0.01000868622213602, + 0.014430887065827847, + 0.008094998076558113, + 0.01111577358096838, + 0.009015281684696674, + 0.016399187967181206, + 0.008977947756648064, + 0.020657802000641823, + 0.02051386423408985, + 0.009215114638209343, + 0.008636618964374065, + 0.006881910841912031, + 0.004555180668830872, + 0.012771441601216793, + 0.007053311914205551, + 0.01786039210855961, + 0.021026408299803734, + 0.01014797668904066, + 0.009751588106155396, + 0.024060241878032684, + 0.013546384871006012, + 0.0096299322322011, + 0.009906573221087456, + 0.1294459104537964, + 0.12707144021987915, + 0.009522375650703907, + 0.01145134586840868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.009794392623007298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296458646655083 + ], + "opposite_logprob_tokens": [ + -3.773296356201172 + ], + "logprob_reward": -0.023296458646655083, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296458646655083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14041398465633392, + 0.08170125633478165, + 0.05345021188259125, + 0.058170709758996964, + 0.0332920216023922, + 0.11074560880661011, + 0.10121363401412964, + 0.23172400891780853, + 0.16097234189510345, + 0.028316248208284378 + ], + "credit_map": [ + 0.013126823119819164, + 0.02399536967277527, + 0.03479966148734093, + 0.010249975137412548, + 0.02068554237484932, + 0.008813275024294853, + 0.016125382855534554, + 0.012617952190339565, + 0.016917357221245766, + 0.013789962977170944, + 0.010584152303636074, + 0.008477919735014439, + 0.014985562302172184, + 0.005286571104079485, + 0.005799444857984781, + 0.00586028303951025, + 0.0052640619687736034, + 0.010391337797045708, + 0.009725960902869701, + 0.010711660608649254, + 0.006940638180822134, + 0.0037984787486493587, + 0.003606013022363186, + 0.0030120613519102335, + 0.006851335521787405, + 0.009229290299117565, + 0.014786383137106895, + 0.008538969792425632, + 0.005944550968706608, + 0.004120948724448681, + 0.004302680492401123, + 0.004396546166390181, + 0.003948675002902746, + 0.004325907677412033, + 0.004650453571230173, + 0.003595929592847824, + 0.004980659577995539, + 0.0030669402331113815, + 0.0031079589389264584, + 0.005615494213998318, + 0.012405455112457275, + 0.010508740320801735, + 0.022652501240372658, + 0.01619870588183403, + 0.013656031340360641, + 0.013281837105751038, + 0.011801696382462978, + 0.010240637697279453, + 0.012099242769181728, + 0.02051670104265213, + 0.01178719662129879, + 0.010965543799102306, + 0.01703905127942562, + 0.009043848142027855, + 0.010287984274327755, + 0.0094740716740489, + 0.019291291013360023, + 0.012806287035346031, + 0.012441548518836498, + 0.009209934622049332, + 0.07974818348884583, + 0.07249555736780167, + 0.015045610256493092, + 0.010685594752430916, + 0.010534633882343769, + 0.008010944351553917, + 0.011605177074670792, + 0.00887032225728035, + 0.040652479976415634, + 0.049110449850559235, + 0.0104373749345541, + 0.021750973537564278, + 0.002359876176342368, + 0.005694211460649967, + 0.002948591485619545, + 0.0040781861171126366, + 0.003180821193382144, + 0.002771897241473198, + 0.004105200059711933, + 0.003177464473992586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.023296458646655083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290008574724197 + ], + "opposite_logprob_tokens": [ + -2.2379000186920166 + ], + "logprob_reward": -0.11290008574724197, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290008574724197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031210968270897865, + 0.039544593542814255, + 0.046885546296834946, + 0.059792410582304, + 0.027504805475473404, + 0.026034511625766754, + 0.10473014414310455, + 0.1533728539943695, + 0.05767562985420227, + 0.4532485008239746 + ], + "credit_map": [ + 0.00517409248277545, + 0.004572554491460323, + 0.005576323252171278, + 0.003292217617854476, + 0.0030923979356884956, + 0.0024004799779504538, + 0.00402831519022584, + 0.003074588719755411, + 0.00872093066573143, + 0.006050703581422567, + 0.007999253459274769, + 0.003167984541505575, + 0.0037280339747667313, + 0.004696316551417112, + 0.0022532411385327578, + 0.002928132191300392, + 0.004925666376948357, + 0.010020819492638111, + 0.007581633515655994, + 0.008919896557927132, + 0.006125698797404766, + 0.0030050703790038824, + 0.0032096353825181723, + 0.0030971274245530367, + 0.005906821694225073, + 0.010654288344085217, + 0.014044045470654964, + 0.00864231027662754, + 0.0065636481158435345, + 0.005434256047010422, + 0.003200216917321086, + 0.005346827208995819, + 0.003918093163520098, + 0.005074213724583387, + 0.004265203606337309, + 0.0028694297652691603, + 0.0032046048436313868, + 0.001973554026335478, + 0.002341133076697588, + 0.00385857536457479, + 0.002662060549482703, + 0.00409583980217576, + 0.0038518314249813557, + 0.002461058786138892, + 0.00481083570048213, + 0.002838150830939412, + 0.0022679241374135017, + 0.0030468134209513664, + 0.017724405974149704, + 0.017569072544574738, + 0.015192173421382904, + 0.010168012231588364, + 0.011449256911873817, + 0.009800742380321026, + 0.010313215665519238, + 0.012513271532952785, + 0.014004584401845932, + 0.010218380950391293, + 0.022160902619361877, + 0.006222517229616642, + 0.04464614391326904, + 0.03924532234668732, + 0.008797723799943924, + 0.0080772889778018, + 0.005487519782036543, + 0.006599001120775938, + 0.010011596605181694, + 0.009235390461981297, + 0.005083722528070211, + 0.00522319320589304, + 0.008674779906868935, + 0.007360432762652636, + 0.03245272859930992, + 0.017783932387828827, + 0.01222539134323597, + 0.013077635318040848, + 0.176131471991539, + 0.17436113953590393, + 0.012054442428052425, + 0.01516178622841835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.11290008574724197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a mid-range shot featuring a young Miao girl at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the girl wear an intricate silver crown with butterfly and floral patterns, along with silver chains and flakes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl wearing a dark blue double-breasted blouse and pleated long skirt with embroidery of mountains, rivers, and animals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a weathered dark brown Miao wooden stilt house directly in front of the girl with rolling green mountains and morning mist in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera gently pan sideways with the subject while the girl performs a graceful spin with her arms extended?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038080889731645584 + ], + "opposite_logprob_tokens": [ + -3.288080930709839 + ], + "logprob_reward": -0.038080889731645584, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038080889731645584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09147758781909943, + 0.053479962050914764, + 0.05307907983660698, + 0.08166888356208801, + 0.07596298307180405, + 0.06168945133686066, + 0.19840078055858612, + 0.15205872058868408, + 0.05801810324192047, + 0.17416448891162872 + ], + "credit_map": [ + 0.010894511826336384, + 0.01349075697362423, + 0.01763525977730751, + 0.009088977240025997, + 0.014742964878678322, + 0.005377696827054024, + 0.011927273124456406, + 0.00832014437764883, + 0.008011309430003166, + 0.009909098967909813, + 0.009406541474163532, + 0.005682905670255423, + 0.008631013333797455, + 0.003710703691467643, + 0.003555241273716092, + 0.004573144484311342, + 0.00609578425064683, + 0.010461430996656418, + 0.009040795266628265, + 0.010362531989812851, + 0.005948744248598814, + 0.003919454291462898, + 0.0032281577587127686, + 0.004022175911813974, + 0.009194198064506054, + 0.013114280998706818, + 0.013155651278793812, + 0.006989236921072006, + 0.011261846870183945, + 0.00803370401263237, + 0.008924917317926884, + 0.010995039716362953, + 0.007320488337427378, + 0.009244497865438461, + 0.009978735819458961, + 0.008321587927639484, + 0.012304059229791164, + 0.006682346109300852, + 0.009262369945645332, + 0.012848902493715286, + 0.005476833321154118, + 0.00982687808573246, + 0.009640080854296684, + 0.00909897219389677, + 0.007224156055599451, + 0.005334801971912384, + 0.007501253858208656, + 0.007586466148495674, + 0.02080618217587471, + 0.03948212042450905, + 0.024531839415431023, + 0.021116705611348152, + 0.03897141292691231, + 0.016996070742607117, + 0.018549658358097076, + 0.01794677972793579, + 0.012254668399691582, + 0.008815944194793701, + 0.013111421838402748, + 0.008128587156534195, + 0.046290501952171326, + 0.044460538774728775, + 0.011322079226374626, + 0.007674961816519499, + 0.005239246413111687, + 0.005730828735977411, + 0.007748749107122421, + 0.005762776825577021, + 0.008673162199556828, + 0.009997567161917686, + 0.005955766886472702, + 0.008910003118216991, + 0.013472170569002628, + 0.006465173792093992, + 0.007231457158923149, + 0.0058687347918748856, + 0.06547709554433823, + 0.06377974152565002, + 0.006130976136773825, + 0.005739137530326843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.038080889731645584 + } + ], + "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/0000155/group_10/index_tensor([1531, 1531])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_tensor([1531, 1531])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fd81e7f9ba580af2ea947dcde72b64eb5e7fc32c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_tensor([1531, 1531])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, the video depicts a scene of minimalist perfection, rendered in surreal detail and filmed in high-speed slow motion. A vibrant sprig of mint floats on the surface of crystal-clear water, against a dark, out-of-focus background. Every detail of the mint is visible: the sharp, jagged edges of the emerald green leaves, the delicate veins of the branches, and the subtle drops of dew clinging to the surface. Suddenly, an unseen force strikes the sprig, causing it to tremble subtly. This impact triggers the video's core action: an intricate series of concentric ripples that spread outward with astonishing clarity. In super-slow motion, the camera captures the formation and expansion of each ripple, its crests catching the clean, bright studio light that shimmers like liquid glass and distorts the reflection of the mint beneath. The entire event unfolds with a tranquil, almost meditative quality, focusing the viewer's attention on the pure, elegant physics of the moment, creating an atmosphere of focused, tranquil beauty.", + "seed": 29017, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_1531_seed_29017_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": -0.00987091762945056, + "vlm_logprob": -0.00987091762945056, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.00987091762945056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.005682623479515314 + ], + "opposite_logprob_tokens": [ + -5.380682468414307 + ], + "logprob_reward": -0.005682623479515314, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005682623479515314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1162056028842926, + 0.13991743326187134, + 0.07208728045225143, + 0.08715119957923889, + 0.1952117383480072, + 0.05668181926012039, + 0.03830965235829353, + 0.07217387855052948, + 0.20150993764400482, + 0.02075142227113247 + ], + "credit_map": [ + 0.012360160239040852, + 0.019544286653399467, + 0.011312827467918396, + 0.01575039140880108, + 0.01449287123978138, + 0.013129761442542076, + 0.016423819586634636, + 0.013191483914852142, + 0.03739247843623161, + 0.010617010295391083, + 0.01892365701496601, + 0.014574804343283176, + 0.009969010017812252, + 0.01574419066309929, + 0.012713054195046425, + 0.019983232021331787, + 0.013325097039341927, + 0.011698182672262192, + 0.00784816313534975, + 0.006498322822153568, + 0.00907996017485857, + 0.010558709502220154, + 0.006559036206454039, + 0.00651980796828866, + 0.012989119626581669, + 0.009697004221379757, + 0.011826110072433949, + 0.005851096473634243, + 0.014147858135402203, + 0.013706054538488388, + 0.007134172134101391, + 0.011799789033830166, + 0.04940158501267433, + 0.00774395139887929, + 0.08289388567209244, + 0.019975095987319946, + 0.004819733090698719, + 0.010557519271969795, + 0.005753401201218367, + 0.014066566713154316, + 0.0073448182083666325, + 0.007441115099936724, + 0.008227069862186909, + 0.004761767573654652, + 0.009662861935794353, + 0.005612251348793507, + 0.00776853458955884, + 0.005863400641828775, + 0.005434202961623669, + 0.0055557312443852425, + 0.004163334146142006, + 0.0026262798346579075, + 0.006679598707705736, + 0.004421020857989788, + 0.004265040159225464, + 0.005164442118257284, + 0.007240142207592726, + 0.010118323378264904, + 0.005542648956179619, + 0.006125729065388441, + 0.013980250805616379, + 0.0061034392565488815, + 0.014159937389194965, + 0.00890340842306614, + 0.023737628012895584, + 0.011637227609753609, + 0.05278706178069115, + 0.020182469859719276, + 0.020894106477499008, + 0.017656980082392693, + 0.016862761229276657, + 0.0377516895532608, + 0.0024996919091790915, + 0.002023964188992977, + 0.002775243716314435, + 0.0022284958977252245, + 0.003006458980962634, + 0.0023504041600972414, + 0.0029024274554103613, + 0.0029647364281117916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.005682623479515314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.006749569438397884 + ], + "opposite_logprob_tokens": [ + -5.006749629974365 + ], + "logprob_reward": -0.006749569438397884, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006749569438397884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0541512668132782, + 0.2848089039325714, + 0.06317088752985, + 0.10107910633087158, + 0.10707473009824753, + 0.04318801313638687, + 0.0348246805369854, + 0.07731485366821289, + 0.17150592803955078, + 0.06288156658411026 + ], + "credit_map": [ + 0.007844539359211922, + 0.006608032155781984, + 0.008164022117853165, + 0.006772376596927643, + 0.005257769953459501, + 0.005677406210452318, + 0.006151613313704729, + 0.0076755136251449585, + 0.09078019112348557, + 0.02145528793334961, + 0.025216680020093918, + 0.015284090302884579, + 0.009630104526877403, + 0.020566433668136597, + 0.019086893647909164, + 0.08278924971818924, + 0.009788391180336475, + 0.009196316823363304, + 0.008321781642735004, + 0.005678119603544474, + 0.008859670720994473, + 0.0078033278696238995, + 0.006636550184339285, + 0.006886739749461412, + 0.014750663191080093, + 0.012807696126401424, + 0.009331506676971912, + 0.012387976050376892, + 0.018166862428188324, + 0.012372417375445366, + 0.009265041910111904, + 0.011996952816843987, + 0.01898111030459404, + 0.012569124810397625, + 0.029738394543528557, + 0.022496270015835762, + 0.004730426240712404, + 0.004029054660350084, + 0.0069803884252905846, + 0.007549969479441643, + 0.006451867986470461, + 0.005328150931745768, + 0.006051518954336643, + 0.003882582997903228, + 0.006383086089044809, + 0.005655261222273111, + 0.004944328218698502, + 0.004491216503083706, + 0.005629024934023619, + 0.00494494428858161, + 0.003996085841208696, + 0.003122067777439952, + 0.005552250891923904, + 0.0029826934915035963, + 0.004533253610134125, + 0.004064362030476332, + 0.00961537379771471, + 0.008126773871481419, + 0.008026082068681717, + 0.007134819868952036, + 0.012404931709170341, + 0.00881276000291109, + 0.00903402827680111, + 0.014160092920064926, + 0.014839756302535534, + 0.006977713666856289, + 0.03806878253817558, + 0.016760053113102913, + 0.01350017823278904, + 0.014532508328557014, + 0.01407031062990427, + 0.05275664106011391, + 0.007429242599755526, + 0.006580711808055639, + 0.01020556315779686, + 0.00852969754487276, + 0.007725224830210209, + 0.0057048792950809, + 0.009668235667049885, + 0.007038016803562641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.006749569438397884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.011090893298387527 + ], + "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": [ + 0.22005786001682281, + 0.1450425535440445, + 0.14245285093784332, + 0.04977084696292877, + 0.16409951448440552, + 0.0474049337208271, + 0.03883913531899452, + 0.06886136531829834, + 0.07279998809099197, + 0.05067093297839165 + ], + "credit_map": [ + 0.0221334807574749, + 0.0397028774023056, + 0.019662143662571907, + 0.029797129333019257, + 0.02855723537504673, + 0.02260098047554493, + 0.028738519176840782, + 0.028865482658147812, + 0.03786030784249306, + 0.010810901410877705, + 0.012129534035921097, + 0.014082645997405052, + 0.007901524193584919, + 0.01167547982186079, + 0.01075757760554552, + 0.039824582636356354, + 0.025388842448592186, + 0.020566262304782867, + 0.01056811772286892, + 0.011481448076665401, + 0.014478222467005253, + 0.039051905274391174, + 0.008292945101857185, + 0.012625107541680336, + 0.007336985319852829, + 0.0051027932204306126, + 0.00890593696385622, + 0.005803021602332592, + 0.0062355403788387775, + 0.006096275523304939, + 0.004129400011152029, + 0.006160893943160772, + 0.045171163976192474, + 0.011855745688080788, + 0.058913279324769974, + 0.011127541773021221, + 0.006837280932813883, + 0.011643090285360813, + 0.0076550585217773914, + 0.010896354913711548, + 0.006446266546845436, + 0.009066353552043438, + 0.005989627446979284, + 0.003425968810915947, + 0.006195292808115482, + 0.006499515846371651, + 0.005627240054309368, + 0.0041546691209077835, + 0.00699484720826149, + 0.004944461397826672, + 0.00447875028476119, + 0.0027292859740555286, + 0.005934045650064945, + 0.003816658863797784, + 0.004048026632517576, + 0.005893060006201267, + 0.007267434149980545, + 0.009115329012274742, + 0.007764645852148533, + 0.005249509122222662, + 0.012322424910962582, + 0.00503395264968276, + 0.011725042015314102, + 0.010383028537034988, + 0.007970982231199741, + 0.0031543399672955275, + 0.022213365882635117, + 0.005313877481967211, + 0.003924046643078327, + 0.0048204748891294, + 0.0058012455701828, + 0.019601652398705482, + 0.005252724513411522, + 0.0051041957922279835, + 0.007546564098447561, + 0.0045347679406404495, + 0.006580844055861235, + 0.0057638040743768215, + 0.00818522647023201, + 0.007702806033194065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011090893298387527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0760977566242218, + 0.08644026517868042, + 0.10335296392440796, + 0.1387479454278946, + 0.09415221214294434, + 0.04663699492812157, + 0.0378393679857254, + 0.028741518035531044, + 0.36620867252349854, + 0.021782323718070984 + ], + "credit_map": [ + 0.010048985481262207, + 0.009972884319722652, + 0.009514338336884975, + 0.008182737044990063, + 0.009325330145657063, + 0.007762227673083544, + 0.009814796037971973, + 0.011476448737084866, + 0.012601148337125778, + 0.007935518398880959, + 0.011281474493443966, + 0.008974381722509861, + 0.005894200410693884, + 0.008081650361418724, + 0.00814095325767994, + 0.023530932143330574, + 0.019230453297495842, + 0.016567328944802284, + 0.013578253798186779, + 0.009320630691945553, + 0.011300751008093357, + 0.014991667121648788, + 0.008697912096977234, + 0.009665954858064651, + 0.019399166107177734, + 0.010898279026150703, + 0.02093249373137951, + 0.011166864074766636, + 0.018460875377058983, + 0.022713767364621162, + 0.010705862194299698, + 0.02447064407169819, + 0.026351764798164368, + 0.006715057417750359, + 0.02658909000456333, + 0.006559684872627258, + 0.005026443861424923, + 0.010711129754781723, + 0.00518713379278779, + 0.007011904381215572, + 0.007924015633761883, + 0.006396945100277662, + 0.006202271208167076, + 0.0030329066794365644, + 0.007695857435464859, + 0.005197474732995033, + 0.004284807480871677, + 0.0059027159586548805, + 0.006394116207957268, + 0.005666228476911783, + 0.0048574707470834255, + 0.0028163064271211624, + 0.00512319291010499, + 0.003999574575573206, + 0.0034638612996786833, + 0.005518611054867506, + 0.003247555810958147, + 0.0028659102972596884, + 0.005221379920840263, + 0.0025904662907123566, + 0.004013578407466412, + 0.002871596487239003, + 0.004166408907622099, + 0.0037646209821105003, + 0.024628590792417526, + 0.00840961467474699, + 0.12985631823539734, + 0.02590011991560459, + 0.01389519777148962, + 0.014665570110082626, + 0.0255802720785141, + 0.12327298521995544, + 0.0021546955686062574, + 0.0026582905557006598, + 0.002784640993922949, + 0.0025806487537920475, + 0.0029602223075926304, + 0.002346506342291832, + 0.003712996141985059, + 0.0025843221228569746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.005250830203294754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.020580671727657318 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.020580671727657318, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580671727657318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10595998167991638, + 0.09395866841077805, + 0.07280361652374268, + 0.0301149170845747, + 0.3277300298213959, + 0.07288577407598495, + 0.03837158530950546, + 0.03204641118645668, + 0.20796696841716766, + 0.018162082880735397 + ], + "credit_map": [ + 0.012236752547323704, + 0.016471968963742256, + 0.012516622431576252, + 0.012353943660855293, + 0.010453327558934689, + 0.011616351082921028, + 0.014132185839116573, + 0.016178827732801437, + 0.01391084585338831, + 0.008883468806743622, + 0.01693171076476574, + 0.013273644261062145, + 0.006463132333010435, + 0.014887300319969654, + 0.00839017704129219, + 0.01121838390827179, + 0.011921259574592113, + 0.011898133903741837, + 0.009131748229265213, + 0.004602568689733744, + 0.009133853949606419, + 0.01331612840294838, + 0.0053921048529446125, + 0.007407810539007187, + 0.0040788850747048855, + 0.0031470078974962234, + 0.0052529447712004185, + 0.0028414539992809296, + 0.004811813589185476, + 0.0036678118631243706, + 0.002691234229132533, + 0.0036237656604498625, + 0.07604982703924179, + 0.01654738001525402, + 0.13268133997917175, + 0.040847357362508774, + 0.009216438047587872, + 0.017319176346063614, + 0.011341463774442673, + 0.02372702956199646, + 0.011419526301324368, + 0.007304383907467127, + 0.011598559096455574, + 0.004980578552931547, + 0.011992421932518482, + 0.010879426263272762, + 0.005383577197790146, + 0.009327294304966927, + 0.006031570490449667, + 0.003941960632801056, + 0.005939848255366087, + 0.002692761132493615, + 0.004192947875708342, + 0.005427855998277664, + 0.0032561728730797768, + 0.006888463161885738, + 0.0036907598841935396, + 0.0032075715716928244, + 0.0038498977664858103, + 0.002688305452466011, + 0.005053075961768627, + 0.0025641517713665962, + 0.006278967950493097, + 0.004713676404207945, + 0.017518892884254456, + 0.0070772250182926655, + 0.0699501782655716, + 0.01663096807897091, + 0.011816897429525852, + 0.012268836610019207, + 0.01455043163150549, + 0.05815352499485016, + 0.0019192537292838097, + 0.00187356595415622, + 0.002487477846443653, + 0.0019175406778231263, + 0.0028038702439516783, + 0.0021205528173595667, + 0.0025669976603239775, + 0.0024728220887482166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.020580671727657318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, the video depicts a scene of minimalist perfection, rendered in surreal detail and filmed in high-speed slow motion. A vibrant sprig of mint floats on the surface of crystal-clear water, against a dark, out-of-focus background. Every detail of the mint is visible: the sharp, jagged edges of the emerald green leaves, the delicate veins of the branches, and the subtle drops of dew clinging to the surface. Suddenly, an unseen force strikes the sprig, causing it to tremble subtly. This impact triggers the video's core action: an intricate series of concentric ripples that spread outward with astonishing clarity. In super-slow motion, the camera captures the formation and expansion of each ripple, its crests catching the clean, bright studio light that shimmers like liquid glass and distorts the reflection of the mint beneath. The entire event unfolds with a tranquil, almost meditative quality, focusing the viewer's attention on the pure, elegant physics of the moment, creating an atmosphere of focused, tranquil beauty.", + "seed": 29018, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_1531_seed_29018_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": -0.008859568461775779, + "vlm_logprob": -0.008859568461775779, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.008859568461775779, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.0038615912199020386 + ], + "opposite_logprob_tokens": [ + -5.878861427307129 + ], + "logprob_reward": -0.0038615912199020386, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0038615912199020386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046670056879520416, + 0.45730409026145935, + 0.014745447784662247, + 0.02362213283777237, + 0.036956869065761566, + 0.050103697925806046, + 0.17633189260959625, + 0.0950273647904396, + 0.08605260401964188, + 0.013185924850404263 + ], + "credit_map": [ + 0.011311673559248447, + 0.007406207732856274, + 0.005449396558105946, + 0.005044059362262487, + 0.0035153948701918125, + 0.00513759721070528, + 0.004287165589630604, + 0.004518562462180853, + 0.05460088327527046, + 0.03513241186738014, + 0.12620805203914642, + 0.01995483785867691, + 0.015630153939127922, + 0.17211167514324188, + 0.009995820932090282, + 0.023670261725783348, + 0.0018081223824992776, + 0.001639971393160522, + 0.0019544141832739115, + 0.0019729326013475657, + 0.0009763819398358464, + 0.0014423199463635683, + 0.001501109218224883, + 0.0034501960035413504, + 0.006169856525957584, + 0.002998544368892908, + 0.0029445216059684753, + 0.0018532411195337772, + 0.0017745227087289095, + 0.0036668621469289064, + 0.0017894821939989924, + 0.0024251018185168505, + 0.00636080838739872, + 0.0053865802474319935, + 0.005293711554259062, + 0.0032230489887297153, + 0.0035682092420756817, + 0.004443804267793894, + 0.004732981789857149, + 0.003947723191231489, + 0.008126475848257542, + 0.0051785847172141075, + 0.009098557755351067, + 0.0037674594204872847, + 0.0051043229177594185, + 0.008846046403050423, + 0.0055070118978619576, + 0.0044752382673323154, + 0.05384630337357521, + 0.00996884610503912, + 0.007435167208313942, + 0.0046466956846416, + 0.0058456650003790855, + 0.07055570930242538, + 0.008698725141584873, + 0.015334772877395153, + 0.008854960091412067, + 0.006007143761962652, + 0.028183022513985634, + 0.004754415713250637, + 0.006009937729686499, + 0.026878543198108673, + 0.008033237420022488, + 0.006306095514446497, + 0.00857562106102705, + 0.003217387478798628, + 0.033674221485853195, + 0.003903829725459218, + 0.0019761621952056885, + 0.029347071424126625, + 0.0022237279918044806, + 0.003134586848318577, + 0.0020845148246735334, + 0.002107766689732671, + 0.0011418677167966962, + 0.0019475198350846767, + 0.000985134276561439, + 0.001346866600215435, + 0.0018852887442335486, + 0.0016869660466909409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.0038615912199020386 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.005960783455520868 + ], + "opposite_logprob_tokens": [ + -5.130960941314697 + ], + "logprob_reward": -0.005960783455520868, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960783455520868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016299834474921227, + 0.3559196889400482, + 0.013064923696219921, + 0.007411627098917961, + 0.006425454281270504, + 0.007292360998690128, + 0.02727806754410267, + 0.017033107578754425, + 0.06683246046304703, + 0.4824425280094147 + ], + "credit_map": [ + 0.0021826624870300293, + 0.0017989024054259062, + 0.0021372861228883266, + 0.0025831314269453287, + 0.001225392334163189, + 0.0013211622135713696, + 0.0027239827904850245, + 0.0023273152764886618, + 0.04262985289096832, + 0.02760399878025055, + 0.09770767390727997, + 0.015578345395624638, + 0.012213654816150665, + 0.13619264960289001, + 0.006733656395226717, + 0.01725984737277031, + 0.0012040759902447462, + 0.0014732865383848548, + 0.001971985213458538, + 0.002039107959717512, + 0.0008242024341598153, + 0.001700287451967597, + 0.0014726199442520738, + 0.0023793589789420366, + 0.0011032313341274858, + 0.0008142573060467839, + 0.001545220147818327, + 0.0007205220172181726, + 0.0007545152329839766, + 0.00098393764346838, + 0.0006150300032459199, + 0.0008749126573093235, + 0.001036889967508614, + 0.0007765372865833342, + 0.0016330513171851635, + 0.0005516205565072596, + 0.0005014307098463178, + 0.0006794121582061052, + 0.0005513541982509196, + 0.0006951577961444855, + 0.0012361279223114252, + 0.0007817483274266124, + 0.0010206833248957992, + 0.0006927170907147229, + 0.0007661407580599189, + 0.0010916567407548428, + 0.0008373416494578123, + 0.0008659448358230293, + 0.008040149696171284, + 0.0013564887922257185, + 0.001582882134243846, + 0.0007641804404556751, + 0.0015940398443490267, + 0.009681356139481068, + 0.0023041442036628723, + 0.0019548276904970407, + 0.0015653965529054403, + 0.0016205451684072614, + 0.003560231300070882, + 0.0029156499076634645, + 0.001313774730078876, + 0.0026768639218062162, + 0.0018445395398885012, + 0.0015361062251031399, + 0.0064368597231805325, + 0.00260313437320292, + 0.026468709111213684, + 0.0028836033307015896, + 0.0014859993243589997, + 0.02258194424211979, + 0.0018901913426816463, + 0.0024820188991725445, + 0.15945854783058167, + 0.03311700001358986, + 0.11570296436548233, + 0.023749476298689842, + 0.015220286324620247, + 0.07712113112211227, + 0.027257150039076805, + 0.03081596828997135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.005960783455520868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.008645367808640003 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645367808640003, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645367808640003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0541885606944561, + 0.49333271384239197, + 0.011312990449368954, + 0.02073863334953785, + 0.014224261976778507, + 0.013553815893828869, + 0.14432315528392792, + 0.03973333537578583, + 0.08515184372663498, + 0.12344074249267578 + ], + "credit_map": [ + 0.007858579978346825, + 0.006719591561704874, + 0.006990090943872929, + 0.008721327409148216, + 0.004587238654494286, + 0.0048489258624613285, + 0.006784480530768633, + 0.0076783280819654465, + 0.058889783918857574, + 0.03857489302754402, + 0.13459090888500214, + 0.02140137180685997, + 0.017202718183398247, + 0.19027985632419586, + 0.009180830791592598, + 0.02321236953139305, + 0.0011315083829686046, + 0.0012166252126917243, + 0.001489302609115839, + 0.0014733157586306334, + 0.0010092740412801504, + 0.0012016920372843742, + 0.001701640896499157, + 0.002089631510898471, + 0.0027144434861838818, + 0.0026323520578444004, + 0.004491116851568222, + 0.0023931795731186867, + 0.001816989155486226, + 0.0023206823971122503, + 0.0017742359777912498, + 0.0025956332683563232, + 0.001949671539478004, + 0.0015604771906509995, + 0.0036505134776234627, + 0.001175112440250814, + 0.0009734511259011924, + 0.0019157796632498503, + 0.0014212895184755325, + 0.0015779666136950254, + 0.0023216481786221266, + 0.0014897751389071345, + 0.0015612761490046978, + 0.0015475467080250382, + 0.0011325771920382977, + 0.002108400920405984, + 0.001149702351540327, + 0.0022428887896239758, + 0.04996168240904808, + 0.007348326966166496, + 0.0054465122520923615, + 0.0031688690651208162, + 0.0037901727482676506, + 0.06019918993115425, + 0.0050787110812962055, + 0.00932969618588686, + 0.0037754313088953495, + 0.0025827146600931883, + 0.013290375471115112, + 0.0017822495428845286, + 0.0012351969489827752, + 0.012746415100991726, + 0.002510634483769536, + 0.0018103168113157153, + 0.007471492979675531, + 0.004020052962005138, + 0.030245112255215645, + 0.005201707594096661, + 0.002447313629090786, + 0.028089111670851707, + 0.0036314765457063913, + 0.00404557678848505, + 0.039515554904937744, + 0.009576719254255295, + 0.028431903570890427, + 0.007838287390768528, + 0.004861052148044109, + 0.018516480922698975, + 0.007211051415652037, + 0.007489695679396391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.008645367808640003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.0052507114596664906 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.0052507114596664906, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052507114596664906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026823390275239944, + 0.5314723253250122, + 0.02264847606420517, + 0.025751229375600815, + 0.02927355468273163, + 0.026103755459189415, + 0.1056118905544281, + 0.08975256234407425, + 0.1175541803240776, + 0.025008689612150192 + ], + "credit_map": [ + 0.00295747397467494, + 0.0028702588751912117, + 0.0033811370376497507, + 0.003348967060446739, + 0.0023945593275129795, + 0.0030634142458438873, + 0.00406372407451272, + 0.0047438545152544975, + 0.06225398927927017, + 0.04134920984506607, + 0.14915797114372253, + 0.02328549325466156, + 0.018557380884885788, + 0.19825494289398193, + 0.011321953497827053, + 0.02729140967130661, + 0.0029556541703641415, + 0.0031191008165478706, + 0.0038183811120688915, + 0.0025662188418209553, + 0.0014057166408747435, + 0.002847841940820217, + 0.0021703424863517284, + 0.003765220521017909, + 0.004175640642642975, + 0.002982116537168622, + 0.003418851178139448, + 0.003114639548584819, + 0.002800358459353447, + 0.0028165271505713463, + 0.003174674231559038, + 0.003268422791734338, + 0.005638600792735815, + 0.003986384253948927, + 0.005268475506454706, + 0.0030728785786777735, + 0.002259041415527463, + 0.003191237570717931, + 0.0028362455777823925, + 0.0030206902883946896, + 0.00357269961386919, + 0.0032080025412142277, + 0.0043537188321352005, + 0.002405332401394844, + 0.0019947581458836794, + 0.00348991178907454, + 0.003060509217903018, + 0.004018824547529221, + 0.02828109636902809, + 0.00695952819660306, + 0.0055257538333535194, + 0.003160767490044236, + 0.0049368999898433685, + 0.03943382948637009, + 0.00463245902210474, + 0.012681552208960056, + 0.00860856007784605, + 0.0070698256604373455, + 0.024686235934495926, + 0.011924385093152523, + 0.004246882628649473, + 0.021696550771594048, + 0.005751254502683878, + 0.005768873728811741, + 0.011674567125737667, + 0.004431179724633694, + 0.045472875237464905, + 0.0050338031724095345, + 0.003057969268411398, + 0.04033564776182175, + 0.0029991990886628628, + 0.004548938013613224, + 0.005984759423881769, + 0.002405119128525257, + 0.0036299193743616343, + 0.002261035842821002, + 0.0019561536610126495, + 0.0026726401410996914, + 0.003432332770898938, + 0.002666726242750883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.0052507114596664906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04635367542505264, + 0.09825544059276581, + 0.02499307133257389, + 0.12773767113685608, + 0.05667963624000549, + 0.03752783313393593, + 0.35433509945869446, + 0.07130851596593857, + 0.1634613573551178, + 0.01934768445789814 + ], + "credit_map": [ + 0.008938848040997982, + 0.006198920775204897, + 0.0055721113458275795, + 0.00545749394223094, + 0.0031095780432224274, + 0.0053058345802128315, + 0.0050421105697751045, + 0.0067287771962583065, + 0.010656245052814484, + 0.007696976885199547, + 0.026587815955281258, + 0.005118329543620348, + 0.0043114409781992435, + 0.03458040952682495, + 0.0033013627398759127, + 0.006002867128700018, + 0.002892999444156885, + 0.002944555599242449, + 0.0031175895128399134, + 0.0033589894883334637, + 0.0020198591519147158, + 0.002890106989070773, + 0.0028446021024137735, + 0.004924369975924492, + 0.013031438924372196, + 0.01698298379778862, + 0.028890445828437805, + 0.01569538377225399, + 0.010137943550944328, + 0.012626728974282742, + 0.012604412622749805, + 0.017768343910574913, + 0.0113590769469738, + 0.0044270046055316925, + 0.019065305590629578, + 0.004067297093570232, + 0.0042291018180549145, + 0.004172320012003183, + 0.004561333451420069, + 0.0047981999814510345, + 0.005254075396806002, + 0.003509447444230318, + 0.008465449325740337, + 0.0023473012261092663, + 0.004206375684589148, + 0.00596696138381958, + 0.004513096995651722, + 0.003265125211328268, + 0.11372777819633484, + 0.020382985472679138, + 0.014822464436292648, + 0.007932483218610287, + 0.011488666757941246, + 0.14518068730831146, + 0.013431796804070473, + 0.027368206530809402, + 0.0074839419685304165, + 0.004999441094696522, + 0.016992827877402306, + 0.0041688233613967896, + 0.007116261869668961, + 0.016026904806494713, + 0.007864711806178093, + 0.006655607372522354, + 0.015589098446071148, + 0.006955939810723066, + 0.06402593851089478, + 0.007057772483676672, + 0.004393236245959997, + 0.053468722850084305, + 0.005463112611323595, + 0.006507537327706814, + 0.004661570303142071, + 0.0015824784059077501, + 0.002682285150513053, + 0.0015207832911983132, + 0.001454001059755683, + 0.0023412411101162434, + 0.0026105688884854317, + 0.0024947563651949167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.020579388365149498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, the video depicts a scene of minimalist perfection, rendered in surreal detail and filmed in high-speed slow motion. A vibrant sprig of mint floats on the surface of crystal-clear water, against a dark, out-of-focus background. Every detail of the mint is visible: the sharp, jagged edges of the emerald green leaves, the delicate veins of the branches, and the subtle drops of dew clinging to the surface. Suddenly, an unseen force strikes the sprig, causing it to tremble subtly. This impact triggers the video's core action: an intricate series of concentric ripples that spread outward with astonishing clarity. In super-slow motion, the camera captures the formation and expansion of each ripple, its crests catching the clean, bright studio light that shimmers like liquid glass and distorts the reflection of the mint beneath. The entire event unfolds with a tranquil, almost meditative quality, focusing the viewer's attention on the pure, elegant physics of the moment, creating an atmosphere of focused, tranquil beauty.", + "seed": 29019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_1531_seed_29019_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": -0.009105908405035733, + "vlm_logprob": -0.009105908405035733, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.009105908405035733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.004909484181553125 + ], + "opposite_logprob_tokens": [ + -5.754909515380859 + ], + "logprob_reward": -0.004909484181553125, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004909484181553125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2343406230211258, + 0.3282985985279083, + 0.05888119712471962, + 0.08119874447584152, + 0.08925262838602066, + 0.06449359655380249, + 0.06015671417117119, + 0.03096814826130867, + 0.017649812623858452, + 0.034760016947984695 + ], + "credit_map": [ + 0.02751222439110279, + 0.03629814088344574, + 0.045483432710170746, + 0.013316791504621506, + 0.02716091461479664, + 0.04583132266998291, + 0.0242355614900589, + 0.014502215199172497, + 0.07116112112998962, + 0.11037577688694, + 0.016608627513051033, + 0.01586194522678852, + 0.03272387757897377, + 0.03898152709007263, + 0.028086522594094276, + 0.014499199576675892, + 0.01647539995610714, + 0.004952108021825552, + 0.007911499589681625, + 0.004059464670717716, + 0.0042546889744699, + 0.005943419877439737, + 0.006092605646699667, + 0.009192006662487984, + 0.018501417711377144, + 0.010659372434020042, + 0.005269013345241547, + 0.007227183319628239, + 0.009048910811543465, + 0.008477719500660896, + 0.010834503918886185, + 0.011180618777871132, + 0.013162228278815746, + 0.01482138130813837, + 0.008132610470056534, + 0.006738830823451281, + 0.013579586520791054, + 0.01023094728589058, + 0.012323177419602871, + 0.010263865813612938, + 0.008565105497837067, + 0.011514492332935333, + 0.006864636670798063, + 0.006950162351131439, + 0.008936789818108082, + 0.008014260791242123, + 0.006516541354358196, + 0.007131610531359911, + 0.010255441069602966, + 0.010699555277824402, + 0.0065579842776060104, + 0.0061163343489170074, + 0.008020679466426373, + 0.006927062291651964, + 0.004456734284758568, + 0.00712292455136776, + 0.004189274739474058, + 0.003402219619601965, + 0.0039036045782268047, + 0.005138796288520098, + 0.0037092382553964853, + 0.0037393937818706036, + 0.003331094281747937, + 0.0035545260179787874, + 0.002906834241002798, + 0.0031637318897992373, + 0.0014868943253532052, + 0.0017217271961271763, + 0.0021481397561728954, + 0.0028979165945202112, + 0.0016435966826975346, + 0.0016809712396934628, + 0.0040022097527980804, + 0.0046460265293717384, + 0.004159196745604277, + 0.005455113016068935, + 0.004747484344989061, + 0.005425191950052977, + 0.003143381793051958, + 0.003181411186233163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.004909484181553125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.004652152303606272 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004652152303606272, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004652152303606272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23014718294143677, + 0.24022041261196136, + 0.1506456583738327, + 0.1306401491165161, + 0.06791144609451294, + 0.03423163294792175, + 0.05564906448125839, + 0.03541975095868111, + 0.021075496450066566, + 0.03405928984284401 + ], + "credit_map": [ + 0.026683634147047997, + 0.03425214812159538, + 0.04618031531572342, + 0.015780208632349968, + 0.02570708654820919, + 0.04772024229168892, + 0.020874040201306343, + 0.012949489988386631, + 0.05140520632266998, + 0.05993550270795822, + 0.018334239721298218, + 0.011575940065085888, + 0.026430346071720123, + 0.026053767651319504, + 0.026728784665465355, + 0.019756611436605453, + 0.032493334263563156, + 0.016468727961182594, + 0.021615272387862206, + 0.014752731658518314, + 0.009928673505783081, + 0.01293553039431572, + 0.016563046723604202, + 0.02588832937180996, + 0.03599557280540466, + 0.015300674363970757, + 0.00675387866795063, + 0.006477499846369028, + 0.009360786527395248, + 0.040855687111616135, + 0.009168243035674095, + 0.006727788597345352, + 0.01050746813416481, + 0.012652170844376087, + 0.005793597549200058, + 0.004963820334523916, + 0.010224820114672184, + 0.00805574655532837, + 0.008951912634074688, + 0.006761904340237379, + 0.004572927486151457, + 0.0071854800917208195, + 0.0028174410108476877, + 0.0022841247264295816, + 0.005811997689306736, + 0.005103490315377712, + 0.003627976169809699, + 0.0028281938284635544, + 0.006582216359674931, + 0.00522683234885335, + 0.005892961286008358, + 0.00778689980506897, + 0.006900548469275236, + 0.006241749040782452, + 0.0058798035606741905, + 0.01113804616034031, + 0.0056383684277534485, + 0.0048898253589868546, + 0.0032034895848482847, + 0.004226327408105135, + 0.005677113309502602, + 0.004940515384078026, + 0.003399800742045045, + 0.0034443067852407694, + 0.003519752761349082, + 0.003317976137623191, + 0.002035918179899454, + 0.0023689577355980873, + 0.0024845812004059553, + 0.0027212691493332386, + 0.0018050925573334098, + 0.0028219465166330338, + 0.00479108514264226, + 0.005010139662772417, + 0.004824432544410229, + 0.003995385952293873, + 0.0036716421600431204, + 0.005316684488207102, + 0.0035159685648977757, + 0.002933947369456291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.004652152303606272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.006748621817678213 + ], + "opposite_logprob_tokens": [ + -5.006748676300049 + ], + "logprob_reward": -0.006748621817678213, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748621817678213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2532085180282593, + 0.2289784848690033, + 0.1651386171579361, + 0.15505863726139069, + 0.035059522837400436, + 0.027386760339140892, + 0.04216597229242325, + 0.030923545360565186, + 0.0237279050052166, + 0.03835202753543854 + ], + "credit_map": [ + 0.02798076719045639, + 0.037538059055805206, + 0.0640619769692421, + 0.017037099227309227, + 0.03113308548927307, + 0.03700608015060425, + 0.021514179185032845, + 0.016937263309955597, + 0.029082583263516426, + 0.03339933976531029, + 0.02479895018041134, + 0.020229792222380638, + 0.026924647390842438, + 0.03899810463190079, + 0.027338091284036636, + 0.028206976130604744, + 0.0450533889234066, + 0.019281746819615364, + 0.013397078029811382, + 0.012134695425629616, + 0.018530944362282753, + 0.015335270203649998, + 0.015255104750394821, + 0.026150386780500412, + 0.05786193162202835, + 0.009596615098416805, + 0.004680478014051914, + 0.0021391964983195066, + 0.005918842274695635, + 0.06293879449367523, + 0.004250011406838894, + 0.007672769948840141, + 0.006244081072509289, + 0.00548501405864954, + 0.003507859306409955, + 0.0037278439849615097, + 0.003391028381884098, + 0.004223817028105259, + 0.003697114298120141, + 0.004782762844115496, + 0.00401511462405324, + 0.005507690366357565, + 0.0029926090501248837, + 0.0021301552187651396, + 0.004141786601394415, + 0.003144933143630624, + 0.00299749500118196, + 0.002456977963447571, + 0.004819465335458517, + 0.007215706165879965, + 0.003604469820857048, + 0.0043045468628406525, + 0.006554929073899984, + 0.005386797245591879, + 0.0052415067330002785, + 0.0050385501235723495, + 0.004580256529152393, + 0.004030194133520126, + 0.0030883359722793102, + 0.003652263432741165, + 0.004384723957628012, + 0.004054273013025522, + 0.003086426295340061, + 0.004047072026878595, + 0.002984491176903248, + 0.0052834139205515385, + 0.0016792662208899856, + 0.002322737593203783, + 0.003939859569072723, + 0.003993152640759945, + 0.001680436311289668, + 0.001844546408392489, + 0.004045885056257248, + 0.0072984835132956505, + 0.005218800622969866, + 0.005059465300291777, + 0.003515088465064764, + 0.006181912962347269, + 0.004198078997433186, + 0.0028343144804239273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006748621817678213 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.008636976592242718 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636976592242718, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636976592242718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11941500753164291, + 0.1308814138174057, + 0.2760888338088989, + 0.15716449916362762, + 0.037303388118743896, + 0.05710848420858383, + 0.052433572709560394, + 0.019465899094939232, + 0.07778850197792053, + 0.07235033810138702 + ], + "credit_map": [ + 0.012057223357260227, + 0.02106919139623642, + 0.018994376063346863, + 0.007978442125022411, + 0.015324855223298073, + 0.02583165653049946, + 0.01056969165802002, + 0.007589570712298155, + 0.022460948675870895, + 0.04084327816963196, + 0.00856151431798935, + 0.0073252832517027855, + 0.0140194958075881, + 0.01338263414800167, + 0.015129579231142998, + 0.00915868952870369, + 0.07966218143701553, + 0.02885543368756771, + 0.02989191561937332, + 0.015624212101101875, + 0.01740591786801815, + 0.02840099111199379, + 0.019672926515340805, + 0.05657527595758438, + 0.05461585894227028, + 0.012205192819237709, + 0.0061757611110806465, + 0.003259690012782812, + 0.009237865917384624, + 0.05327334627509117, + 0.00917055830359459, + 0.009226227179169655, + 0.006201594602316618, + 0.004796317778527737, + 0.004699391312897205, + 0.004256886430084705, + 0.00343148922547698, + 0.00521403644233942, + 0.0034808130003511906, + 0.005222859792411327, + 0.008906522765755653, + 0.011715947650372982, + 0.005080454517155886, + 0.003628826467320323, + 0.009842172265052795, + 0.007322093937546015, + 0.005458403378725052, + 0.005154064856469631, + 0.006105402484536171, + 0.009605450555682182, + 0.00507362000644207, + 0.005863750819116831, + 0.007878128439188004, + 0.006645469460636377, + 0.005807042587548494, + 0.005454704165458679, + 0.004012132063508034, + 0.00231374753639102, + 0.001848437706939876, + 0.0016361133893951774, + 0.0022263890132308006, + 0.0027839411050081253, + 0.002003601286560297, + 0.002641537692397833, + 0.009745622053742409, + 0.019663646817207336, + 0.004195918329060078, + 0.0043645999394357204, + 0.014717443846166134, + 0.014902164228260517, + 0.007212481461465359, + 0.0029866283293813467, + 0.009793626144528389, + 0.013383629731833935, + 0.009564999490976334, + 0.007356257177889347, + 0.008073512464761734, + 0.011302442289888859, + 0.006695767864584923, + 0.006180104333907366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.008636976592242718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.020582307130098343 + ], + "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": [ + 0.15032878518104553, + 0.1640344262123108, + 0.16892622411251068, + 0.208530992269516, + 0.06075518578290939, + 0.052573587745428085, + 0.04475533589720726, + 0.03700687363743782, + 0.04688923433423042, + 0.06619932502508163 + ], + "credit_map": [ + 0.013873405754566193, + 0.027300871908664703, + 0.019894516095519066, + 0.01004862692207098, + 0.0190435778349638, + 0.034596215933561325, + 0.01586170308291912, + 0.009709861129522324, + 0.01937875524163246, + 0.04001757875084877, + 0.016368497163057327, + 0.012572436593472958, + 0.017762212082743645, + 0.028113795444369316, + 0.014540847390890121, + 0.015280306339263916, + 0.043577857315540314, + 0.018648754805326462, + 0.015429641120135784, + 0.010694197379052639, + 0.011647217907011509, + 0.019788555800914764, + 0.0156540647149086, + 0.033485930413007736, + 0.07798586040735245, + 0.014808623120188713, + 0.006188674829900265, + 0.0034747375175356865, + 0.009088654071092606, + 0.08017038553953171, + 0.006947557907551527, + 0.009866506792604923, + 0.007195890415459871, + 0.010815199464559555, + 0.007558136247098446, + 0.007263250648975372, + 0.007551334798336029, + 0.007306221406906843, + 0.0069980802945792675, + 0.0060670748353004456, + 0.008374150842428207, + 0.01013134978711605, + 0.005646855104714632, + 0.0046398937702178955, + 0.00680188462138176, + 0.006562750320881605, + 0.0038555420469492674, + 0.006561160087585449, + 0.007178858853876591, + 0.005584921222180128, + 0.006029859185218811, + 0.0059271520003676414, + 0.005213344469666481, + 0.006364534143358469, + 0.0046080476604402065, + 0.0038486213888972998, + 0.00655634980648756, + 0.004351634066551924, + 0.0039805518463253975, + 0.0045263515785336494, + 0.00409712502732873, + 0.00527421897277236, + 0.00299751665443182, + 0.005223125219345093, + 0.0066604954190552235, + 0.011071430519223213, + 0.0030164942145347595, + 0.0023648701608181, + 0.007831694558262825, + 0.008847190998494625, + 0.003729145508259535, + 0.0033679131884127855, + 0.009138353168964386, + 0.011875682510435581, + 0.0075186267495155334, + 0.005383514333516359, + 0.009461464360356331, + 0.01156958844512701, + 0.006534289103001356, + 0.004717799834907055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.020582307130098343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in extreme close-up from a slightly elevated angle with a static camera, the video depicts a scene of minimalist perfection, rendered in surreal detail and filmed in high-speed slow motion. A vibrant sprig of mint floats on the surface of crystal-clear water, against a dark, out-of-focus background. Every detail of the mint is visible: the sharp, jagged edges of the emerald green leaves, the delicate veins of the branches, and the subtle drops of dew clinging to the surface. Suddenly, an unseen force strikes the sprig, causing it to tremble subtly. This impact triggers the video's core action: an intricate series of concentric ripples that spread outward with astonishing clarity. In super-slow motion, the camera captures the formation and expansion of each ripple, its crests catching the clean, bright studio light that shimmers like liquid glass and distorts the reflection of the mint beneath. The entire event unfolds with a tranquil, almost meditative quality, focusing the viewer's attention on the pure, elegant physics of the moment, creating an atmosphere of focused, tranquil beauty.", + "seed": 29020, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_10/index_1531_seed_29020_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": -0.008050943817943334, + "vlm_logprob": -0.008050943817943334, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.008050943817943334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.0053130853921175 + ], + "opposite_logprob_tokens": [ + -5.880312919616699 + ], + "logprob_reward": -0.0053130853921175, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0053130853921175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09744476526975632, + 0.04330064356327057, + 0.07222539931535721, + 0.33446183800697327, + 0.08450556546449661, + 0.16355367004871368, + 0.02687431313097477, + 0.060002218931913376, + 0.05612878128886223, + 0.06150282174348831 + ], + "credit_map": [ + 0.019022967666387558, + 0.007513710763305426, + 0.02749725803732872, + 0.008856135420501232, + 0.0076529053039848804, + 0.012514312751591206, + 0.00898767076432705, + 0.005399811547249556, + 0.006842219270765781, + 0.004141970071941614, + 0.009436670690774918, + 0.003574714530259371, + 0.006368149537593126, + 0.0050416309386491776, + 0.004033861216157675, + 0.003861424047499895, + 0.011899830773472786, + 0.006546321790665388, + 0.015432982705533504, + 0.004642057698220015, + 0.007799630519002676, + 0.0094585120677948, + 0.007144861854612827, + 0.0093012023717165, + 0.08490170538425446, + 0.02420222945511341, + 0.09086920320987701, + 0.02500065416097641, + 0.017506277188658714, + 0.034689176827669144, + 0.026873217895627022, + 0.030419372022151947, + 0.014477085322141647, + 0.00892773736268282, + 0.017107225954532623, + 0.007131362799555063, + 0.010947191156446934, + 0.008718308992683887, + 0.009678366594016552, + 0.007518291473388672, + 0.04128052294254303, + 0.012211774475872517, + 0.018067430704832077, + 0.015770265832543373, + 0.009264537133276463, + 0.01604386791586876, + 0.013200365006923676, + 0.03771489858627319, + 0.0040670884773135185, + 0.0031250242609530687, + 0.004479927942156792, + 0.0027678257320076227, + 0.003548959270119667, + 0.003643297590315342, + 0.0030606056097894907, + 0.0021815835498273373, + 0.02082415483891964, + 0.004232603125274181, + 0.016264386475086212, + 0.0030900666024535894, + 0.003747177543118596, + 0.0037082210183143616, + 0.004318964667618275, + 0.003816646058112383, + 0.010847003199160099, + 0.005771160591393709, + 0.008698456920683384, + 0.006178239826112986, + 0.005163016729056835, + 0.008310175500810146, + 0.00545975286513567, + 0.005700972862541676, + 0.011335944756865501, + 0.007276458665728569, + 0.008945534005761147, + 0.0059827351942658424, + 0.006917867809534073, + 0.009325905703008175, + 0.007201910484582186, + 0.0045164660550653934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0053130853921175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.0052717006765306 + ], + "opposite_logprob_tokens": [ + -5.255271911621094 + ], + "logprob_reward": -0.0052717006765306, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052717006765306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12315642833709717, + 0.06957132369279861, + 0.1697043627500534, + 0.032781727612018585, + 0.07603593915700912, + 0.11521308124065399, + 0.03403304144740105, + 0.2711111307144165, + 0.05328623950481415, + 0.05510673671960831 + ], + "credit_map": [ + 0.01917009986937046, + 0.01236957497894764, + 0.01949291303753853, + 0.008955718018114567, + 0.015285412780940533, + 0.02628197707235813, + 0.013345078565180302, + 0.00825566053390503, + 0.011406447738409042, + 0.004911123774945736, + 0.014232391491532326, + 0.005618005059659481, + 0.007641367148607969, + 0.010969216004014015, + 0.0065818107686936855, + 0.008210962638258934, + 0.027638589963316917, + 0.0165732279419899, + 0.035220831632614136, + 0.011842082254588604, + 0.0161022637039423, + 0.026433877646923065, + 0.01394769735634327, + 0.02194579876959324, + 0.00468855956569314, + 0.0020600503776222467, + 0.007222178857773542, + 0.0033090366050601006, + 0.002861968707293272, + 0.004929765593260527, + 0.004251816775649786, + 0.0034583488013595343, + 0.013502407819032669, + 0.008372197858989239, + 0.0184775460511446, + 0.006272586993873119, + 0.00817885436117649, + 0.0073089031502604485, + 0.008697492070496082, + 0.005225950386375189, + 0.0178847536444664, + 0.008416741155087948, + 0.018479859456419945, + 0.012677276507019997, + 0.008202391676604748, + 0.015240805223584175, + 0.012681619264185429, + 0.021629638969898224, + 0.00488974479958415, + 0.004114081617444754, + 0.005342390853911638, + 0.005070861428976059, + 0.0035122851841151714, + 0.0029600539710372686, + 0.004852648358792067, + 0.0032909796573221684, + 0.1033540591597557, + 0.014864182099699974, + 0.08103889971971512, + 0.009890680201351643, + 0.009260163642466068, + 0.015140904113650322, + 0.014829572290182114, + 0.022732680663466454, + 0.008969536051154137, + 0.004736755508929491, + 0.007873890921473503, + 0.00424372311681509, + 0.006042359862476587, + 0.009350504726171494, + 0.006218000780791044, + 0.005851465743035078, + 0.008272443898022175, + 0.004706404637545347, + 0.007797390688210726, + 0.0047515942715108395, + 0.005415585357695818, + 0.008251751773059368, + 0.009159436449408531, + 0.006752127315849066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0052717006765306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.008652222342789173 + ], + "opposite_logprob_tokens": [ + -4.758652210235596 + ], + "logprob_reward": -0.008652222342789173, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652222342789173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09316051006317139, + 0.020895548164844513, + 0.02334887534379959, + 0.2632990777492523, + 0.04237734526395798, + 0.16820752620697021, + 0.041891910135746, + 0.2260279804468155, + 0.0411364808678627, + 0.07965482771396637 + ], + "credit_map": [ + 0.016828259453177452, + 0.009869476780295372, + 0.0209820456802845, + 0.007660527713596821, + 0.011475402861833572, + 0.011691711843013763, + 0.008909178897738457, + 0.005743898916989565, + 0.0025933696888387203, + 0.0014789270935580134, + 0.0035417804028838873, + 0.0017542581772431731, + 0.0025070051196962595, + 0.004211894236505032, + 0.0023150271736085415, + 0.0024932846426963806, + 0.004220770206302404, + 0.0021323778200894594, + 0.004543370567262173, + 0.0019059173064306378, + 0.0026314223650842905, + 0.0033883836586028337, + 0.002455837791785598, + 0.002070796210318804, + 0.06017978861927986, + 0.019615057855844498, + 0.06875710189342499, + 0.02065768837928772, + 0.016721708700060844, + 0.030349966138601303, + 0.0231133084744215, + 0.023904448375105858, + 0.008797840215265751, + 0.004314099904149771, + 0.007751591969281435, + 0.00372677412815392, + 0.0042467075400054455, + 0.004740526434034109, + 0.0037170613650232553, + 0.005082744173705578, + 0.04351593554019928, + 0.012408154085278511, + 0.019452808424830437, + 0.01398293487727642, + 0.0088915154337883, + 0.016078179702162743, + 0.013610703870654106, + 0.04026728868484497, + 0.005984630901366472, + 0.004867429379373789, + 0.006123097147792578, + 0.005517043638974428, + 0.004994553979486227, + 0.004838039632886648, + 0.0055922954343259335, + 0.003974820952862501, + 0.09007825702428818, + 0.01093453448265791, + 0.06943168491125107, + 0.006190729793161154, + 0.005622121039777994, + 0.01273152232170105, + 0.009451719000935555, + 0.021587396040558815, + 0.005985681898891926, + 0.005082096438854933, + 0.00623321533203125, + 0.003445612033829093, + 0.004793859552592039, + 0.005591973662376404, + 0.005562121048569679, + 0.004441921133548021, + 0.012601272203028202, + 0.012149595655500889, + 0.010004252195358276, + 0.009838111698627472, + 0.01225913967937231, + 0.008421458303928375, + 0.009953415021300316, + 0.004427584819495678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.008652222342789173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.0028211576864123344 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028211576864123344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028211576864123344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02457050047814846, + 0.020437199622392654, + 0.11568128317594528, + 0.4851064085960388, + 0.027947379276156425, + 0.03464500233530998, + 0.06456457078456879, + 0.1797865778207779, + 0.03271449729800224, + 0.01454663835465908 + ], + "credit_map": [ + 0.0036507423501461744, + 0.0021014935337007046, + 0.0035629102494567633, + 0.002733479952439666, + 0.0029826981481164694, + 0.003915935289114714, + 0.0035645179450511932, + 0.0020587220788002014, + 0.0029899494256824255, + 0.0025478096213191748, + 0.0034436320420354605, + 0.001803967752493918, + 0.002206810750067234, + 0.00312341982498765, + 0.002194635570049286, + 0.002126975217834115, + 0.021679190918803215, + 0.008211306296288967, + 0.02835002914071083, + 0.008908331394195557, + 0.010176975280046463, + 0.010237626731395721, + 0.015378742478787899, + 0.012739083729684353, + 0.11400381475687027, + 0.040863119065761566, + 0.12386678159236908, + 0.038601282984018326, + 0.03275512531399727, + 0.05127069354057312, + 0.04379125311970711, + 0.0399542897939682, + 0.005538126453757286, + 0.0038512088358402252, + 0.005692548584192991, + 0.0024871397763490677, + 0.0032066842541098595, + 0.0022532811854034662, + 0.0030329525470733643, + 0.0018854369409382343, + 0.00597467040643096, + 0.0027493673842400312, + 0.0058196550235152245, + 0.0030754529871046543, + 0.0024632641579955816, + 0.004722247831523418, + 0.003133899299427867, + 0.006706444546580315, + 0.009260283783078194, + 0.006766878068447113, + 0.01071478333324194, + 0.009657412767410278, + 0.007227080874145031, + 0.006462303455919027, + 0.008798348717391491, + 0.005677476059645414, + 0.07152518630027771, + 0.00881325826048851, + 0.05452904850244522, + 0.004854772239923477, + 0.004030653741210699, + 0.010435781441628933, + 0.007609458640217781, + 0.017988435924053192, + 0.004392897244542837, + 0.005002282559871674, + 0.004999315831810236, + 0.0026569515466690063, + 0.004707415588200092, + 0.0035462374798953533, + 0.0044368174858391285, + 0.002972579561173916, + 0.0026853936724364758, + 0.0014613330131396651, + 0.0021282306406646967, + 0.0011496335500851274, + 0.001549123553559184, + 0.002248910954222083, + 0.0020611989311873913, + 0.0012628131080418825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.0028211576864123344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant sprig of mint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the mint sprig floating on the surface of crystal-clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the leaves of the mint emerald green with visible sharp, jagged edges and delicate veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mint sprig tremble and create an intricate series of concentric ripples spreading outward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static, shot from a slightly elevated angle, in super-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": [ + -0.018196552991867065 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196552991867065, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196552991867065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052103180438280106, + 0.012546987272799015, + 0.04538068547844887, + 0.6122652292251587, + 0.012582038529217243, + 0.05050504580140114, + 0.03164688125252724, + 0.13716216385364532, + 0.024270309135317802, + 0.021537497639656067 + ], + "credit_map": [ + 0.007355564273893833, + 0.005238326266407967, + 0.007528485730290413, + 0.0046267444267869, + 0.0044821868650615215, + 0.011563296429812908, + 0.007606020662933588, + 0.003702564397826791, + 0.0018937840359285474, + 0.001403647824190557, + 0.002301130909472704, + 0.0011757287429645658, + 0.0015348896849900484, + 0.00206936988979578, + 0.0010833012638613582, + 0.0010851363185793161, + 0.006653907243162394, + 0.0037808469496667385, + 0.010243536904454231, + 0.0030364361591637135, + 0.003862585872411728, + 0.007391332648694515, + 0.004073433578014374, + 0.006338608451187611, + 0.14192932844161987, + 0.04818181321024895, + 0.15896819531917572, + 0.04822694882750511, + 0.03963574022054672, + 0.06645961105823517, + 0.05550180375576019, + 0.05336178094148636, + 0.0016344928881153464, + 0.0017350510461255908, + 0.0026445328257977962, + 0.0015572966076433659, + 0.0018607915844768286, + 0.0010723520535975695, + 0.0012639330234378576, + 0.0008135895477607846, + 0.006948855239897966, + 0.00486903078854084, + 0.011038484051823616, + 0.005713596940040588, + 0.00465170806273818, + 0.0070487093180418015, + 0.005010652355849743, + 0.005224012769758701, + 0.004472002852708101, + 0.003798418678343296, + 0.004754926543682814, + 0.004858086351305246, + 0.0037664170376956463, + 0.0028447371441870928, + 0.004544015508145094, + 0.002608278766274452, + 0.055641964077949524, + 0.00642531318590045, + 0.04279138520359993, + 0.0034618491772562265, + 0.0026893322356045246, + 0.00750529021024704, + 0.005173787008970976, + 0.013473253697156906, + 0.004045923240482807, + 0.0028793138917535543, + 0.0034764267038553953, + 0.0017605031607672572, + 0.0029164159204810858, + 0.0036102321464568377, + 0.003074249718338251, + 0.0025072440039366484, + 0.0033507368061691523, + 0.003217191668227315, + 0.0025853600818663836, + 0.0027001930866390467, + 0.0029156797099858522, + 0.002407785737887025, + 0.0029247610364109278, + 0.0014357927720993757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.018196552991867065 + } + ], + "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/0000155/group_11/index_tensor([881, 881])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_11/index_tensor([881, 881])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cfa88af921205bd739546bea925e7323e098742d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_11/index_tensor([881, 881])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, overhead tracking shot follows a massive dragon as it slithers through a chaotic sea of storm clouds, a cacophony of inky black and purplish purple. Its serpentine form is covered in countless shield-sized cyan scales, which reflect a cool, moist metallic sheen in the intermittent twilight. Dazzling flashes of blue-white lightning tear through the sky without warning, each flash momentarily illuminating the dragon's powerful muscular features, its horns carved from ancient jade, and its mane dancing like white flames in the wind. The dragon swirls and unwinds with a powerful yet unhurried demeanor. Its golden eyes, gleaming with ancient wisdom, gaze impassively ahead, its long whiskers swaying with the currents. Its massive form sometimes disappears into the swirling clouds, sometimes suddenly bursting forth, stirring the thick mist, as if performing a solemn ritual. The whole picture is full of mythical grandeur and awe-inspiring natural power. Against the background of cold colors, the cyan color of the dragon's body and the white color of the lightning form a strong visual impact, creating an epic, tension-filled and solemn atmosphere.", + "seed": 24975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_11/index_881_seed_24975_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": -0.12965677808970213, + "vlm_logprob": -0.12965677808970213, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.12965677808970213, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.01388814952224493 + ], + "opposite_logprob_tokens": [ + -4.388888359069824 + ], + "logprob_reward": -0.01388814952224493, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01388814952224493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13339145481586456, + 0.0490664467215538, + 0.06242629140615463, + 0.4066014587879181, + 0.11730294674634933, + 0.0738249123096466, + 0.048584748059511185, + 0.055984899401664734, + 0.022750575095415115, + 0.03006625734269619 + ], + "credit_map": [ + 0.009892755188047886, + 0.013155657798051834, + 0.018596533685922623, + 0.014342344366014004, + 0.024691523984074593, + 0.018506286665797234, + 0.010988830588757992, + 0.02321750484406948, + 0.0035486710257828236, + 0.006827968638390303, + 0.005673625506460667, + 0.009546156041324139, + 0.004772975575178862, + 0.005514740012586117, + 0.005775622092187405, + 0.00740668224170804, + 0.006824391428381205, + 0.007046887651085854, + 0.00805413443595171, + 0.009470470249652863, + 0.00795134250074625, + 0.00655872980132699, + 0.008458404801785946, + 0.008061924017965794, + 0.02073543146252632, + 0.009215922094881535, + 0.1587061583995819, + 0.02664937451481819, + 0.015381723642349243, + 0.021438775584101677, + 0.017405003309249878, + 0.13706904649734497, + 0.015703385695815086, + 0.01320315059274435, + 0.017921363934874535, + 0.010029401630163193, + 0.019059596583247185, + 0.011870825663208961, + 0.014930378645658493, + 0.014584827236831188, + 0.0063827186822891235, + 0.009892101399600506, + 0.007736958097666502, + 0.013183345086872578, + 0.011433187872171402, + 0.007862246595323086, + 0.007320807781070471, + 0.010013533756136894, + 0.0053214021027088165, + 0.006572816986590624, + 0.006751158274710178, + 0.005460523068904877, + 0.006364379543811083, + 0.004346886649727821, + 0.007056009955704212, + 0.0067115668207407, + 0.0057597472332417965, + 0.005488808732479811, + 0.006203097756952047, + 0.011126632802188396, + 0.004918789491057396, + 0.006961729843169451, + 0.0045556980185210705, + 0.01097039319574833, + 0.002061276463791728, + 0.0029760643374174833, + 0.003066478995606303, + 0.003823877777904272, + 0.0025403674226254225, + 0.0019952962175011635, + 0.0027796283829957247, + 0.003507582936435938, + 0.003273797919973731, + 0.003923133946955204, + 0.004347243811935186, + 0.0038278319407254457, + 0.0034723440185189247, + 0.005628600716590881, + 0.002430964959785342, + 0.0031623372342437506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.01388814952224493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.06208301708102226 + ], + "opposite_logprob_tokens": [ + -2.8120830059051514 + ], + "logprob_reward": -0.06208301708102226, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208301708102226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24543407559394836, + 0.047736842185258865, + 0.15057210624217987, + 0.269766628742218, + 0.04546660557389259, + 0.06642071157693863, + 0.036103829741477966, + 0.08326444029808044, + 0.029122835025191307, + 0.026112008839845657 + ], + "credit_map": [ + 0.018446259200572968, + 0.027715427801012993, + 0.04273984953761101, + 0.02882734313607216, + 0.034643903374671936, + 0.03172074630856514, + 0.019342176616191864, + 0.04199834540486336, + 0.0040471237152814865, + 0.009425352327525616, + 0.006756604183465242, + 0.00551246153190732, + 0.004146438557654619, + 0.005871032830327749, + 0.00478463526815176, + 0.00719319237396121, + 0.011973533779382706, + 0.020848825573921204, + 0.01693989709019661, + 0.020758438855409622, + 0.01636420004069805, + 0.014915227890014648, + 0.020190393552184105, + 0.028581563383340836, + 0.015515945851802826, + 0.008498817682266235, + 0.10435424000024796, + 0.016850998625159264, + 0.009512760676443577, + 0.016283171251416206, + 0.012116541154682636, + 0.08663415163755417, + 0.006171440705657005, + 0.004835254978388548, + 0.006430067587643862, + 0.005855857394635677, + 0.00429389625787735, + 0.005171203054487705, + 0.0059370724484324455, + 0.006771810818463564, + 0.005544234998524189, + 0.007979077287018299, + 0.009367428719997406, + 0.009097723290324211, + 0.009099675342440605, + 0.008602047339081764, + 0.006301209796220064, + 0.010429308749735355, + 0.0038419354241341352, + 0.004799315240234137, + 0.006166825070977211, + 0.004497107584029436, + 0.0032963030971586704, + 0.0024787625297904015, + 0.004571099299937487, + 0.0064524817280471325, + 0.008093561977148056, + 0.00851709395647049, + 0.009008019231259823, + 0.016136227175593376, + 0.009777223691344261, + 0.006308493670076132, + 0.009433586150407791, + 0.015990227460861206, + 0.0031396897975355387, + 0.0037948591634631157, + 0.0042600990273058414, + 0.004058346152305603, + 0.004228999838232994, + 0.0034351360518485308, + 0.002474229084327817, + 0.003731474746018648, + 0.0029298088047653437, + 0.0025646451395004988, + 0.0036041957791894674, + 0.002924119820818305, + 0.003189565846696496, + 0.00617833761498332, + 0.0016626709839329123, + 0.0030586642678827047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.06208301708102226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.0549808144569397 + ], + "opposite_logprob_tokens": [ + -2.929980754852295 + ], + "logprob_reward": -0.0549808144569397, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549808144569397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29189467430114746, + 0.0685231164097786, + 0.0581345334649086, + 0.23615318536758423, + 0.04853544011712074, + 0.1263020634651184, + 0.041032105684280396, + 0.06650619208812714, + 0.026896143332123756, + 0.03602258116006851 + ], + "credit_map": [ + 0.021723518148064613, + 0.02551889978349209, + 0.04207763820886612, + 0.02952507510781288, + 0.061587195843458176, + 0.04700947925448418, + 0.02426997385919094, + 0.04018288478255272, + 0.005009600892663002, + 0.010741226375102997, + 0.008269168436527252, + 0.011922342702746391, + 0.006212920881807804, + 0.007501945830881596, + 0.007806936278939247, + 0.011058975011110306, + 0.004482020623981953, + 0.00884169340133667, + 0.006773496512323618, + 0.006670401897281408, + 0.008571592159569263, + 0.008248195052146912, + 0.007124353665858507, + 0.0074227782897651196, + 0.012609031982719898, + 0.012533390894532204, + 0.0770605057477951, + 0.018514173105359077, + 0.011042640544474125, + 0.022450832650065422, + 0.013854692690074444, + 0.06808792054653168, + 0.005141223780810833, + 0.006723007187247276, + 0.005292037967592478, + 0.006438926327973604, + 0.006704109720885754, + 0.00477314367890358, + 0.0059001147747039795, + 0.007562878541648388, + 0.010815064422786236, + 0.015136269852519035, + 0.016428209841251373, + 0.01922871544957161, + 0.018360363319516182, + 0.013569037429988384, + 0.015608939342200756, + 0.01715545915067196, + 0.004298302344977856, + 0.004872999154031277, + 0.006255042739212513, + 0.004765225574374199, + 0.005864419508725405, + 0.0033086510375142097, + 0.005064566154032946, + 0.0066028982400894165, + 0.005793516989797354, + 0.007905704900622368, + 0.00787681806832552, + 0.012119133956730366, + 0.00853799656033516, + 0.0054018511436879635, + 0.006392824463546276, + 0.012478343211114407, + 0.002103277947753668, + 0.0028293372597545385, + 0.004698798526078463, + 0.0033435423392802477, + 0.004182545933872461, + 0.003174379700794816, + 0.002276086248457432, + 0.0042881746776402, + 0.0045389579609036446, + 0.005504524800926447, + 0.004830928053706884, + 0.004293615464121103, + 0.004007768351584673, + 0.005924420431256294, + 0.0024434970691800117, + 0.004478868562728167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0549808144569397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.474274218082428 + ], + "opposite_logprob_tokens": [ + -0.974274218082428 + ], + "logprob_reward": -0.474274218082428, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474274218082428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28165188431739807, + 0.1385769546031952, + 0.03829529508948326, + 0.08917785435914993, + 0.04161541536450386, + 0.038408152759075165, + 0.11627867817878723, + 0.14980171620845795, + 0.0821068212389946, + 0.024087168276309967 + ], + "credit_map": [ + 0.030356308445334435, + 0.031197793781757355, + 0.04889027029275894, + 0.03396570682525635, + 0.024582548066973686, + 0.026977229863405228, + 0.021519936621189117, + 0.06416211277246475, + 0.011813275516033173, + 0.012851794250309467, + 0.016298118978738785, + 0.020907875150442123, + 0.014704410918056965, + 0.014676736667752266, + 0.02120080217719078, + 0.026123959571123123, + 0.0032337470911443233, + 0.0030935918912291527, + 0.0044574495404958725, + 0.004825361538678408, + 0.005713362246751785, + 0.0035124230198562145, + 0.005294312722980976, + 0.008165049366652966, + 0.005286180879920721, + 0.0058008343912661076, + 0.027267813682556152, + 0.006429821718484163, + 0.006415966432541609, + 0.010173122398555279, + 0.005730769131332636, + 0.022073350846767426, + 0.004518152680248022, + 0.005964074283838272, + 0.006816709879785776, + 0.004055705852806568, + 0.005277161952108145, + 0.0045250835828483105, + 0.003786038840189576, + 0.006672493647783995, + 0.003984116483479738, + 0.005313233472406864, + 0.006903548259288073, + 0.006107537541538477, + 0.0035375652369111776, + 0.004407418891787529, + 0.002994915936142206, + 0.0051598213613033295, + 0.01400790922343731, + 0.014164242893457413, + 0.021419942378997803, + 0.014832885935902596, + 0.006471040192991495, + 0.008601716719567776, + 0.012215680442750454, + 0.024565264582633972, + 0.012721927836537361, + 0.01589074544608593, + 0.023710278794169426, + 0.029843196272850037, + 0.015620503574609756, + 0.013304749503731728, + 0.013887131586670876, + 0.024823196232318878, + 0.007588325999677181, + 0.01176262367516756, + 0.011248820461332798, + 0.01207641325891018, + 0.01349268201738596, + 0.007515058387070894, + 0.006506115663796663, + 0.01191679947078228, + 0.001457857433706522, + 0.0012131128460168839, + 0.0039049030747264624, + 0.0017723157070577145, + 0.00277105625718832, + 0.008780132979154587, + 0.0013124140677973628, + 0.0028753785882145166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.474274218082428 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.04305769130587578 + ], + "opposite_logprob_tokens": [ + -3.168057680130005 + ], + "logprob_reward": -0.04305769130587578, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305769130587578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21883873641490936, + 0.10424678772687912, + 0.11868873983621597, + 0.059496499598026276, + 0.029660001397132874, + 0.07540886849164963, + 0.05463960021734238, + 0.19885341823101044, + 0.09246925264596939, + 0.04769809916615486 + ], + "credit_map": [ + 0.02139328420162201, + 0.026852024719119072, + 0.03346017375588417, + 0.024047210812568665, + 0.022439872846007347, + 0.024534443393349648, + 0.017911406233906746, + 0.048200320452451706, + 0.006826114375144243, + 0.013341580517590046, + 0.012619096785783768, + 0.019290732219815254, + 0.01083964854478836, + 0.010102174244821072, + 0.01637478545308113, + 0.014852664433419704, + 0.010759789496660233, + 0.010032381862401962, + 0.013244658708572388, + 0.012700020335614681, + 0.01446215808391571, + 0.011816217564046383, + 0.018147077411413193, + 0.027526430785655975, + 0.004407661966979504, + 0.0050270711071789265, + 0.015382598154246807, + 0.005362454801797867, + 0.006118388380855322, + 0.007091250736266375, + 0.006435585208237171, + 0.009671487845480442, + 0.0035738754086196423, + 0.004305106122046709, + 0.004872661549597979, + 0.0025888513773679733, + 0.0035628986079245806, + 0.0037459263112396, + 0.0031114930752664804, + 0.003899186383932829, + 0.0082538528367877, + 0.012372289784252644, + 0.015160318464040756, + 0.011052710004150867, + 0.006965347565710545, + 0.007955183275043964, + 0.005721875466406345, + 0.007927287369966507, + 0.0065526170656085014, + 0.00759923504665494, + 0.011350422166287899, + 0.006813465617597103, + 0.0043188901618123055, + 0.0047195362858474255, + 0.005219205282628536, + 0.008066229522228241, + 0.016889849677681923, + 0.02648308500647545, + 0.019826630130410194, + 0.023350214585661888, + 0.03133575618267059, + 0.02975456602871418, + 0.024599507451057434, + 0.02661382220685482, + 0.008199733681976795, + 0.010082103312015533, + 0.012555200606584549, + 0.01796782575547695, + 0.009936091490089893, + 0.009527306072413921, + 0.011157863773405552, + 0.01304312702268362, + 0.004077570978552103, + 0.004671962931752205, + 0.0074849724769592285, + 0.004726131912320852, + 0.0052339863032102585, + 0.012696154415607452, + 0.0031670466996729374, + 0.005640273913741112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.04305769130587578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, overhead tracking shot follows a massive dragon as it slithers through a chaotic sea of storm clouds, a cacophony of inky black and purplish purple. Its serpentine form is covered in countless shield-sized cyan scales, which reflect a cool, moist metallic sheen in the intermittent twilight. Dazzling flashes of blue-white lightning tear through the sky without warning, each flash momentarily illuminating the dragon's powerful muscular features, its horns carved from ancient jade, and its mane dancing like white flames in the wind. The dragon swirls and unwinds with a powerful yet unhurried demeanor. Its golden eyes, gleaming with ancient wisdom, gaze impassively ahead, its long whiskers swaying with the currents. Its massive form sometimes disappears into the swirling clouds, sometimes suddenly bursting forth, stirring the thick mist, as if performing a solemn ritual. The whole picture is full of mythical grandeur and awe-inspiring natural power. Against the background of cold colors, the cyan color of the dragon's body and the white color of the lightning form a strong visual impact, creating an epic, tension-filled and solemn atmosphere.", + "seed": 24976, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_11/index_881_seed_24976_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": -0.024512211792171, + "vlm_logprob": -0.024512211792171, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.024512211792171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.010120504535734653 + ], + "opposite_logprob_tokens": [ + -4.885120391845703 + ], + "logprob_reward": -0.010120504535734653, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010120504535734653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1086658388376236, + 0.04654074087738991, + 0.2432665079832077, + 0.09078042954206467, + 0.15842632949352264, + 0.09273423254489899, + 0.029045345261693, + 0.03897686302661896, + 0.10919263958930969, + 0.08237112313508987 + ], + "credit_map": [ + 0.01026266161352396, + 0.017978888005018234, + 0.014164170250296593, + 0.012274011969566345, + 0.013376890681684017, + 0.012548735365271568, + 0.01221124455332756, + 0.015849240124225616, + 0.003978244494646788, + 0.0059925224632024765, + 0.0067330701276659966, + 0.007699405774474144, + 0.0034469279926270247, + 0.005182444583624601, + 0.00500065041705966, + 0.008507476188242435, + 0.013989700004458427, + 0.008924407884478569, + 0.0920039564371109, + 0.05933761224150658, + 0.006904302164912224, + 0.01735503226518631, + 0.023875322192907333, + 0.020876191556453705, + 0.009141059592366219, + 0.004761173855513334, + 0.01876950077712536, + 0.009240693412721157, + 0.004428392741829157, + 0.009250742383301258, + 0.0052764154970645905, + 0.029912451282143593, + 0.015461398288607597, + 0.02978699468076229, + 0.01799672655761242, + 0.014660638757050037, + 0.02394798770546913, + 0.020638111978769302, + 0.021755099296569824, + 0.014179367572069168, + 0.012714989483356476, + 0.010933213867247105, + 0.016082845628261566, + 0.009250317700207233, + 0.007224099710583687, + 0.01752367429435253, + 0.009447426535189152, + 0.009557666257023811, + 0.0025059268809854984, + 0.0021726444829255342, + 0.003930128179490566, + 0.0028427380602806807, + 0.002938356250524521, + 0.006185536272823811, + 0.004064198117703199, + 0.004405816551297903, + 0.005144948605448008, + 0.002498860936611891, + 0.005827682092785835, + 0.007682241965085268, + 0.0029153628274798393, + 0.006173588801175356, + 0.003189411014318466, + 0.005544766318053007, + 0.016886377707123756, + 0.011541065759956837, + 0.010962662287056446, + 0.012588098645210266, + 0.014925499446690083, + 0.01177485752850771, + 0.01733972690999508, + 0.01317435223609209, + 0.009506477043032646, + 0.013416861183941364, + 0.015088210813701153, + 0.010172910057008266, + 0.01040156651288271, + 0.008035543374717236, + 0.008563991636037827, + 0.007185567636042833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.010120504535734653 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.02981741353869438 + ], + "opposite_logprob_tokens": [ + -3.5298173427581787 + ], + "logprob_reward": -0.02981741353869438, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981741353869438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13172176480293274, + 0.08437324315309525, + 0.17296947538852692, + 0.15169845521450043, + 0.05981890857219696, + 0.12093313038349152, + 0.029185425490140915, + 0.039516132324934006, + 0.11517634242773056, + 0.09460712969303131 + ], + "credit_map": [ + 0.009343447163701057, + 0.01745627075433731, + 0.02165573090314865, + 0.020983662456274033, + 0.012054599821567535, + 0.011305165477097034, + 0.018144510686397552, + 0.020778365433216095, + 0.007591163273900747, + 0.010096556507050991, + 0.011666763573884964, + 0.01256704144179821, + 0.006840054411441088, + 0.013004046864807606, + 0.008063576184213161, + 0.01454402320086956, + 0.013042417354881763, + 0.008895164355635643, + 0.054662495851516724, + 0.039762090891599655, + 0.005777339916676283, + 0.014577580615878105, + 0.02549237571656704, + 0.010759979486465454, + 0.016707004979252815, + 0.018765131011605263, + 0.015502383932471275, + 0.01949799805879593, + 0.01658996008336544, + 0.02313116565346718, + 0.010943394154310226, + 0.030561411753296852, + 0.00570506788790226, + 0.010896366089582443, + 0.007512395735830069, + 0.006066183093935251, + 0.008549787104129791, + 0.009152819402515888, + 0.006594175472855568, + 0.005342109128832817, + 0.014652557671070099, + 0.017891181632876396, + 0.013611583039164543, + 0.014335783198475838, + 0.011117130517959595, + 0.02362041175365448, + 0.0126792062073946, + 0.013025260530412197, + 0.002272095065563917, + 0.002634545089676976, + 0.0038352091796696186, + 0.0029097474180161953, + 0.003722679801285267, + 0.004691681358963251, + 0.004908236674964428, + 0.004211230203509331, + 0.005727603565901518, + 0.002566311974078417, + 0.0065954760648310184, + 0.007352381944656372, + 0.003730597672984004, + 0.0049786497838795185, + 0.0032863060478121042, + 0.0052788034081459045, + 0.01006287056952715, + 0.013434120453894138, + 0.017652850598096848, + 0.012644301168620586, + 0.01739792712032795, + 0.014339611865580082, + 0.01805299147963524, + 0.011591657064855099, + 0.012830566614866257, + 0.015777425840497017, + 0.02248559333384037, + 0.008113968186080456, + 0.009660561569035053, + 0.008652145974338055, + 0.0087276641279459, + 0.008359194733202457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.02981741353869438 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.02635768987238407 + ], + "opposite_logprob_tokens": [ + -3.651357650756836 + ], + "logprob_reward": -0.02635768987238407, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635768987238407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0921211838722229, + 0.07916567474603653, + 0.14149706065654755, + 0.09832434356212616, + 0.1670040339231491, + 0.18763917684555054, + 0.039890632033348083, + 0.054149266332387924, + 0.07748640328645706, + 0.06272223591804504 + ], + "credit_map": [ + 0.011397887952625751, + 0.011086776852607727, + 0.011935606598854065, + 0.015251731500029564, + 0.008034867234528065, + 0.009344609454274178, + 0.01261663157492876, + 0.012453073635697365, + 0.0064321779645979404, + 0.009108173660933971, + 0.010670102201402187, + 0.013698907569050789, + 0.006269328761845827, + 0.01103547029197216, + 0.009555968455970287, + 0.012395543977618217, + 0.010576671920716763, + 0.008430734276771545, + 0.04752793163061142, + 0.03303312882781029, + 0.005190395284444094, + 0.010587549768388271, + 0.0147300586104393, + 0.011420589871704578, + 0.009203137829899788, + 0.010035596787929535, + 0.013644978404045105, + 0.010365297086536884, + 0.009700777940452099, + 0.01400448102504015, + 0.00759331276640296, + 0.023776765912771225, + 0.016096288338303566, + 0.021861696615815163, + 0.026042213663458824, + 0.019556961953639984, + 0.021741783246397972, + 0.024069391191005707, + 0.025411546230316162, + 0.012224150821566582, + 0.02349948324263096, + 0.021805817261338234, + 0.023745696991682053, + 0.01883447915315628, + 0.03149476647377014, + 0.026042969897389412, + 0.022851191461086273, + 0.01936477981507778, + 0.004287293646484613, + 0.0025628639850765467, + 0.00460638664662838, + 0.0036708328407257795, + 0.0042541190050542355, + 0.008547931909561157, + 0.006175779737532139, + 0.005785422399640083, + 0.005069926381111145, + 0.005325711332261562, + 0.006486593279987574, + 0.008089753799140453, + 0.006060631014406681, + 0.010482151992619038, + 0.004583469592034817, + 0.008051032200455666, + 0.01423004548996687, + 0.008061627857387066, + 0.010814494453370571, + 0.007563542574644089, + 0.009302153252065182, + 0.010046935640275478, + 0.008714191615581512, + 0.008753413334488869, + 0.007765301503241062, + 0.012133238837122917, + 0.014407352544367313, + 0.005895300302654505, + 0.007797510363161564, + 0.004665770567953587, + 0.0046487171202898026, + 0.005409051198512316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.02635768987238407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.038089267909526825 + ], + "opposite_logprob_tokens": [ + -3.2880892753601074 + ], + "logprob_reward": -0.038089267909526825, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089267909526825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05448993295431137, + 0.08529609441757202, + 0.5751512050628662, + 0.1375441551208496, + 0.028642088174819946, + 0.037718046456575394, + 0.018858524039387703, + 0.014005112461745739, + 0.025594698265194893, + 0.022700143977999687 + ], + "credit_map": [ + 0.006272546947002411, + 0.007773215416818857, + 0.007907509803771973, + 0.006417965982109308, + 0.004828380886465311, + 0.005427523981779814, + 0.006711255759000778, + 0.009151529520750046, + 0.007131804712116718, + 0.010225329548120499, + 0.012106307782232761, + 0.010435649193823338, + 0.008877784013748169, + 0.015175925567746162, + 0.008065778762102127, + 0.013277516700327396, + 0.03376735746860504, + 0.019946156069636345, + 0.2214384824037552, + 0.15126928687095642, + 0.016414327546954155, + 0.02974085509777069, + 0.06868504732847214, + 0.03388968110084534, + 0.014217846095561981, + 0.014955905266106129, + 0.018031613901257515, + 0.01665985956788063, + 0.012751711532473564, + 0.019066141918301582, + 0.00984965544193983, + 0.032011423259973526, + 0.0043692742474377155, + 0.0042134542018175125, + 0.003930917475372553, + 0.0029238455463200808, + 0.002838496118783951, + 0.004735845606774092, + 0.003305890830233693, + 0.0023243643809109926, + 0.005995248910039663, + 0.005603447090834379, + 0.004679475910961628, + 0.00387440319173038, + 0.0036551544908434153, + 0.0061178975738584995, + 0.003916638437658548, + 0.003875777591019869, + 0.00195369403809309, + 0.0014518180396407843, + 0.0024853793438524008, + 0.0016445437213405967, + 0.0016817585565149784, + 0.00518963485956192, + 0.0014485878637060523, + 0.003003107849508524, + 0.0014328599208965898, + 0.0011316129239276052, + 0.0020322627387940884, + 0.00240109721198678, + 0.0012212732108309865, + 0.002708876971155405, + 0.0012924085604026914, + 0.0017847202252596617, + 0.004080763552337885, + 0.0028771194629371166, + 0.0034160916693508625, + 0.002176526468247175, + 0.00463431142270565, + 0.0025329000782221556, + 0.0029135800432413816, + 0.002963404403999448, + 0.0024896429385989904, + 0.00310736196115613, + 0.004421711433678865, + 0.002729737898334861, + 0.0030758874490857124, + 0.0022525046952068806, + 0.002212373074144125, + 0.0024109247606247663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.038089267909526825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.018176183104515076 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018176183104515076, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176183104515076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1249442920088768, + 0.030421925708651543, + 0.5174354910850525, + 0.10522854328155518, + 0.04326566308736801, + 0.04465361312031746, + 0.0399697907269001, + 0.03897864744067192, + 0.028871893882751465, + 0.026230165734887123 + ], + "credit_map": [ + 0.013651812449097633, + 0.01918746903538704, + 0.016682898625731468, + 0.016052121296525, + 0.018067406490445137, + 0.01180713064968586, + 0.01176510564982891, + 0.01773034781217575, + 0.0029022754170000553, + 0.004702565260231495, + 0.004499216563999653, + 0.005366052035242319, + 0.0030159023590385914, + 0.0024986399803310633, + 0.0034072583075612783, + 0.0040300157852470875, + 0.028100499883294106, + 0.016153478994965553, + 0.20166857540607452, + 0.1341211199760437, + 0.013934968039393425, + 0.026469865813851357, + 0.0685613751411438, + 0.028425585478544235, + 0.012867799960076809, + 0.008619815111160278, + 0.021591028198599815, + 0.009893704205751419, + 0.006058899685740471, + 0.008980116806924343, + 0.006461316253989935, + 0.03075586073100567, + 0.005072475876659155, + 0.005198057275265455, + 0.006955354008823633, + 0.004506368190050125, + 0.004187438637018204, + 0.008391358889639378, + 0.0053784651681780815, + 0.0035761401522904634, + 0.008533986285328865, + 0.006336017046123743, + 0.004624406341463327, + 0.005953527055680752, + 0.0034219601657241583, + 0.007256283424794674, + 0.003895766334608197, + 0.004631667397916317, + 0.0037158720660954714, + 0.002428723033517599, + 0.006438565440475941, + 0.003922693431377411, + 0.0036963834427297115, + 0.008881689980626106, + 0.005118276458233595, + 0.00576758710667491, + 0.0040137614123523235, + 0.004144743084907532, + 0.00643238564953208, + 0.004893493838608265, + 0.003887208877131343, + 0.006172697525471449, + 0.003728818381205201, + 0.005705540534108877, + 0.004745717626065016, + 0.0034600135404616594, + 0.004059763625264168, + 0.0030831252224743366, + 0.0031443249899894, + 0.0030040356796234846, + 0.0045528654009103775, + 0.002822049194946885, + 0.003319279756397009, + 0.003951185382902622, + 0.004403390921652317, + 0.0036480464041233063, + 0.003638036549091339, + 0.0024446197785437107, + 0.0028203825931996107, + 0.0020052241161465645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.018176183104515076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, overhead tracking shot follows a massive dragon as it slithers through a chaotic sea of storm clouds, a cacophony of inky black and purplish purple. Its serpentine form is covered in countless shield-sized cyan scales, which reflect a cool, moist metallic sheen in the intermittent twilight. Dazzling flashes of blue-white lightning tear through the sky without warning, each flash momentarily illuminating the dragon's powerful muscular features, its horns carved from ancient jade, and its mane dancing like white flames in the wind. The dragon swirls and unwinds with a powerful yet unhurried demeanor. Its golden eyes, gleaming with ancient wisdom, gaze impassively ahead, its long whiskers swaying with the currents. Its massive form sometimes disappears into the swirling clouds, sometimes suddenly bursting forth, stirring the thick mist, as if performing a solemn ritual. The whole picture is full of mythical grandeur and awe-inspiring natural power. Against the background of cold colors, the cyan color of the dragon's body and the white color of the lightning form a strong visual impact, creating an epic, tension-filled and solemn atmosphere.", + "seed": 24977, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_11/index_881_seed_24977_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": -0.33430233635008333, + "vlm_logprob": -0.33430233635008333, + "vlm_yes_no_margin": 2.0, + "vlm_token_credit": -0.33430233635008333, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.027151787653565407 + ], + "opposite_logprob_tokens": [ + -3.6521518230438232 + ], + "logprob_reward": -0.027151787653565407, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027151787653565407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12985435128211975, + 0.11326836794614792, + 0.043274205178022385, + 0.08915633708238602, + 0.0844317153096199, + 0.13609537482261658, + 0.2418196201324463, + 0.02410326525568962, + 0.0437505878508091, + 0.09424615651369095 + ], + "credit_map": [ + 0.007419647183269262, + 0.02696780301630497, + 0.006385372020304203, + 0.006230392027646303, + 0.021489067003130913, + 0.031693823635578156, + 0.011458370834589005, + 0.018209869042038918, + 0.009360916912555695, + 0.012272199615836143, + 0.012531481683254242, + 0.013469678349792957, + 0.014323861338198185, + 0.022258583456277847, + 0.014341644942760468, + 0.01470999326556921, + 0.00397410849109292, + 0.005629694554954767, + 0.007443009410053492, + 0.00586136756464839, + 0.007706411648541689, + 0.003362331772223115, + 0.004614390432834625, + 0.004682890139520168, + 0.007957403548061848, + 0.010549698024988174, + 0.0122213801369071, + 0.012071379460394382, + 0.010674855671823025, + 0.010295812040567398, + 0.01305427122861147, + 0.012331532314419746, + 0.008922534063458443, + 0.012622537091374397, + 0.00980579573661089, + 0.010162441059947014, + 0.007732790894806385, + 0.00841311365365982, + 0.01306995190680027, + 0.01370254810899496, + 0.025554968044161797, + 0.02229376696050167, + 0.017571207135915756, + 0.01723370887339115, + 0.011269262991845608, + 0.015106902457773685, + 0.010896757245063782, + 0.016168802976608276, + 0.026940446346998215, + 0.02740093693137169, + 0.03185324743390083, + 0.031090354546904564, + 0.031555213034152985, + 0.02258865348994732, + 0.03806870058178902, + 0.032322075217962265, + 0.002089656423777342, + 0.0013563932152464986, + 0.007425628136843443, + 0.0017090210458263755, + 0.001102473004721105, + 0.0018869278719648719, + 0.0012147539528086782, + 0.007318411022424698, + 0.004485583398491144, + 0.00465288246050477, + 0.006437451113015413, + 0.005185116548091173, + 0.004315635655075312, + 0.007455604150891304, + 0.005490383133292198, + 0.00572793185710907, + 0.021000096574425697, + 0.004488801117986441, + 0.006145707797259092, + 0.009500152431428432, + 0.006584463641047478, + 0.005691292695701122, + 0.03476288914680481, + 0.006072761956602335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.027151787653565407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.07900331914424896 + ], + "opposite_logprob_tokens": [ + -2.57900333404541 + ], + "logprob_reward": -0.07900331914424896, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900331914424896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32343846559524536, + 0.12752851843833923, + 0.030192455276846886, + 0.07790042459964752, + 0.08861599117517471, + 0.11502271890640259, + 0.050898753106594086, + 0.030909080058336258, + 0.11624173820018768, + 0.03925187885761261 + ], + "credit_map": [ + 0.027721406891942024, + 0.06389787048101425, + 0.019503960385918617, + 0.024717135354876518, + 0.03745574131608009, + 0.06253041326999664, + 0.04102296754717827, + 0.04658897593617439, + 0.012182952836155891, + 0.03030947968363762, + 0.030691398307681084, + 0.014952627010643482, + 0.009744555689394474, + 0.00916812103241682, + 0.010924253612756729, + 0.009555123746395111, + 0.003919726237654686, + 0.005270143039524555, + 0.004728201311081648, + 0.004195879213511944, + 0.0028852219693362713, + 0.0035798747558146715, + 0.0027856891974806786, + 0.0028277200181037188, + 0.008857660926878452, + 0.009592907503247261, + 0.011907492764294147, + 0.010982872918248177, + 0.007644189987331629, + 0.006748449057340622, + 0.011837870813906193, + 0.01032897736877203, + 0.008570551872253418, + 0.010060410015285015, + 0.011479560285806656, + 0.011775000020861626, + 0.009468842297792435, + 0.0104710403829813, + 0.013728626072406769, + 0.013061964884400368, + 0.02389066480100155, + 0.020650595426559448, + 0.015104544349014759, + 0.014028494246304035, + 0.0085794348269701, + 0.010832424275577068, + 0.010225735604763031, + 0.011710821650922298, + 0.0065078046172857285, + 0.006385164335370064, + 0.007656107656657696, + 0.005751663353294134, + 0.00526728481054306, + 0.004954250995069742, + 0.005804827902466059, + 0.008571651764214039, + 0.002787964418530464, + 0.002585610141977668, + 0.008435278199613094, + 0.002362270839512348, + 0.0019569830037653446, + 0.002835607621818781, + 0.001560365897603333, + 0.008385003544390202, + 0.009938797913491726, + 0.011644366197288036, + 0.01503310352563858, + 0.013996049761772156, + 0.013906341046094894, + 0.01995125599205494, + 0.01639402285218239, + 0.015377810224890709, + 0.006333052646368742, + 0.0027394236531108618, + 0.0040117064490914345, + 0.0042741019278764725, + 0.00310897477902472, + 0.005534456577152014, + 0.00978846475481987, + 0.003461697604507208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.07900331914424896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.08906760066747665 + ], + "opposite_logprob_tokens": [ + -2.4640676975250244 + ], + "logprob_reward": -0.08906760066747665, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906760066747665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17357847094535828, + 0.15717090666294098, + 0.02757369540631771, + 0.09791519492864609, + 0.13550052046775818, + 0.06186556816101074, + 0.09091781079769135, + 0.053418342024087906, + 0.07306428998708725, + 0.12899522483348846 + ], + "credit_map": [ + 0.016996808350086212, + 0.03249628096818924, + 0.01544281654059887, + 0.014253812842071056, + 0.036378249526023865, + 0.01997818611562252, + 0.02163800410926342, + 0.01639430969953537, + 0.014000918716192245, + 0.027672547847032547, + 0.02265152893960476, + 0.01820027455687523, + 0.016093973070383072, + 0.024419941008090973, + 0.017637718468904495, + 0.016494005918502808, + 0.0030937758274376392, + 0.004713739268481731, + 0.0035437894985079765, + 0.0036399103701114655, + 0.004361314233392477, + 0.002948951441794634, + 0.0029374100267887115, + 0.0023348049726337194, + 0.009288882836699486, + 0.014161958359181881, + 0.014641646295785904, + 0.014753414317965508, + 0.011328840628266335, + 0.009115857072174549, + 0.012903263792395592, + 0.011721336282789707, + 0.013653945177793503, + 0.0120348259806633, + 0.014794707298278809, + 0.01954253762960434, + 0.016979385167360306, + 0.01582857221364975, + 0.023045815527439117, + 0.019620725885033607, + 0.01103112380951643, + 0.012448898516595364, + 0.008086051791906357, + 0.008306124247610569, + 0.00443435600027442, + 0.005888751707971096, + 0.0042902156710624695, + 0.007380045484751463, + 0.011602239683270454, + 0.007783286273479462, + 0.014129750430583954, + 0.007955058477818966, + 0.010072439908981323, + 0.0077424850314855576, + 0.014900904148817062, + 0.01673164591193199, + 0.004448241088539362, + 0.0030768902506679296, + 0.016101792454719543, + 0.0038250109646469355, + 0.002914851065725088, + 0.004631367512047291, + 0.0026116257067769766, + 0.015808559954166412, + 0.006432086229324341, + 0.008017209358513355, + 0.01147015206515789, + 0.011092325672507286, + 0.006554265972226858, + 0.010175593197345734, + 0.01017199270427227, + 0.009150664322078228, + 0.02469194307923317, + 0.008444149978458881, + 0.010198515839874744, + 0.014132875017821789, + 0.011065172031521797, + 0.012393230572342873, + 0.03801782801747322, + 0.010051509365439415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.08906760066747665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062702655792236 + ], + "opposite_logprob_tokens": [ + -0.2812703251838684 + ], + "logprob_reward": -1.4062702655792236, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062702655792236, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21237485110759735, + 0.12197812646627426, + 0.1962217539548874, + 0.04349219799041748, + 0.06757355481386185, + 0.058005549013614655, + 0.08110801875591278, + 0.06741528958082199, + 0.030956942588090897, + 0.12087372690439224 + ], + "credit_map": [ + 0.017617490142583847, + 0.041969019919633865, + 0.02424512803554535, + 0.016532346606254578, + 0.028290729969739914, + 0.03365249186754227, + 0.02522396109998226, + 0.02484368532896042, + 0.011104822158813477, + 0.0347016267478466, + 0.029909877106547356, + 0.014035796746611595, + 0.009933173656463623, + 0.006879448890686035, + 0.007811333984136581, + 0.0076020448468625546, + 0.02054101787507534, + 0.02684156782925129, + 0.03433432802557945, + 0.022809689864516258, + 0.024269474670290947, + 0.018282249569892883, + 0.0218521598726511, + 0.02729126438498497, + 0.005436996463686228, + 0.0059233191423118114, + 0.006215100176632404, + 0.006399096921086311, + 0.003979250323027372, + 0.0038125705905258656, + 0.006549357436597347, + 0.005176509264856577, + 0.008319571614265442, + 0.0091676851734519, + 0.009301052428781986, + 0.012408958747982979, + 0.005448805168271065, + 0.005551433190703392, + 0.009471516124904156, + 0.007904532365500927, + 0.01323970127850771, + 0.008977890945971012, + 0.007803061977028847, + 0.005773988086730242, + 0.004872032906860113, + 0.005509672686457634, + 0.005657443776726723, + 0.006171754561364651, + 0.01083199679851532, + 0.008857116103172302, + 0.01044644508510828, + 0.00951221864670515, + 0.010234626941382885, + 0.009479700587689877, + 0.011634845286607742, + 0.010111070238053799, + 0.008361510932445526, + 0.004570884630084038, + 0.013578987680375576, + 0.006106119137257338, + 0.0035188491456210613, + 0.006234207656234503, + 0.006314288824796677, + 0.018730437383055687, + 0.0036977401468902826, + 0.005040404852479696, + 0.005046559497714043, + 0.004766085185110569, + 0.002552361460402608, + 0.002930745715275407, + 0.003711463650688529, + 0.0032115825451910496, + 0.02451314777135849, + 0.00755201606079936, + 0.010072797536849976, + 0.016649674624204636, + 0.008138691075146198, + 0.0070359776727855206, + 0.038404688239097595, + 0.00850672647356987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -1.4062702655792236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.0700187087059021 + ], + "opposite_logprob_tokens": [ + -2.695018768310547 + ], + "logprob_reward": -0.0700187087059021, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700187087059021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18361550569534302, + 0.09998080879449844, + 0.14306211471557617, + 0.04157430678606033, + 0.1063799113035202, + 0.06850679218769073, + 0.09391998499631882, + 0.03879556804895401, + 0.03979536518454552, + 0.18436968326568604 + ], + "credit_map": [ + 0.01830282062292099, + 0.03464701399207115, + 0.013643928803503513, + 0.013720648363232613, + 0.026532337069511414, + 0.02907923422753811, + 0.02177993766963482, + 0.025909561663866043, + 0.008556249551475048, + 0.014725925400853157, + 0.018389923498034477, + 0.011425485834479332, + 0.011803814209997654, + 0.015307988971471786, + 0.00965423695743084, + 0.010117176920175552, + 0.013765426352620125, + 0.020144909620285034, + 0.023159347474575043, + 0.01855388656258583, + 0.01891753450036049, + 0.016109200194478035, + 0.016338929533958435, + 0.016072871163487434, + 0.005168064963072538, + 0.005571797024458647, + 0.005420267581939697, + 0.004786335863173008, + 0.00482714781537652, + 0.0037734084762632847, + 0.005990579258650541, + 0.006036701612174511, + 0.01526461262255907, + 0.01241137646138668, + 0.013456347398459911, + 0.01371314749121666, + 0.01075819693505764, + 0.010918112471699715, + 0.017658477649092674, + 0.012199632823467255, + 0.0126288877800107, + 0.009645066224038601, + 0.008238059468567371, + 0.011298894882202148, + 0.005113511346280575, + 0.006966674700379372, + 0.005827500484883785, + 0.008788185194134712, + 0.011617864482104778, + 0.010623249225318432, + 0.013762001879513264, + 0.0111324992030859, + 0.010388532653450966, + 0.009019231423735619, + 0.013449561782181263, + 0.013927038758993149, + 0.004005522932857275, + 0.0029628202319145203, + 0.006637982092797756, + 0.004578622989356518, + 0.002551676006987691, + 0.004114869050681591, + 0.005210234317928553, + 0.0087338387966156, + 0.0063264621421694756, + 0.006203868426382542, + 0.006259050685912371, + 0.004761277232319117, + 0.003076715860515833, + 0.004949573427438736, + 0.0037756392266601324, + 0.004442775622010231, + 0.04225374013185501, + 0.009039188735187054, + 0.011278749443590641, + 0.021119195967912674, + 0.012469270266592503, + 0.007023845333606005, + 0.06901843845844269, + 0.012167233042418957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0700187087059021 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "dynamic, overhead tracking shot follows a massive dragon as it slithers through a chaotic sea of storm clouds, a cacophony of inky black and purplish purple. Its serpentine form is covered in countless shield-sized cyan scales, which reflect a cool, moist metallic sheen in the intermittent twilight. Dazzling flashes of blue-white lightning tear through the sky without warning, each flash momentarily illuminating the dragon's powerful muscular features, its horns carved from ancient jade, and its mane dancing like white flames in the wind. The dragon swirls and unwinds with a powerful yet unhurried demeanor. Its golden eyes, gleaming with ancient wisdom, gaze impassively ahead, its long whiskers swaying with the currents. Its massive form sometimes disappears into the swirling clouds, sometimes suddenly bursting forth, stirring the thick mist, as if performing a solemn ritual. The whole picture is full of mythical grandeur and awe-inspiring natural power. Against the background of cold colors, the cyan color of the dragon's body and the white color of the lightning form a strong visual impact, creating an epic, tension-filled and solemn atmosphere.", + "seed": 24978, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_11/index_881_seed_24978_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": -0.3143813632428646, + "vlm_logprob": -0.3143813632428646, + "vlm_yes_no_margin": 1.975, + "vlm_token_credit": -0.3143813632428646, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.019371140748262405 + ], + "opposite_logprob_tokens": [ + -4.019371032714844 + ], + "logprob_reward": -0.019371140748262405, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019371140748262405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11515171080827713, + 0.04634194076061249, + 0.0770920142531395, + 0.045228999108076096, + 0.0818830206990242, + 0.07722949981689453, + 0.03580937534570694, + 0.1655033975839615, + 0.2859869599342346, + 0.0697731301188469 + ], + "credit_map": [ + 0.010849591344594955, + 0.015627814456820488, + 0.01648194156587124, + 0.009427572600543499, + 0.015720319002866745, + 0.016805505380034447, + 0.013067214749753475, + 0.01717175543308258, + 0.004393274430185556, + 0.005318618845194578, + 0.0058763111010193825, + 0.007766322232782841, + 0.005088827572762966, + 0.0050810533575713634, + 0.003998978063464165, + 0.008818551898002625, + 0.007024907041341066, + 0.00677621690556407, + 0.009178186766803265, + 0.010759183205664158, + 0.00962921418249607, + 0.007061474956572056, + 0.011907960288226604, + 0.01475487183779478, + 0.005228627473115921, + 0.005867636296898127, + 0.006688572932034731, + 0.004818566609174013, + 0.005640486255288124, + 0.007280881516635418, + 0.0049411035142838955, + 0.0047631217166781425, + 0.007045586593449116, + 0.008954266086220741, + 0.008175983093678951, + 0.010922364890575409, + 0.009071513079106808, + 0.018242858350276947, + 0.011079363524913788, + 0.008391088806092739, + 0.006493001244962215, + 0.005032832268625498, + 0.008361407555639744, + 0.007536256220191717, + 0.01566605642437935, + 0.009497634135186672, + 0.010984391905367374, + 0.013657920062541962, + 0.003130802186205983, + 0.003657288383692503, + 0.004236183129251003, + 0.00391507800668478, + 0.004379835445433855, + 0.007319920230656862, + 0.005575092975050211, + 0.003595177549868822, + 0.012286187149584293, + 0.019427860155701637, + 0.018987424671649933, + 0.021764110773801804, + 0.01980871707201004, + 0.03557600826025009, + 0.01987624168395996, + 0.01777682825922966, + 0.026862865313887596, + 0.032927900552749634, + 0.037925686687231064, + 0.05015077441930771, + 0.04078102856874466, + 0.03338129445910454, + 0.03457298502326012, + 0.029384411871433258, + 0.005937748122960329, + 0.008029812946915627, + 0.009417114779353142, + 0.007660564035177231, + 0.010822645388543606, + 0.009952426888048649, + 0.005234526004642248, + 0.01271829754114151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.019371140748262405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.1799088716506958 + ], + "opposite_logprob_tokens": [ + -1.8049088716506958 + ], + "logprob_reward": -0.1799088716506958, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799088716506958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10277655720710754, + 0.07886867225170135, + 0.0962461456656456, + 0.09435322135686874, + 0.07975737005472183, + 0.06493719667196274, + 0.05275510624051094, + 0.16678623855113983, + 0.13199011981487274, + 0.1315292865037918 + ], + "credit_map": [ + 0.01075260154902935, + 0.011444482952356339, + 0.01564396172761917, + 0.009720571339130402, + 0.015048947185277939, + 0.01577669195830822, + 0.01116211898624897, + 0.013227188028395176, + 0.008126316592097282, + 0.008365939371287823, + 0.011002185754477978, + 0.012732905335724354, + 0.010299320332705975, + 0.007916109636425972, + 0.005019397009164095, + 0.015406504273414612, + 0.008782914839684963, + 0.009205101057887077, + 0.010324910283088684, + 0.013224278576672077, + 0.011857007630169392, + 0.010567705146968365, + 0.016777286306023598, + 0.015506959520280361, + 0.011766912415623665, + 0.010143117979168892, + 0.013246672227978706, + 0.00974959321320057, + 0.010616891086101532, + 0.012251490727066994, + 0.015257597900927067, + 0.01132095605134964, + 0.006104812026023865, + 0.00789017416536808, + 0.009515007957816124, + 0.009747877717018127, + 0.009510582312941551, + 0.015672508627176285, + 0.012145068496465683, + 0.009171347133815289, + 0.005337949842214584, + 0.008116612210869789, + 0.007185821421444416, + 0.008904450573027134, + 0.014067022129893303, + 0.006676428485661745, + 0.007926338352262974, + 0.006722582038491964, + 0.005289160180836916, + 0.007932483218610287, + 0.00594328111037612, + 0.0057790097780525684, + 0.006696145515888929, + 0.008947952650487423, + 0.006308114621788263, + 0.005858961492776871, + 0.01784435845911503, + 0.016368715092539787, + 0.02120499685406685, + 0.020458657294511795, + 0.01776685193181038, + 0.03543771430850029, + 0.022308750078082085, + 0.01539621502161026, + 0.013898523524403572, + 0.022725412622094154, + 0.017821433022618294, + 0.017581317573785782, + 0.015954067930579185, + 0.014062397181987762, + 0.017856668680906296, + 0.012090321630239487, + 0.011296546086668968, + 0.015693234279751778, + 0.018281713128089905, + 0.010514816269278526, + 0.02111157216131687, + 0.019102955237030983, + 0.011555830016732216, + 0.023972632363438606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.1799088716506958 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.07900959998369217 + ], + "opposite_logprob_tokens": [ + -2.579009532928467 + ], + "logprob_reward": -0.07900959998369217, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900959998369217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19152678549289703, + 0.033256303519010544, + 0.09436057507991791, + 0.029376551508903503, + 0.11472709476947784, + 0.07637116312980652, + 0.09506324678659439, + 0.21878042817115784, + 0.09572020173072815, + 0.05081764608621597 + ], + "credit_map": [ + 0.014727414585649967, + 0.02233007736504078, + 0.02858635038137436, + 0.01750263199210167, + 0.025528088212013245, + 0.02967155911028385, + 0.019414076581597328, + 0.033766601234674454, + 0.003318474628031254, + 0.004657017067074776, + 0.0044153006747365, + 0.003962540067732334, + 0.0037188187707215548, + 0.00352348736487329, + 0.002928768750280142, + 0.006731895264238119, + 0.009358062408864498, + 0.010075472295284271, + 0.011071004904806614, + 0.013378857634961605, + 0.01220657303929329, + 0.011982453987002373, + 0.01483981404453516, + 0.011448340490460396, + 0.0038133100606501102, + 0.003510572947561741, + 0.0033422717824578285, + 0.003991102334111929, + 0.0031087761744856834, + 0.004739856813102961, + 0.003229551250115037, + 0.003641110844910145, + 0.010509701445698738, + 0.01189421582967043, + 0.012838785536587238, + 0.01248129177838564, + 0.014654514379799366, + 0.021969957277178764, + 0.014812572859227657, + 0.015566052868962288, + 0.00514109106734395, + 0.00755442725494504, + 0.008272325620055199, + 0.01146608218550682, + 0.019081272184848785, + 0.008506684564054012, + 0.008905855007469654, + 0.0074434285052120686, + 0.007901502773165703, + 0.015061641111969948, + 0.011982000432908535, + 0.010883129201829433, + 0.010282501578330994, + 0.015382247045636177, + 0.014644037000834942, + 0.00892618391662836, + 0.026621684432029724, + 0.018369825556874275, + 0.024664724245667458, + 0.025432387366890907, + 0.020458009093999863, + 0.04623515531420708, + 0.01851634494960308, + 0.038482293486595154, + 0.009113873355090618, + 0.013986670412123203, + 0.012243763543665409, + 0.012945745140314102, + 0.013719496317207813, + 0.01174800656735897, + 0.012022688053548336, + 0.00993996299803257, + 0.005005206447094679, + 0.005850410088896751, + 0.007014652714133263, + 0.007580998353660107, + 0.0061199357733130455, + 0.007053720764815807, + 0.005263119470328093, + 0.006929597817361355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.07900959998369217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2235954999923706 + ], + "opposite_logprob_tokens": [ + -0.3485954999923706 + ], + "logprob_reward": -1.2235954999923706, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2235954999923706, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3470490276813507, + 0.08816762268543243, + 0.09452159702777863, + 0.04343622922897339, + 0.040674131363630295, + 0.0922805592417717, + 0.07542549073696136, + 0.05863826349377632, + 0.10402363538742065, + 0.055783502757549286 + ], + "credit_map": [ + 0.02480938844382763, + 0.044915709644556046, + 0.053916431963443756, + 0.030571475625038147, + 0.0500514879822731, + 0.052972253412008286, + 0.03476790338754654, + 0.05504434183239937, + 0.008558507077395916, + 0.011306234635412693, + 0.011912831105291843, + 0.007137726992368698, + 0.01598198525607586, + 0.0084341149777174, + 0.00549300154671073, + 0.019343215972185135, + 0.008525965735316277, + 0.009584223851561546, + 0.011095735244452953, + 0.013217811472713947, + 0.013224322348833084, + 0.009992439299821854, + 0.01673622615635395, + 0.012144860811531544, + 0.004687198903411627, + 0.006462239660322666, + 0.007008484099060297, + 0.003927154932171106, + 0.003942732233554125, + 0.008411338552832603, + 0.003601258620619774, + 0.005395817570388317, + 0.0038459477946162224, + 0.004127472639083862, + 0.005247314926236868, + 0.005492531228810549, + 0.005516250152140856, + 0.005014092195779085, + 0.006472330074757338, + 0.004958188161253929, + 0.011761106550693512, + 0.009589476510882378, + 0.009215720929205418, + 0.00951460376381874, + 0.011027324013411999, + 0.009739321656525135, + 0.013512258417904377, + 0.017920739948749542, + 0.006176200229674578, + 0.01286356057971716, + 0.008919354528188705, + 0.007563761901110411, + 0.01064187940210104, + 0.011463127098977566, + 0.009962652809917927, + 0.007834945805370808, + 0.004557640291750431, + 0.007189755793660879, + 0.006671520881354809, + 0.0083094397559762, + 0.006739887408912182, + 0.011296278797090054, + 0.006668899208307266, + 0.007204839028418064, + 0.01072690635919571, + 0.012060397304594517, + 0.01517645362764597, + 0.014061559922993183, + 0.013763437047600746, + 0.01328214444220066, + 0.009775912389159203, + 0.015176807530224323, + 0.005847535561770201, + 0.006006657145917416, + 0.006309999618679285, + 0.00562514690682292, + 0.007988483645021915, + 0.008188387379050255, + 0.005261726677417755, + 0.010555561631917953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -1.2235954999923706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 dragon with a serpentine form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the dragon's scales cyan, shield-sized, and reflecting a cool metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the dragon have horns carved from ancient jade and a mane resembling white flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a chaotic sea of storm clouds with inky black and purplish purple colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video feature an overhead tracking shot of the dragon moving through the 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": [ + -0.07002170383930206 + ], + "opposite_logprob_tokens": [ + -2.695021629333496 + ], + "logprob_reward": -0.07002170383930206, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002170383930206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17595835030078888, + 0.05521874502301216, + 0.0960923284292221, + 0.05790667608380318, + 0.1290103793144226, + 0.1722181886434555, + 0.12237957119941711, + 0.026154272258281708, + 0.06674938648939133, + 0.09831210225820541 + ], + "credit_map": [ + 0.014514508657157421, + 0.02108066901564598, + 0.02811003290116787, + 0.01666124165058136, + 0.024828627705574036, + 0.02517756260931492, + 0.018390856683254242, + 0.027194872498512268, + 0.005421149078756571, + 0.005652175284922123, + 0.007684985641390085, + 0.008278582245111465, + 0.007078446913510561, + 0.005676792468875647, + 0.003111692378297448, + 0.012314921244978905, + 0.006832631770521402, + 0.009237109683454037, + 0.010486496612429619, + 0.01295668538659811, + 0.012444923631846905, + 0.009760420769453049, + 0.019579168409109116, + 0.014794888906180859, + 0.009232603944838047, + 0.006812917534261942, + 0.0073079015128314495, + 0.005661449860781431, + 0.006573606748133898, + 0.009598868899047375, + 0.005832601338624954, + 0.006886724848300219, + 0.012797132134437561, + 0.014845493249595165, + 0.011052711866796017, + 0.018602006137371063, + 0.015346020460128784, + 0.02572735585272312, + 0.016683388501405716, + 0.013956278562545776, + 0.014905434101819992, + 0.01358035858720541, + 0.01646597310900688, + 0.018692631274461746, + 0.02862168475985527, + 0.02267678640782833, + 0.022487599402666092, + 0.034787703305482864, + 0.010569588281214237, + 0.024730293080210686, + 0.013948526233434677, + 0.013378946110606194, + 0.018303602933883667, + 0.015840517356991768, + 0.013596195727586746, + 0.01201190147548914, + 0.002439473057165742, + 0.0033632458653301, + 0.0028348728083074093, + 0.0032957573421299458, + 0.002947361208498478, + 0.004117598291486502, + 0.003801840590313077, + 0.0033541256561875343, + 0.006178824231028557, + 0.008128885179758072, + 0.0110239889472723, + 0.009956815280020237, + 0.00792454183101654, + 0.007766061928123236, + 0.00839738454669714, + 0.007372882682830095, + 0.007701231632381678, + 0.009796129539608955, + 0.013374699279665947, + 0.011677540838718414, + 0.012310034595429897, + 0.013484838418662548, + 0.009489803574979305, + 0.02047782763838768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.07002170383930206 + } + ], + "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/0000155/group_12/index_tensor([999, 999])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_12/index_tensor([999, 999])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5b0f00026e05af67d45174fb99bede646692e42c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_12/index_tensor([999, 999])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a bizarre car carved from a giant potato as it bumps along a narrow, dirt country road. At the wheel, a man wearing old-fashioned aviator goggles and windswept hair, his face radiating pure joy and excitement. He's dressed in a frayed brown leather jacket and a plaid flannel shirt. The \"potato car\" retains its pristine, bumpy brown surface, even revealing a few newly sprouted buds. The car's body is crudely carved to its basic outline, with four rickety, old-fashioned spoked wheels of varying sizes wedged deep into the potato flesh. With every bump, the car trembles slightly like jelly. The man clings to a steering wheel, also made of a small potato, his body swaying violently, yet he laughs with delight. A light cloud of dust rises from behind. The warm golden afternoon sun fills the frame, lengthening the comical shadows cast by vehicles on the dusty road. Against the backdrop of rolling green hills and azure sky, the whole scene is permeated with a whimsical, joyful atmosphere full of freedom and adventure.", + "seed": 32063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_12/index_999_seed_32063_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": -0.09499657643027604, + "vlm_logprob": -0.09499657643027604, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.09499657643027604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43294432759284973 + ], + "opposite_logprob_tokens": [ + -1.0579442977905273 + ], + "logprob_reward": -0.43294432759284973, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43294432759284973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2338975965976715, + 0.07408462464809418, + 0.0321245938539505, + 0.10170737653970718, + 0.03491009771823883, + 0.1117810532450676, + 0.1445349007844925, + 0.026145193725824356, + 0.1980978399515152, + 0.04271670803427696 + ], + "credit_map": [ + 0.0694810003042221, + 0.06840069591999054, + 0.012944329530000687, + 0.024474218487739563, + 0.01866094022989273, + 0.010640094988048077, + 0.01688847877085209, + 0.01240784116089344, + 0.00813829991966486, + 0.006157549563795328, + 0.008677762933075428, + 0.011102690361440182, + 0.015674451366066933, + 0.005081050097942352, + 0.011820323765277863, + 0.007432497106492519, + 0.004025481641292572, + 0.004980042576789856, + 0.003630956867709756, + 0.004885696806013584, + 0.0035598604008555412, + 0.003854803740978241, + 0.0037881082389503717, + 0.0033996424172073603, + 0.009533483535051346, + 0.011089665815234184, + 0.012162121944129467, + 0.019305909052491188, + 0.012056161649525166, + 0.010882196947932243, + 0.015598435886204243, + 0.011079397983849049, + 0.004694875329732895, + 0.006404871121048927, + 0.00352250668220222, + 0.0049786376766860485, + 0.004642942454665899, + 0.003156599821522832, + 0.003646292258054018, + 0.003863370744511485, + 0.011325448751449585, + 0.01383624505251646, + 0.01475734356790781, + 0.016633065417408943, + 0.015373357571661472, + 0.011245188303291798, + 0.016467180103063583, + 0.012143224477767944, + 0.010859047994017601, + 0.011436089873313904, + 0.019083980470895767, + 0.02347094751894474, + 0.014507906511425972, + 0.014551842585206032, + 0.02480010874569416, + 0.02582496590912342, + 0.003278759540989995, + 0.003666071454063058, + 0.0023687556385993958, + 0.004756297450512648, + 0.0031772684305906296, + 0.0023316156584769487, + 0.003606189740821719, + 0.0029602358117699623, + 0.02692772075533867, + 0.02958686836063862, + 0.021445278078317642, + 0.04297275096178055, + 0.01863824389874935, + 0.018393397331237793, + 0.020753275603055954, + 0.019380297511816025, + 0.006393548101186752, + 0.0048348200507462025, + 0.0037060792092233896, + 0.007537085562944412, + 0.005727998446673155, + 0.006771519314497709, + 0.004860359709709883, + 0.0028852957766503096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.43294432759284973 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006760936230421066 + ], + "opposite_logprob_tokens": [ + -5.006761074066162 + ], + "logprob_reward": -0.006760936230421066, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006760936230421066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15914899110794067, + 0.10216302424669266, + 0.042948462069034576, + 0.04538516327738762, + 0.03344107046723366, + 0.033358559012413025, + 0.17252810299396515, + 0.15918226540088654, + 0.14156708121299744, + 0.11027725040912628 + ], + "credit_map": [ + 0.029843637719750404, + 0.02671900764107704, + 0.012740937992930412, + 0.019108908250927925, + 0.027950888499617577, + 0.01278537604957819, + 0.016436591744422913, + 0.013563642278313637, + 0.012659015133976936, + 0.010822388343513012, + 0.011310464702546597, + 0.01462780311703682, + 0.017720872536301613, + 0.00993167795240879, + 0.014674368314445019, + 0.010416432283818722, + 0.004699802491813898, + 0.00551564572378993, + 0.0063046542927622795, + 0.005069784354418516, + 0.0035489085130393505, + 0.0046134041622281075, + 0.003982407506555319, + 0.009213853627443314, + 0.004046137444674969, + 0.004550086800009012, + 0.006582195404917002, + 0.008862652815878391, + 0.0036389881279319525, + 0.005067942664027214, + 0.006232098676264286, + 0.006405063439160585, + 0.004793227184563875, + 0.003619550261646509, + 0.004442177247256041, + 0.0028005719650536776, + 0.003498974023386836, + 0.004792978521436453, + 0.0032319091260433197, + 0.006261680740863085, + 0.003613556968048215, + 0.004185553174465895, + 0.0026923578698188066, + 0.00571275781840086, + 0.006656351499259472, + 0.003105637850239873, + 0.0039023850113153458, + 0.0034899567253887653, + 0.014794663526117802, + 0.012587206438183784, + 0.019390398636460304, + 0.03753918781876564, + 0.017327163368463516, + 0.013721427880227566, + 0.032747432589530945, + 0.024420641362667084, + 0.02217070944607258, + 0.02308175340294838, + 0.013753438368439674, + 0.026483146473765373, + 0.013700730167329311, + 0.023141415789723396, + 0.02148957923054695, + 0.015361485071480274, + 0.022325269877910614, + 0.021403057500720024, + 0.014729278162121773, + 0.02629408799111843, + 0.012047377415001392, + 0.018215740099549294, + 0.013600124977529049, + 0.012952141463756561, + 0.03461355343461037, + 0.006327335257083178, + 0.00419009942561388, + 0.006309186574071646, + 0.00597786670550704, + 0.039767105132341385, + 0.008328517898917198, + 0.004763585515320301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.006760936230421066 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083937250077724 + ], + "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": [ + 0.15991367399692535, + 0.06010071188211441, + 0.026045244187116623, + 0.05517737567424774, + 0.031597819179296494, + 0.05806457996368408, + 0.3588252067565918, + 0.03789853677153587, + 0.15960335731506348, + 0.052773598581552505 + ], + "credit_map": [ + 0.025887683033943176, + 0.026549912989139557, + 0.01916707307100296, + 0.018026988953351974, + 0.02216395176947117, + 0.012123680673539639, + 0.01828397437930107, + 0.017710383981466293, + 0.00793767161667347, + 0.005258746910840273, + 0.009086531586945057, + 0.010107346810400486, + 0.00730557506904006, + 0.006545106880366802, + 0.006465672980993986, + 0.007394047919660807, + 0.0030284449458122253, + 0.003572737565264106, + 0.0024170337710529566, + 0.004100991878658533, + 0.003045062767341733, + 0.0036577493883669376, + 0.0031398367136716843, + 0.003083382500335574, + 0.005365224089473486, + 0.00616760877892375, + 0.006681170780211687, + 0.01192270778119564, + 0.005714196711778641, + 0.00433807447552681, + 0.007509585004299879, + 0.007478798273950815, + 0.00426925253123045, + 0.0054290080443024635, + 0.0032124656718224287, + 0.004781460855156183, + 0.004276704974472523, + 0.003719814121723175, + 0.0031811753287911415, + 0.002727930899709463, + 0.007610885426402092, + 0.008017626591026783, + 0.0076372441835701466, + 0.008569909259676933, + 0.005248602479696274, + 0.006199241615831852, + 0.0075196209363639355, + 0.007261442951858044, + 0.036376893520355225, + 0.028568506240844727, + 0.03593917563557625, + 0.08466386049985886, + 0.03190832585096359, + 0.023813476786017418, + 0.0730568915605545, + 0.044498030096292496, + 0.004225814249366522, + 0.005155397113412619, + 0.004140021279454231, + 0.005455738864839077, + 0.004932081792503595, + 0.005595226772129536, + 0.0041602361015975475, + 0.004234013147652149, + 0.02507382072508335, + 0.027359966188669205, + 0.015988456085324287, + 0.029310068115592003, + 0.013949149288237095, + 0.017883500084280968, + 0.015675822272896767, + 0.014362544752657413, + 0.012672943994402885, + 0.0036123842000961304, + 0.003184752305969596, + 0.006302986294031143, + 0.004810729529708624, + 0.01404929906129837, + 0.005782980006188154, + 0.0023575176019221544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "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 main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581139251589775 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581139251589775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581139251589775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0989573746919632, + 0.11191368103027344, + 0.041114382445812225, + 0.03784991055727005, + 0.042856745421886444, + 0.19749000668525696, + 0.23187747597694397, + 0.10124996304512024, + 0.11835066974163055, + 0.018339810892939568 + ], + "credit_map": [ + 0.030441155657172203, + 0.03126389533281326, + 0.005096142180263996, + 0.00744623551145196, + 0.0060905651189386845, + 0.0064998953603208065, + 0.005921525880694389, + 0.006197964772582054, + 0.013252084143459797, + 0.01162975374609232, + 0.013337644748389721, + 0.02320536971092224, + 0.011345985345542431, + 0.01466026995331049, + 0.010982750914990902, + 0.013499826192855835, + 0.005521166138350964, + 0.005140944384038448, + 0.00534426886588335, + 0.006277311593294144, + 0.002954747062176466, + 0.004705838393419981, + 0.0037090121768414974, + 0.0074610929004848, + 0.0036369606386870146, + 0.005392628721892834, + 0.003931356128305197, + 0.00948763731867075, + 0.0037831352092325687, + 0.0031355253886431456, + 0.00438473979011178, + 0.004097931087017059, + 0.005912160035222769, + 0.005906916223466396, + 0.004741725046187639, + 0.005945795215666294, + 0.0055781262926757336, + 0.0038743591867387295, + 0.006427357904613018, + 0.004470302257686853, + 0.02149633690714836, + 0.03031863272190094, + 0.015820320695638657, + 0.021749600768089294, + 0.05271574482321739, + 0.01800132729113102, + 0.022648436948657036, + 0.014739617705345154, + 0.024222170934081078, + 0.019329382106661797, + 0.025789503008127213, + 0.02270970679819584, + 0.029598550871014595, + 0.032596901059150696, + 0.02522018551826477, + 0.05241107568144798, + 0.017962303012609482, + 0.019603995606303215, + 0.008843994699418545, + 0.014379692263901234, + 0.006344464607536793, + 0.01686347834765911, + 0.007979920133948326, + 0.009272110648453236, + 0.018655797466635704, + 0.015614724718034267, + 0.012009418569505215, + 0.020302196964621544, + 0.015449983067810535, + 0.013284330256283283, + 0.01422443613409996, + 0.008809777908027172, + 0.0036128549836575985, + 0.0011705734068527818, + 0.0029928209260106087, + 0.0020114032085984945, + 0.0013829735107719898, + 0.0037779384292662144, + 0.0021151539403945208, + 0.0012760920217260718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.020581139251589775 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003612541826441884 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.003612541826441884, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003612541826441884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.299237459897995, + 0.06745660305023193, + 0.035476796329021454, + 0.026937101036310196, + 0.09448637813329697, + 0.06211310997605324, + 0.17972007393836975, + 0.07555817067623138, + 0.07647659629583359, + 0.08253779262304306 + ], + "credit_map": [ + 0.09769945591688156, + 0.0979425311088562, + 0.014991449192166328, + 0.024510936811566353, + 0.01739731803536415, + 0.017581971362233162, + 0.014076687395572662, + 0.015037097036838531, + 0.007618246600031853, + 0.007018727250397205, + 0.008235764689743519, + 0.01594083197414875, + 0.007415685337036848, + 0.00787335354834795, + 0.007661009673029184, + 0.00569298816844821, + 0.0042760116048157215, + 0.003983188420534134, + 0.005218161270022392, + 0.004878027364611626, + 0.003890618681907654, + 0.004015354439616203, + 0.003582014935091138, + 0.005633418448269367, + 0.0028694544453173876, + 0.0029770280234515667, + 0.0029015159234404564, + 0.004846855532377958, + 0.0030326915439218283, + 0.0019134506583213806, + 0.005132712889462709, + 0.003263392485678196, + 0.013895596377551556, + 0.022072460502386093, + 0.010671965777873993, + 0.010404502972960472, + 0.007387896534055471, + 0.013503579422831535, + 0.008029567077755928, + 0.008520807139575481, + 0.006528209894895554, + 0.0058507537469267845, + 0.005883842706680298, + 0.011526310816407204, + 0.009668517857789993, + 0.0068544126115739346, + 0.007303878664970398, + 0.008497188799083233, + 0.019668078050017357, + 0.015678297728300095, + 0.016290370374917984, + 0.04352489486336708, + 0.016102485358715057, + 0.01258684042841196, + 0.036920491605997086, + 0.018948612734675407, + 0.011576312594115734, + 0.011875826865434647, + 0.0062738568522036076, + 0.012070056982338428, + 0.005972417537122965, + 0.014152907766401768, + 0.006293595768511295, + 0.007343189790844917, + 0.011215470731258392, + 0.013535924255847931, + 0.00841133575886488, + 0.010296592488884926, + 0.007973292842507362, + 0.010008322075009346, + 0.007900074124336243, + 0.007135576568543911, + 0.024165987968444824, + 0.004432479850947857, + 0.003869730746373534, + 0.004608931951224804, + 0.005574514158070087, + 0.028912657871842384, + 0.006109854206442833, + 0.004863640293478966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.003612541826441884 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a bizarre car carved from a giant potato as it bumps along a narrow, dirt country road. At the wheel, a man wearing old-fashioned aviator goggles and windswept hair, his face radiating pure joy and excitement. He's dressed in a frayed brown leather jacket and a plaid flannel shirt. The \"potato car\" retains its pristine, bumpy brown surface, even revealing a few newly sprouted buds. The car's body is crudely carved to its basic outline, with four rickety, old-fashioned spoked wheels of varying sizes wedged deep into the potato flesh. With every bump, the car trembles slightly like jelly. The man clings to a steering wheel, also made of a small potato, his body swaying violently, yet he laughs with delight. A light cloud of dust rises from behind. The warm golden afternoon sun fills the frame, lengthening the comical shadows cast by vehicles on the dusty road. Against the backdrop of rolling green hills and azure sky, the whole scene is permeated with a whimsical, joyful atmosphere full of freedom and adventure.", + "seed": 32064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_12/index_999_seed_32064_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": -0.012487516086548568, + "vlm_logprob": -0.012487516086548568, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.012487516086548568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022116679698228836 + ], + "opposite_logprob_tokens": [ + -3.8971166610717773 + ], + "logprob_reward": -0.022116679698228836, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022116679698228836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26008328795433044, + 0.07129886001348495, + 0.0324929878115654, + 0.22094190120697021, + 0.06325788050889969, + 0.06720397621393204, + 0.0569356270134449, + 0.03888232633471489, + 0.05101454257965088, + 0.1378885954618454 + ], + "credit_map": [ + 0.031017307192087173, + 0.022979211062192917, + 0.029783733189105988, + 0.030875978991389275, + 0.03154951333999634, + 0.03633583337068558, + 0.03589663282036781, + 0.04164505377411842, + 0.009762049652636051, + 0.008565113879740238, + 0.007735542953014374, + 0.009093078784644604, + 0.00955833587795496, + 0.00644060131162405, + 0.0112797562032938, + 0.008864377625286579, + 0.004963589366525412, + 0.003732487792149186, + 0.005034414120018482, + 0.00438824575394392, + 0.003109012031927705, + 0.002930109156295657, + 0.003802634309977293, + 0.004532493185251951, + 0.019829915836453438, + 0.02270006388425827, + 0.028020944446325302, + 0.02596384659409523, + 0.028992563486099243, + 0.01920079067349434, + 0.049089860171079636, + 0.02714390866458416, + 0.0061078257858753204, + 0.00711276987567544, + 0.005550145171582699, + 0.008886883035302162, + 0.011485681869089603, + 0.00690515385940671, + 0.01076426263898611, + 0.006445157341659069, + 0.005382372997701168, + 0.00496724434196949, + 0.011616348288953304, + 0.013444101437926292, + 0.005049275234341621, + 0.0064484728500247, + 0.008718910627067089, + 0.011577246710658073, + 0.004491252359002829, + 0.006230156868696213, + 0.005690851714462042, + 0.013968556188046932, + 0.005542919971048832, + 0.00519122090190649, + 0.008779183961451054, + 0.00704148318618536, + 0.004094521980732679, + 0.00401259446516633, + 0.002973695518448949, + 0.0060921525582671165, + 0.005370070226490498, + 0.005370816215872765, + 0.0059002176858484745, + 0.005068258848041296, + 0.003911168780177832, + 0.0059729162603616714, + 0.005784638226032257, + 0.00921635702252388, + 0.00933068711310625, + 0.006899410858750343, + 0.0057642278261482716, + 0.00413513695821166, + 0.011155250482261181, + 0.04719247668981552, + 0.0055117495357990265, + 0.012420319952070713, + 0.007165290415287018, + 0.041128598153591156, + 0.006464563310146332, + 0.006850346457213163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.022116679698228836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798288345336914 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009798288345336914, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798288345336914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19267618656158447, + 0.15801309049129486, + 0.06481098383665085, + 0.10763430595397949, + 0.12032684683799744, + 0.08102061599493027, + 0.1387389451265335, + 0.03408535569906235, + 0.04332476109266281, + 0.05936887860298157 + ], + "credit_map": [ + 0.024046529084444046, + 0.01788141578435898, + 0.022755000740289688, + 0.0294985082000494, + 0.021564897149801254, + 0.019999265670776367, + 0.026871977373957634, + 0.030058590695261955, + 0.018313948065042496, + 0.020577523857355118, + 0.01726389303803444, + 0.02775365300476551, + 0.01972557045519352, + 0.014116878621280193, + 0.02179427072405815, + 0.01846734993159771, + 0.007815616205334663, + 0.007434495259076357, + 0.007849621586501598, + 0.008912799879908562, + 0.008260433562099934, + 0.008147942833602428, + 0.008317079395055771, + 0.008072988130152225, + 0.009319067001342773, + 0.011033543385565281, + 0.010191596113145351, + 0.012310737743973732, + 0.016933681443333626, + 0.009683119133114815, + 0.026874922215938568, + 0.011287638917565346, + 0.010646199807524681, + 0.015226681716740131, + 0.01094796136021614, + 0.011610005050897598, + 0.021804776042699814, + 0.014379170723259449, + 0.022304730489850044, + 0.013407318852841854, + 0.008171409368515015, + 0.012077242136001587, + 0.010303560644388199, + 0.01258406788110733, + 0.008753456175327301, + 0.009539344348013401, + 0.011117692105472088, + 0.008473845198750496, + 0.01333615556359291, + 0.015756534412503242, + 0.01377487275749445, + 0.03543319180607796, + 0.014484226703643799, + 0.012168511748313904, + 0.020898960530757904, + 0.012886491604149342, + 0.0037955360021442175, + 0.003335659857839346, + 0.0041170790791511536, + 0.004820001777261496, + 0.004971734248101711, + 0.004641175270080566, + 0.00454087695106864, + 0.003863292746245861, + 0.003643611678853631, + 0.005152665078639984, + 0.006702024023979902, + 0.006448425818234682, + 0.006544101517647505, + 0.005418207962065935, + 0.005399297922849655, + 0.004016428254544735, + 0.004879893269389868, + 0.006254294421523809, + 0.008836633525788784, + 0.01420869491994381, + 0.00782835390418768, + 0.005741366185247898, + 0.006113568786531687, + 0.00550607405602932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009798288345336914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200001023709774 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200001023709774, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200001023709774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2265404909849167, + 0.14997033774852753, + 0.04272840544581413, + 0.09234218299388885, + 0.057384245097637177, + 0.06652054190635681, + 0.11882589757442474, + 0.04670446366071701, + 0.0661182701587677, + 0.13286519050598145 + ], + "credit_map": [ + 0.03154435753822327, + 0.023993581533432007, + 0.02981056459248066, + 0.02761748433113098, + 0.02696264162659645, + 0.026212641969323158, + 0.031238095834851265, + 0.0291611235588789, + 0.015335453674197197, + 0.014321483671665192, + 0.016522204503417015, + 0.02851221151649952, + 0.019329961389303207, + 0.018472224473953247, + 0.017406901344656944, + 0.020069893449544907, + 0.007955912500619888, + 0.005253025330603123, + 0.006347224116325378, + 0.005596401169896126, + 0.0038810379337519407, + 0.0033213540446013212, + 0.005568336695432663, + 0.004805114585906267, + 0.007725307252258062, + 0.010184716433286667, + 0.010494974441826344, + 0.010049260221421719, + 0.013718456961214542, + 0.007429393474012613, + 0.02381737343966961, + 0.008922708220779896, + 0.005832748953253031, + 0.006274256389588118, + 0.01069546677172184, + 0.008362308144569397, + 0.00773632200434804, + 0.004970871843397617, + 0.0070259179919958115, + 0.0064863525331020355, + 0.00639648362994194, + 0.0065701799467206, + 0.0076310220174491405, + 0.017033370211720467, + 0.005403133574873209, + 0.006662329658865929, + 0.008983299136161804, + 0.007840720936655998, + 0.01411446649581194, + 0.011324673891067505, + 0.01449219137430191, + 0.02855456806719303, + 0.008376238867640495, + 0.014996087178587914, + 0.013582677580416203, + 0.013384995050728321, + 0.004843421280384064, + 0.004965931177139282, + 0.004921563435345888, + 0.008479373529553413, + 0.007372133433818817, + 0.004768416751176119, + 0.00761265866458416, + 0.003740964224562049, + 0.005065681878477335, + 0.006768686696887016, + 0.009466427378356457, + 0.010539944283664227, + 0.009487624280154705, + 0.008146455511450768, + 0.010035114362835884, + 0.00660833390429616, + 0.010486246086657047, + 0.017698079347610474, + 0.01983092539012432, + 0.028619499877095222, + 0.015459798276424408, + 0.01346721313893795, + 0.012470381334424019, + 0.014833061955869198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.014200001023709774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1547488123178482, + 0.09273617714643478, + 0.04966181889176369, + 0.05582893639802933, + 0.12510178983211517, + 0.1258283406496048, + 0.06344757229089737, + 0.12756593525409698, + 0.028975052759051323, + 0.1761055886745453 + ], + "credit_map": [ + 0.022026386111974716, + 0.014458481222391129, + 0.020795980468392372, + 0.02381441928446293, + 0.016437198966741562, + 0.013752643950283527, + 0.019599782302975655, + 0.023863911628723145, + 0.02010703831911087, + 0.010545936413109303, + 0.016759660094976425, + 0.01049448736011982, + 0.00973985344171524, + 0.007313213311135769, + 0.0078103081323206425, + 0.00996567215770483, + 0.008331650868058205, + 0.00663568452000618, + 0.0066572981886565685, + 0.0055724456906318665, + 0.006918526254594326, + 0.0049475799314677715, + 0.004301084671169519, + 0.006297547835856676, + 0.0064215040765702724, + 0.006009973119944334, + 0.007241641171276569, + 0.009153434075415134, + 0.007570484187453985, + 0.004515470936894417, + 0.010477686300873756, + 0.004438740201294422, + 0.013775372877717018, + 0.01760011352598667, + 0.01094980351626873, + 0.014445201493799686, + 0.019073015078902245, + 0.017392730340361595, + 0.02081192284822464, + 0.011053611524403095, + 0.009632013738155365, + 0.012478903867304325, + 0.024545352905988693, + 0.02270716428756714, + 0.010675989091396332, + 0.011972959153354168, + 0.014366242103278637, + 0.019449714571237564, + 0.007426925469189882, + 0.005963534116744995, + 0.00833895243704319, + 0.013070417568087578, + 0.003984847106039524, + 0.0073586259968578815, + 0.0076482780277729034, + 0.009655984118580818, + 0.014128927141427994, + 0.014538208954036236, + 0.015059386380016804, + 0.011576012708246708, + 0.015948215499520302, + 0.02426932565867901, + 0.019737357273697853, + 0.012308496981859207, + 0.0023255597334355116, + 0.003044398967176676, + 0.0051481216214597225, + 0.003804469248279929, + 0.004439866170287132, + 0.0037235254421830177, + 0.0036859912797808647, + 0.0028031188994646072, + 0.0105467289686203, + 0.03731800988316536, + 0.016460813581943512, + 0.03412729129195213, + 0.013713563792407513, + 0.03572713956236839, + 0.014042333699762821, + 0.014169694855809212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.011074034497141838 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248576868325472 + ], + "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": [ + 0.18739406764507294, + 0.08873333781957626, + 0.060341764241456985, + 0.04940907284617424, + 0.12438581138849258, + 0.09517590701580048, + 0.08551852405071259, + 0.07043968886137009, + 0.09356921166181564, + 0.145032599568367 + ], + "credit_map": [ + 0.0392473004758358, + 0.030157316476106644, + 0.0255205649882555, + 0.022080449387431145, + 0.01761273667216301, + 0.0162563007324934, + 0.019610609859228134, + 0.01690879836678505, + 0.02045029029250145, + 0.010483616963028908, + 0.015743233263492584, + 0.012959938496351242, + 0.009502600878477097, + 0.004736488219350576, + 0.00759524293243885, + 0.007261921651661396, + 0.009981852024793625, + 0.008715485222637653, + 0.009357583709061146, + 0.007840870879590511, + 0.005984369665384293, + 0.004957723431289196, + 0.006915749050676823, + 0.0065881311893463135, + 0.007262478116899729, + 0.006065095774829388, + 0.009691775776445866, + 0.006573741789907217, + 0.004242440219968557, + 0.0032634956296533346, + 0.008103721775114536, + 0.004206323530524969, + 0.01873311959207058, + 0.015637392178177834, + 0.015302350744605064, + 0.02070676162838936, + 0.0158334169536829, + 0.011575327254831791, + 0.013978792354464531, + 0.01261864323168993, + 0.01102562714368105, + 0.007613640744239092, + 0.013354053720831871, + 0.015309588983654976, + 0.008874563500285149, + 0.011174391955137253, + 0.01054159551858902, + 0.017282450571656227, + 0.010304218158125877, + 0.007869263179600239, + 0.011467003263533115, + 0.018313221633434296, + 0.0052154031582176685, + 0.011296925134956837, + 0.009176062420010567, + 0.011876426637172699, + 0.007698815781623125, + 0.007798628881573677, + 0.009037374518811703, + 0.008426015265285969, + 0.006680503021925688, + 0.0119822071865201, + 0.008624769747257233, + 0.010191375389695168, + 0.011574935168027878, + 0.010366908274590969, + 0.016316082328557968, + 0.01411735825240612, + 0.009958726353943348, + 0.01292693056166172, + 0.007676216308027506, + 0.010632048361003399, + 0.012194025330245495, + 0.014084743335843086, + 0.022288301959633827, + 0.03096557781100273, + 0.015546437352895737, + 0.015213643200695515, + 0.01466644648462534, + 0.0200734194368124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.005248576868325472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a bizarre car carved from a giant potato as it bumps along a narrow, dirt country road. At the wheel, a man wearing old-fashioned aviator goggles and windswept hair, his face radiating pure joy and excitement. He's dressed in a frayed brown leather jacket and a plaid flannel shirt. The \"potato car\" retains its pristine, bumpy brown surface, even revealing a few newly sprouted buds. The car's body is crudely carved to its basic outline, with four rickety, old-fashioned spoked wheels of varying sizes wedged deep into the potato flesh. With every bump, the car trembles slightly like jelly. The man clings to a steering wheel, also made of a small potato, his body swaying violently, yet he laughs with delight. A light cloud of dust rises from behind. The warm golden afternoon sun fills the frame, lengthening the comical shadows cast by vehicles on the dusty road. Against the backdrop of rolling green hills and azure sky, the whole scene is permeated with a whimsical, joyful atmosphere full of freedom and adventure.", + "seed": 32065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_12/index_999_seed_32065_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": -0.022786738723516463, + "vlm_logprob": -0.022786738723516463, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.022786738723516463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0652417317032814 + ], + "opposite_logprob_tokens": [ + -2.815241813659668 + ], + "logprob_reward": -0.0652417317032814, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0652417317032814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13286888599395752, + 0.06494475156068802, + 0.04443357512354851, + 0.05308844894170761, + 0.07034225761890411, + 0.0239433404058218, + 0.12259268015623093, + 0.031684085726737976, + 0.25001534819602966, + 0.20608659088611603 + ], + "credit_map": [ + 0.008264840580523014, + 0.010257150046527386, + 0.011563185602426529, + 0.017197875306010246, + 0.025511879473924637, + 0.017355425283312798, + 0.027422798797488213, + 0.015295727178454399, + 0.008270430378615856, + 0.01502844039350748, + 0.0055976868607103825, + 0.0065389517694711685, + 0.007485616020858288, + 0.010360542684793472, + 0.005840389057993889, + 0.005822691600769758, + 0.004878858104348183, + 0.004889987874776125, + 0.004226514138281345, + 0.00735462224110961, + 0.007314550690352917, + 0.006035623140633106, + 0.004995373077690601, + 0.004738046787679195, + 0.006807275116443634, + 0.005898462608456612, + 0.004814917221665382, + 0.011043542064726353, + 0.0061243572272360325, + 0.0038276424165815115, + 0.008967170491814613, + 0.005605082958936691, + 0.011642697267234325, + 0.010607171803712845, + 0.008590389043092728, + 0.005796195473521948, + 0.007532423362135887, + 0.010816843248903751, + 0.007706745062023401, + 0.0076497942209243774, + 0.002280389191582799, + 0.0025802389718592167, + 0.002393105300143361, + 0.003997425548732281, + 0.004097977187484503, + 0.0024019775446504354, + 0.0031378536950796843, + 0.0030543734319508076, + 0.012863535434007645, + 0.012658435851335526, + 0.010231591761112213, + 0.014809628948569298, + 0.024558011442422867, + 0.012240353971719742, + 0.025035500526428223, + 0.010195612907409668, + 0.003802775638177991, + 0.0037906765937805176, + 0.0023114453069865704, + 0.004615278448909521, + 0.006474998313933611, + 0.0037609892897307873, + 0.004932621493935585, + 0.001995302736759186, + 0.044255580753088, + 0.02528226189315319, + 0.016298659145832062, + 0.037682946771383286, + 0.017344562336802483, + 0.05362474545836449, + 0.03630950301885605, + 0.019217099994421005, + 0.04932791367173195, + 0.0154183991253376, + 0.009892791509628296, + 0.027064979076385498, + 0.012441309168934822, + 0.04751746729016304, + 0.031216617673635483, + 0.01320710126310587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0652417317032814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801357984542847 + ], + "opposite_logprob_tokens": [ + -4.634801387786865 + ], + "logprob_reward": -0.009801357984542847, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801357984542847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18037769198417664, + 0.10344225913286209, + 0.07994777709245682, + 0.0504121370613575, + 0.0322260744869709, + 0.10385735332965851, + 0.1680038571357727, + 0.08043775707483292, + 0.09234769642353058, + 0.10894729197025299 + ], + "credit_map": [ + 0.014372173696756363, + 0.016458675265312195, + 0.015700895339250565, + 0.028646741062402725, + 0.02690448798239231, + 0.018267005681991577, + 0.036150433123111725, + 0.02387731708586216, + 0.010355272330343723, + 0.01080504059791565, + 0.010957716032862663, + 0.011776141822338104, + 0.026373205706477165, + 0.009538907557725906, + 0.01300850696861744, + 0.010627486743032932, + 0.008278784342110157, + 0.009311161004006863, + 0.0057842060923576355, + 0.009624515660107136, + 0.015486213378608227, + 0.011045084334909916, + 0.012820241041481495, + 0.007597590796649456, + 0.0068792495876550674, + 0.006774976383894682, + 0.005851347930729389, + 0.008487213402986526, + 0.003966436255723238, + 0.004818212706595659, + 0.008296429179608822, + 0.005338279530405998, + 0.00472470186650753, + 0.005097343120723963, + 0.004068102687597275, + 0.0047340854071080685, + 0.0033391721080988646, + 0.0025326109025627375, + 0.004519640002399683, + 0.003210424678400159, + 0.011224291287362576, + 0.010344975627958775, + 0.011133326217532158, + 0.020566444844007492, + 0.014514892362058163, + 0.010676389560103416, + 0.015491805970668793, + 0.009905251674354076, + 0.017533190548419952, + 0.017927397042512894, + 0.012708050198853016, + 0.01951134204864502, + 0.035289790481328964, + 0.016974257305264473, + 0.035301897674798965, + 0.012757963500916958, + 0.011731726117432117, + 0.007731262594461441, + 0.00569396885111928, + 0.011890108697116375, + 0.01156050618737936, + 0.012724671512842178, + 0.012617753818631172, + 0.006487774662673473, + 0.013239500112831593, + 0.010069559328258038, + 0.007191350217908621, + 0.014314444735646248, + 0.007814818061888218, + 0.014499444514513016, + 0.017128443345427513, + 0.008090157993137836, + 0.02539271116256714, + 0.008605679497122765, + 0.006320866756141186, + 0.007877948693931103, + 0.013095838017761707, + 0.027433816343545914, + 0.013385378755629063, + 0.006835076492279768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.009801357984542847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607341133058071 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607341133058071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607341133058071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3624829053878784, + 0.07904740422964096, + 0.07257876545190811, + 0.09483744204044342, + 0.024856485426425934, + 0.08834361284971237, + 0.10931035131216049, + 0.07281053811311722, + 0.0417044498026371, + 0.054028015583753586 + ], + "credit_map": [ + 0.023950934410095215, + 0.028459012508392334, + 0.03892830014228821, + 0.06510267406702042, + 0.062320105731487274, + 0.02857043780386448, + 0.06947016716003418, + 0.04568127542734146, + 0.008191293105483055, + 0.010182336904108524, + 0.007932573556900024, + 0.008539333939552307, + 0.012936105020344257, + 0.010527407750487328, + 0.010564344935119152, + 0.010174013674259186, + 0.007985910400748253, + 0.009149624966084957, + 0.005295144859701395, + 0.008637253195047379, + 0.012194590643048286, + 0.009156884625554085, + 0.009194073267281055, + 0.010965285822749138, + 0.009855423122644424, + 0.009799331426620483, + 0.01029259990900755, + 0.012839848175644875, + 0.013588744215667248, + 0.006561531685292721, + 0.01991475373506546, + 0.011985204182565212, + 0.0028471078258007765, + 0.0030927490442991257, + 0.003410656237974763, + 0.0030485803727060556, + 0.004362120758742094, + 0.003336938563734293, + 0.0027574568521231413, + 0.0020008771680295467, + 0.009087620303034782, + 0.00867121946066618, + 0.009336980059742928, + 0.01825864240527153, + 0.01140027865767479, + 0.009300210513174534, + 0.013984370045363903, + 0.008304288610816002, + 0.01125749759376049, + 0.010796909220516682, + 0.009745323099195957, + 0.012859291397035122, + 0.02290445752441883, + 0.01087753102183342, + 0.01964484341442585, + 0.01122449990361929, + 0.01360959280282259, + 0.006254637613892555, + 0.004094753414392471, + 0.008339375257492065, + 0.008628726936876774, + 0.014810820110142231, + 0.008074494078755379, + 0.008998144418001175, + 0.005732691381126642, + 0.003637977410107851, + 0.0031865998171269894, + 0.004879026673734188, + 0.005972492508590221, + 0.006663813721388578, + 0.008225136436522007, + 0.0034067099913954735, + 0.010201327502727509, + 0.004122392274439335, + 0.0040563009679317474, + 0.010861965827643871, + 0.004154462367296219, + 0.008801301941275597, + 0.008401270024478436, + 0.0034289974719285965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.01607341133058071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818414404988289 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.01818414404988289, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818414404988289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19624300301074982, + 0.08424493670463562, + 0.047079116106033325, + 0.23073014616966248, + 0.05817358195781708, + 0.08017531037330627, + 0.08588852733373642, + 0.08342790603637695, + 0.08064531534910202, + 0.05339214205741882 + ], + "credit_map": [ + 0.014658221043646336, + 0.019569773226976395, + 0.026108041405677795, + 0.032039422541856766, + 0.02182432822883129, + 0.028567295521497726, + 0.03115026280283928, + 0.022325672209262848, + 0.008862394839525223, + 0.012314993888139725, + 0.01070371177047491, + 0.0104245375841856, + 0.012883464805781841, + 0.010612729005515575, + 0.008341556414961815, + 0.010101553052663803, + 0.005084304604679346, + 0.0044076284393668175, + 0.004575147293508053, + 0.006770952604711056, + 0.006077849771827459, + 0.005757114849984646, + 0.008116921409964561, + 0.006289198063313961, + 0.025501878932118416, + 0.023792676627635956, + 0.043342225253582, + 0.028279541060328484, + 0.022692406550049782, + 0.016734300181269646, + 0.03287417069077492, + 0.037512946873903275, + 0.007234412711113691, + 0.005716262850910425, + 0.008529726415872574, + 0.0073946379125118256, + 0.004561121575534344, + 0.00740845687687397, + 0.006648501381278038, + 0.01068046223372221, + 0.007256535813212395, + 0.008676964789628983, + 0.007725495845079422, + 0.016537107527256012, + 0.014730599708855152, + 0.00769171305000782, + 0.011364526115357876, + 0.006192365195602179, + 0.010924022644758224, + 0.008534540422260761, + 0.009497893042862415, + 0.009737806394696236, + 0.01778981275856495, + 0.009134297259151936, + 0.011440595611929893, + 0.00882955826818943, + 0.010901962406933308, + 0.007434808183461428, + 0.007830681279301643, + 0.010684107430279255, + 0.014210306107997894, + 0.010361051186919212, + 0.011413170956075191, + 0.010591821745038033, + 0.016490023583173752, + 0.006068249698728323, + 0.005884441081434488, + 0.009161202237010002, + 0.006984058301895857, + 0.017412034794688225, + 0.009551570750772953, + 0.009093744680285454, + 0.012009291909635067, + 0.004422397818416357, + 0.003737044520676136, + 0.005845136474817991, + 0.003499970305711031, + 0.013559969142079353, + 0.003892752109095454, + 0.006425580475479364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.01818414404988289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633048549294472 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633048549294472, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633048549294472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3526584506034851, + 0.14012260735034943, + 0.05288786068558693, + 0.05383013188838959, + 0.04400232061743736, + 0.09126719832420349, + 0.11168861389160156, + 0.031104223802685738, + 0.06369529664516449, + 0.05874330550432205 + ], + "credit_map": [ + 0.02222723700106144, + 0.02619941718876362, + 0.04070664197206497, + 0.06790819019079208, + 0.054396796971559525, + 0.020544948056340218, + 0.06563781201839447, + 0.05503735691308975, + 0.015961192548274994, + 0.015221554785966873, + 0.014701965264976025, + 0.017269346863031387, + 0.018223214894533157, + 0.017883924767374992, + 0.02048395574092865, + 0.02037745527923107, + 0.006857582367956638, + 0.0048042102716863155, + 0.005237743258476257, + 0.0068025426007807255, + 0.00672053312882781, + 0.008395391516387463, + 0.008655552752315998, + 0.005414302926510572, + 0.005502589978277683, + 0.0037873066030442715, + 0.006399799138307571, + 0.008373967371881008, + 0.005716209765523672, + 0.003923875279724598, + 0.01219476293772459, + 0.00793161429464817, + 0.005220585502684116, + 0.0069670542143285275, + 0.00492381677031517, + 0.0033811975736171007, + 0.0045672873966395855, + 0.008988792076706886, + 0.005040240474045277, + 0.004913344047963619, + 0.014820225536823273, + 0.009256495162844658, + 0.009551032446324825, + 0.015005798079073429, + 0.009680966846644878, + 0.01115912664681673, + 0.012157157063484192, + 0.00963639561086893, + 0.015665093436837196, + 0.00950666330754757, + 0.02078024297952652, + 0.011534745804965496, + 0.012492612935602665, + 0.011816814541816711, + 0.014796800911426544, + 0.015095633454620838, + 0.0036334076430648565, + 0.004048922564834356, + 0.00260086078196764, + 0.0062011717818677425, + 0.003518542507663369, + 0.003146560862660408, + 0.004303367342799902, + 0.0036513893865048885, + 0.01589369960129261, + 0.006025393959134817, + 0.0027138972654938698, + 0.003906590864062309, + 0.00505110714584589, + 0.017892014235258102, + 0.008054069243371487, + 0.004158519208431244, + 0.011968869715929031, + 0.004008469171822071, + 0.0029800389893352985, + 0.006336564663797617, + 0.004699186887592077, + 0.012397903949022293, + 0.009397637099027634, + 0.006954634562134743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004633048549294472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a bizarre car carved from a giant potato as it bumps along a narrow, dirt country road. At the wheel, a man wearing old-fashioned aviator goggles and windswept hair, his face radiating pure joy and excitement. He's dressed in a frayed brown leather jacket and a plaid flannel shirt. The \"potato car\" retains its pristine, bumpy brown surface, even revealing a few newly sprouted buds. The car's body is crudely carved to its basic outline, with four rickety, old-fashioned spoked wheels of varying sizes wedged deep into the potato flesh. With every bump, the car trembles slightly like jelly. The man clings to a steering wheel, also made of a small potato, his body swaying violently, yet he laughs with delight. A light cloud of dust rises from behind. The warm golden afternoon sun fills the frame, lengthening the comical shadows cast by vehicles on the dusty road. Against the backdrop of rolling green hills and azure sky, the whole scene is permeated with a whimsical, joyful atmosphere full of freedom and adventure.", + "seed": 32066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_12/index_999_seed_32066_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": -0.014448216604068875, + "vlm_logprob": -0.014448216604068875, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.014448216604068875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.036516934633255005 + ], + "opposite_logprob_tokens": [ + -3.4115169048309326 + ], + "logprob_reward": -0.036516934633255005, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036516934633255005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18466684222221375, + 0.03525721654295921, + 0.04722204431891441, + 0.34059199690818787, + 0.13576051592826843, + 0.09893257915973663, + 0.04081122204661369, + 0.04628903418779373, + 0.032694343477487564, + 0.037774235010147095 + ], + "credit_map": [ + 0.01353900134563446, + 0.017851123586297035, + 0.018380306661128998, + 0.022878902032971382, + 0.03389029577374458, + 0.016644056886434555, + 0.04033481702208519, + 0.021148329600691795, + 0.003931219223886728, + 0.0034474339336156845, + 0.003540302626788616, + 0.00512743229046464, + 0.0066360970959067345, + 0.0034499114844948053, + 0.0062044053338468075, + 0.00292041408829391, + 0.0044839042238891125, + 0.006386790424585342, + 0.0069877635687589645, + 0.004153235349804163, + 0.006013731937855482, + 0.008061973378062248, + 0.004915669560432434, + 0.006218974478542805, + 0.051032040268182755, + 0.118691086769104, + 0.08046245574951172, + 0.02572007104754448, + 0.023985888808965683, + 0.013293023221194744, + 0.011551376432180405, + 0.015856066718697548, + 0.009111249819397926, + 0.016110075637698174, + 0.020757073536515236, + 0.015840692445635796, + 0.01597769372165203, + 0.01150443498045206, + 0.022170400246977806, + 0.024288887158036232, + 0.007099917624145746, + 0.008977334015071392, + 0.014806404709815979, + 0.011683901771903038, + 0.01821020059287548, + 0.010079264640808105, + 0.016657885164022446, + 0.011417672038078308, + 0.0035680835135281086, + 0.006029508076608181, + 0.0030094117391854525, + 0.00935988686978817, + 0.005727197043597698, + 0.004383689723908901, + 0.00458043348044157, + 0.004153011366724968, + 0.004200538154691458, + 0.006480349227786064, + 0.004541827365756035, + 0.008542867377400398, + 0.006154125556349754, + 0.006211661733686924, + 0.004256118088960648, + 0.005901549942791462, + 0.0025077080354094505, + 0.0027679174672812223, + 0.0036388023290783167, + 0.00353103200905025, + 0.008254677057266235, + 0.004113889764994383, + 0.0041165584698319435, + 0.0037637597415596247, + 0.00397879583761096, + 0.0028790771029889584, + 0.005347004160284996, + 0.005241821054369211, + 0.005479805171489716, + 0.006154497154057026, + 0.0049055577255785465, + 0.003787674242630601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.036516934633255005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764006981626153 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.00764006981626153, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764006981626153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14098049700260162, + 0.08662129938602448, + 0.04318152368068695, + 0.2860046923160553, + 0.054839543998241425, + 0.06342561542987823, + 0.06353949010372162, + 0.07963795959949493, + 0.06800979375839233, + 0.11375955492258072 + ], + "credit_map": [ + 0.011963818222284317, + 0.012194769456982613, + 0.014860056340694427, + 0.019122494384646416, + 0.025104112923145294, + 0.015327485278248787, + 0.02633722871541977, + 0.016070540994405746, + 0.005787889007478952, + 0.008515948429703712, + 0.009804473258554935, + 0.012558909133076668, + 0.016013450920581818, + 0.008242951706051826, + 0.016232311725616455, + 0.009465368464589119, + 0.003381266724318266, + 0.00441268365830183, + 0.007795701269060373, + 0.004375780001282692, + 0.005537916906177998, + 0.006946064997464418, + 0.005537174642086029, + 0.005194942466914654, + 0.045838601887226105, + 0.10429234802722931, + 0.06255544722080231, + 0.019472049549221992, + 0.021037109196186066, + 0.011167561635375023, + 0.008926808834075928, + 0.012714773416519165, + 0.004828108008950949, + 0.0051415106281638145, + 0.004863282199949026, + 0.009416958317160606, + 0.0093739815056324, + 0.005519805010408163, + 0.00862740445882082, + 0.0070684971287846565, + 0.0063256146386265755, + 0.006831515580415726, + 0.010627437382936478, + 0.0071027339436113834, + 0.0087058050557971, + 0.00624370202422142, + 0.00909376610070467, + 0.008495046757161617, + 0.005435854196548462, + 0.00903471652418375, + 0.005746963433921337, + 0.009718047454953194, + 0.015276597812771797, + 0.004671621602028608, + 0.007368618156760931, + 0.0062870760448277, + 0.008105030283331871, + 0.014292551204562187, + 0.0075119840912520885, + 0.012186000123620033, + 0.010531858541071415, + 0.009681487455964088, + 0.009781450033187866, + 0.0075476039201021194, + 0.0044672987423837185, + 0.004069989547133446, + 0.007538776844739914, + 0.007406770717352629, + 0.01370393205434084, + 0.013482355512678623, + 0.008257713168859482, + 0.009082959964871407, + 0.00930813979357481, + 0.008972473442554474, + 0.013134604319930077, + 0.022592246532440186, + 0.016374841332435608, + 0.012403419241309166, + 0.012371942400932312, + 0.018601901829242706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.00764006981626153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076509952545166 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076509952545166, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076509952545166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17898187041282654, + 0.0722508504986763, + 0.11668803542852402, + 0.0690034031867981, + 0.11638324707746506, + 0.1477833390235901, + 0.1182403564453125, + 0.04497602581977844, + 0.04710103198885918, + 0.08859185874462128 + ], + "credit_map": [ + 0.015649013221263885, + 0.01458931528031826, + 0.01852988451719284, + 0.01841716095805168, + 0.043023381382226944, + 0.01813530921936035, + 0.030948618426918983, + 0.019689183682203293, + 0.004123759921640158, + 0.00647633895277977, + 0.008216232061386108, + 0.01308494433760643, + 0.010526136495172977, + 0.005950425751507282, + 0.015932178124785423, + 0.007940832525491714, + 0.008291911333799362, + 0.009253966622054577, + 0.014448113739490509, + 0.009938081726431847, + 0.016290631145238876, + 0.02444557286798954, + 0.01906137354671955, + 0.014958391897380352, + 0.007664387580007315, + 0.0190774817019701, + 0.017079010605812073, + 0.006132570561021566, + 0.004686834756284952, + 0.005275107454508543, + 0.004491348750889301, + 0.004596663173288107, + 0.010665904730558395, + 0.024545826017856598, + 0.01219925470650196, + 0.011258323676884174, + 0.009931422770023346, + 0.009933843277394772, + 0.01644480600953102, + 0.02140386961400509, + 0.022166509181261063, + 0.02475563995540142, + 0.027767973020672798, + 0.014498130418360233, + 0.011848749592900276, + 0.015565343201160431, + 0.014362002722918987, + 0.016818992793560028, + 0.00792563147842884, + 0.014802866615355015, + 0.013890443369746208, + 0.025755710899829865, + 0.01767369545996189, + 0.010837813839316368, + 0.011773903854191303, + 0.015580286271870136, + 0.0033323029056191444, + 0.005166070535778999, + 0.004873065743595362, + 0.010424432344734669, + 0.005547093693166971, + 0.005575654562562704, + 0.004312328528612852, + 0.005745076574385166, + 0.0039229863323271275, + 0.0037055620923638344, + 0.0077936649322509766, + 0.008200728334486485, + 0.006486242637038231, + 0.004883183166384697, + 0.0051568602211773396, + 0.006951802875846624, + 0.0077506923116743565, + 0.006357324775308371, + 0.014544801786541939, + 0.018249325454235077, + 0.00971447117626667, + 0.009245707653462887, + 0.008719255216419697, + 0.014010285958647728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.011076509952545166 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19637136161327362, + 0.1296221762895584, + 0.15891790390014648, + 0.05642426386475563, + 0.08206005394458771, + 0.10632801800966263, + 0.10924889147281647, + 0.04784087464213371, + 0.06567955017089844, + 0.04750685393810272 + ], + "credit_map": [ + 0.021060850471258163, + 0.019407100975513458, + 0.026705259457230568, + 0.018930189311504364, + 0.02709713578224182, + 0.026093952357769012, + 0.03330397233366966, + 0.023772913962602615, + 0.008437360636889935, + 0.011995981447398663, + 0.014613758772611618, + 0.02077723667025566, + 0.019681427627801895, + 0.012501455843448639, + 0.02705317549407482, + 0.014561775140464306, + 0.015648774802684784, + 0.018214590847492218, + 0.026717180386185646, + 0.012500906363129616, + 0.016841977834701538, + 0.035500116646289825, + 0.016750985756516457, + 0.016743365675210953, + 0.007168140262365341, + 0.011415586806833744, + 0.009778025560081005, + 0.007008131593465805, + 0.005676652304828167, + 0.006061140447854996, + 0.005607994273304939, + 0.0037085944786667824, + 0.006420179270207882, + 0.01010239589959383, + 0.009526848793029785, + 0.010780475102365017, + 0.006841131020337343, + 0.007405491080135107, + 0.012978588230907917, + 0.018004940822720528, + 0.011542486026883125, + 0.019862432032823563, + 0.013077715411782265, + 0.010967141017317772, + 0.01974799856543541, + 0.01074559148401022, + 0.012113886885344982, + 0.008270771242678165, + 0.008459706790745258, + 0.014044186100363731, + 0.010211297310888767, + 0.01711297407746315, + 0.01941131241619587, + 0.010996613651514053, + 0.01652902364730835, + 0.012483772821724415, + 0.0034814379177987576, + 0.006276084575802088, + 0.006866813637316227, + 0.009053224697709084, + 0.0077329883351922035, + 0.004683217033743858, + 0.004354176111519337, + 0.005392931401729584, + 0.005426776595413685, + 0.004881335888057947, + 0.00928210187703371, + 0.009697727859020233, + 0.01544744148850441, + 0.00578076858073473, + 0.007310774177312851, + 0.007852624170482159, + 0.007691887207329273, + 0.004494743421673775, + 0.0058383760042488575, + 0.006356006488204002, + 0.004712296184152365, + 0.006161034107208252, + 0.005112180951982737, + 0.007140335161238909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014192949049174786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a car carved from a giant potato?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver a man wearing old-fashioned aviator goggles and a frayed brown leather jacket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car have four rickety, old-fashioned spoked wheels of varying sizes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot taken from a low angle while tracking the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a narrow dirt country road with rolling green hills and an azure sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00281461956910789 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.00281461956910789, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00281461956910789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1941019594669342, + 0.1106005385518074, + 0.05579325929284096, + 0.18893639743328094, + 0.09279517829418182, + 0.1900845617055893, + 0.03778811916708946, + 0.02686549350619316, + 0.04434879869222641, + 0.058685656636953354 + ], + "credit_map": [ + 0.0196151714771986, + 0.01680150255560875, + 0.02774285152554512, + 0.027629753574728966, + 0.024194961413741112, + 0.02051348052918911, + 0.031229235231876373, + 0.02637503482401371, + 0.008557822555303574, + 0.00952215027064085, + 0.013143028132617474, + 0.02006659470498562, + 0.01374899037182331, + 0.01073681004345417, + 0.023257797583937645, + 0.011567356996238232, + 0.005718755070120096, + 0.005463864654302597, + 0.008855639025568962, + 0.006670921575278044, + 0.005650012753903866, + 0.006826964672654867, + 0.007134855259209871, + 0.009472250938415527, + 0.032432034611701965, + 0.06006096303462982, + 0.026477418839931488, + 0.016099464148283005, + 0.01671757735311985, + 0.012497778050601482, + 0.014560597017407417, + 0.01009057741612196, + 0.01059561688452959, + 0.010548343881964684, + 0.019664132967591286, + 0.008738848380744457, + 0.009351241402328014, + 0.008191726170480251, + 0.013506120070815086, + 0.01219914574176073, + 0.025043725967407227, + 0.028241224586963654, + 0.030688220635056496, + 0.03237764909863472, + 0.018474195152521133, + 0.019646095111966133, + 0.01789015159010887, + 0.0177233275026083, + 0.003266362939029932, + 0.0031844819895923138, + 0.005225083325058222, + 0.005845010280609131, + 0.0042793829925358295, + 0.004695609211921692, + 0.003721850924193859, + 0.007570343092083931, + 0.00202280655503273, + 0.002884587738662958, + 0.003207295900210738, + 0.006720769684761763, + 0.003878452815115452, + 0.0026599233970046043, + 0.002529970370233059, + 0.002961688442155719, + 0.004377780482172966, + 0.003117963206022978, + 0.007120805326849222, + 0.00978558324277401, + 0.00575372576713562, + 0.005078899674117565, + 0.003701250534504652, + 0.00541279511526227, + 0.009633149020373821, + 0.005663319956511259, + 0.005710778292268515, + 0.00831996276974678, + 0.0059391409158706665, + 0.00796956941485405, + 0.004753305576741695, + 0.01069643348455429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.00281461956910789 + } + ], + "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/0000155/group_13/index_tensor([420, 420])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_tensor([420, 420])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..62272ff535299a5bf6304fc1d8819097e85ed503 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_tensor([420, 420])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot begins from a slight overhead angle, then steadily pulls back, ultimately culminating in a magnificent long shot. The camera initially focuses on a small, meticulously folded paper boat, crafted from an acceptance letter. The thick, off-white surface of the paper reveals faintly blurred, black Song-style calligraphy, slightly dampened by moisture. Clear creases outline its small, yet resolute outline. Steadily and slowly, it drifts along the gentle ripples of the water. Under the warm, soft golden light of the setting sun, the boat is gilded with a dreamy golden edge. As the camera pulls out, the vast expanse of water within it is gradually revealed. The tranquil, clear, and deep waters reflect the vibrant sunset glow, tinged in warm orange and soft pink-purple. The ripples shimmer where the water meets the sky. Ultimately, the tiny paper boat appears both small and powerful within a vast, vibrant world. The warm tones of the entire scene create a poetic atmosphere of tranquility, hope, and a touch of loneliness.", + "seed": 20085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_420_seed_20085_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": -0.004527107928879559, + "vlm_logprob": -0.004527107928879559, + "vlm_yes_no_margin": 5.575, + "vlm_token_credit": -0.004527107928879559, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.008548925630748272 + ], + "opposite_logprob_tokens": [ + -4.883548736572266 + ], + "logprob_reward": -0.008548925630748272, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008548925630748272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09122483432292938, + 0.1151842325925827, + 0.2542019784450531, + 0.12480787932872772, + 0.09666897356510162, + 0.13142326474189758, + 0.07371554523706436, + 0.043553587049245834, + 0.030724624171853065, + 0.03849514201283455 + ], + "credit_map": [ + 0.017813164740800858, + 0.008455194532871246, + 0.022608807310461998, + 0.009504124522209167, + 0.011893704533576965, + 0.006485632620751858, + 0.005416763946413994, + 0.009047439321875572, + 0.0267951637506485, + 0.00893410574644804, + 0.019693860784173012, + 0.007133502513170242, + 0.007216820493340492, + 0.008068887516856194, + 0.006154150702059269, + 0.031187724322080612, + 0.022058801725506783, + 0.02911933697760105, + 0.04076341912150383, + 0.10687942802906036, + 0.010274169966578484, + 0.00933649018406868, + 0.018817057833075523, + 0.01695326343178749, + 0.021781472489237785, + 0.022776491940021515, + 0.028395945206284523, + 0.015126309357583523, + 0.01074230670928955, + 0.007191258482635021, + 0.011234648525714874, + 0.007559445220977068, + 0.012625059112906456, + 0.019568119198083878, + 0.01247642282396555, + 0.013574458658695221, + 0.014339679852128029, + 0.006908653303980827, + 0.011139062233269215, + 0.006037507206201553, + 0.03261421620845795, + 0.013541738502681255, + 0.03968963027000427, + 0.008708109147846699, + 0.011118810623884201, + 0.00749429315328598, + 0.00969006959348917, + 0.008566375821828842, + 0.02520817704498768, + 0.02019183523952961, + 0.006093943957239389, + 0.0034618417266756296, + 0.003453112207353115, + 0.005270785186439753, + 0.003228584537282586, + 0.006807253696024418, + 0.005370850674808025, + 0.0052313851192593575, + 0.0074998377822339535, + 0.004854203201830387, + 0.005425179377198219, + 0.0061552091501653194, + 0.004142621997743845, + 0.0048742941580712795, + 0.003911755047738552, + 0.0052208309061825275, + 0.003804258769378066, + 0.003784544300287962, + 0.004056315403431654, + 0.0034365872852504253, + 0.0041413214057683945, + 0.0023690033704042435, + 0.004467671737074852, + 0.006232669111341238, + 0.004862755071371794, + 0.0035302871838212013, + 0.005441009532660246, + 0.006038071122020483, + 0.004923421889543533, + 0.002999254735186696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.008548925630748272 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.003619787283241749 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.003619787283241749, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003619787283241749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08050539344549179, + 0.192116379737854, + 0.16562297940254211, + 0.05404943972826004, + 0.033282820135354996, + 0.16628581285476685, + 0.07728542387485504, + 0.0351041778922081, + 0.11609871685504913, + 0.07964880019426346 + ], + "credit_map": [ + 0.009098825976252556, + 0.011969232000410557, + 0.010290302336215973, + 0.009522226639091969, + 0.008436012081801891, + 0.011844336055219173, + 0.009000012651085854, + 0.010344442911446095, + 0.029862485826015472, + 0.030456900596618652, + 0.025102835148572922, + 0.024651246145367622, + 0.022640272974967957, + 0.016165366396307945, + 0.014322998002171516, + 0.02891426719725132, + 0.01590646058320999, + 0.017039289698004723, + 0.026460614055395126, + 0.0642215684056282, + 0.009343833662569523, + 0.00827081874012947, + 0.013757915236055851, + 0.010622480884194374, + 0.00788013357669115, + 0.009943732060492039, + 0.01001930795609951, + 0.006174406502395868, + 0.006183052901178598, + 0.003655117005109787, + 0.005756236147135496, + 0.004437458701431751, + 0.006894758436828852, + 0.0060747480019927025, + 0.005604553502053022, + 0.0028842573519796133, + 0.005030481610447168, + 0.0017830643337219954, + 0.002651915652677417, + 0.0023590377531945705, + 0.04791151359677315, + 0.014156694523990154, + 0.06602226197719574, + 0.008953753858804703, + 0.009257251396775246, + 0.004830657970160246, + 0.005266668740659952, + 0.009887012653052807, + 0.028219735249876976, + 0.022014301270246506, + 0.005006452091038227, + 0.0027876459062099457, + 0.0027959798462688923, + 0.004339991137385368, + 0.003104320727288723, + 0.00901699811220169, + 0.004987041931599379, + 0.005626544822007418, + 0.006028986535966396, + 0.0032590660266578197, + 0.0043960437178611755, + 0.0031405154149979353, + 0.0030189298558980227, + 0.004647047724574804, + 0.020323850214481354, + 0.02228989452123642, + 0.014080751687288284, + 0.009558298625051975, + 0.012891091406345367, + 0.01634526066482067, + 0.010071580298244953, + 0.010537986643612385, + 0.008568577468395233, + 0.011108147911727428, + 0.010909694246947765, + 0.008493070490658283, + 0.01127699390053749, + 0.01150404755026102, + 0.012278414331376553, + 0.005509852897375822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.003619787283241749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.0017124050064012408 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017124050064012408, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017124050064012408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1076163724064827, + 0.27064377069473267, + 0.12177535146474838, + 0.06415979564189911, + 0.04647773876786232, + 0.03960045054554939, + 0.0434194840490818, + 0.04659511148929596, + 0.20080450177192688, + 0.0589073970913887 + ], + "credit_map": [ + 0.02886565774679184, + 0.008376330137252808, + 0.02954663708806038, + 0.007447327021509409, + 0.00942966528236866, + 0.00927650835365057, + 0.00601807190105319, + 0.008656173013150692, + 0.038629110902547836, + 0.010839981958270073, + 0.11289232224225998, + 0.016382338479161263, + 0.010717589408159256, + 0.01416538655757904, + 0.011312037706375122, + 0.05570501834154129, + 0.012876254506409168, + 0.01493819896131754, + 0.02189728245139122, + 0.03463592007756233, + 0.00742732360959053, + 0.005619552452117205, + 0.00895314384251833, + 0.01542768906801939, + 0.01012685801833868, + 0.009221626445651054, + 0.014693780802190304, + 0.007907033897936344, + 0.007427409756928682, + 0.004198178648948669, + 0.006491592153906822, + 0.004093322437256575, + 0.008388230577111244, + 0.00615472299978137, + 0.0061446139588952065, + 0.004819265566766262, + 0.009342364966869354, + 0.0033897734247148037, + 0.004937131889164448, + 0.0033016400411725044, + 0.006036513019353151, + 0.005961947608739138, + 0.0070663862861692905, + 0.004127469379454851, + 0.0045204260386526585, + 0.0033742294181138277, + 0.0053232884965837, + 0.003190194256603718, + 0.00564034515991807, + 0.011578594334423542, + 0.0062068854458630085, + 0.0029873240273445845, + 0.0035502808168530464, + 0.003188219154253602, + 0.0032655876129865646, + 0.007002249360084534, + 0.005444880109280348, + 0.009571937844157219, + 0.008448026143014431, + 0.004460473544895649, + 0.0039138514548540115, + 0.0035836605820804834, + 0.0036351110320538282, + 0.007537176366895437, + 0.0395367369055748, + 0.03719314560294151, + 0.022878332063555717, + 0.01697877235710621, + 0.022134196013212204, + 0.02655722014605999, + 0.01859445497393608, + 0.01693165861070156, + 0.006253390107303858, + 0.0074855550192296505, + 0.007960464805364609, + 0.006781758740544319, + 0.007720833644270897, + 0.008089693263173103, + 0.01036234013736248, + 0.004253363236784935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.0017124050064012408 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.0059419418685138226 + ], + "opposite_logprob_tokens": [ + -5.130941867828369 + ], + "logprob_reward": -0.0059419418685138226, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059419418685138226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04593586549162865, + 0.10116693377494812, + 0.05532122775912285, + 0.050489433109760284, + 0.050097521394491196, + 0.12478596717119217, + 0.33909907937049866, + 0.08980745822191238, + 0.06294728070497513, + 0.08034916967153549 + ], + "credit_map": [ + 0.006105139851570129, + 0.004450778942555189, + 0.006972599774599075, + 0.004799544811248779, + 0.004892706871032715, + 0.007858334109187126, + 0.004579134751111269, + 0.006277626845985651, + 0.012472273781895638, + 0.005917106289416552, + 0.04073312133550644, + 0.006008113734424114, + 0.0035890822764486074, + 0.0054731001146137714, + 0.004288011696189642, + 0.02268613502383232, + 0.006729471031576395, + 0.006750489119440317, + 0.01433841697871685, + 0.006222511641681194, + 0.003735595615580678, + 0.004207920748740435, + 0.004197569098323584, + 0.009139257483184338, + 0.008050370961427689, + 0.009512484073638916, + 0.0069886124692857265, + 0.006088623311370611, + 0.006999857723712921, + 0.00338139315135777, + 0.0059966216795146465, + 0.0034714709036052227, + 0.012991036288440228, + 0.008841591887176037, + 0.00902589876204729, + 0.0057951463386416435, + 0.003382737748324871, + 0.0024684991221874952, + 0.0033371944446116686, + 0.004255422856658697, + 0.02921992540359497, + 0.010673911310732365, + 0.0462070032954216, + 0.007622729986906052, + 0.008638248778879642, + 0.0049953097477555275, + 0.0074857319705188274, + 0.009943118318915367, + 0.12474258989095688, + 0.10600215941667557, + 0.01790853589773178, + 0.011436540633440018, + 0.013774687424302101, + 0.01709953509271145, + 0.011987481266260147, + 0.036147601902484894, + 0.010689790360629559, + 0.018452847376465797, + 0.0194283127784729, + 0.008931105956435204, + 0.0074063921347260475, + 0.0069929021410644054, + 0.006243812385946512, + 0.011662298813462257, + 0.009623215533792973, + 0.009760740213096142, + 0.012646527960896492, + 0.006986407097429037, + 0.005904663819819689, + 0.0055332910269498825, + 0.009184019640088081, + 0.0033084233291447163, + 0.008011662401258945, + 0.010297254659235477, + 0.009581862948834896, + 0.009913465939462185, + 0.00868359487503767, + 0.009044784121215343, + 0.01895892806351185, + 0.005857622250914574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.0059419418685138226 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.002812479855492711 + ], + "opposite_logprob_tokens": [ + -5.877812385559082 + ], + "logprob_reward": -0.002812479855492711, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002812479855492711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18971134722232819, + 0.15415827929973602, + 0.08190533518791199, + 0.08441329002380371, + 0.049692414700984955, + 0.1260477751493454, + 0.08247755467891693, + 0.08131404221057892, + 0.05533719062805176, + 0.0949428528547287 + ], + "credit_map": [ + 0.05202759802341461, + 0.017239896580576897, + 0.05848544090986252, + 0.013267655856907368, + 0.01634090021252632, + 0.012556951493024826, + 0.008235767483711243, + 0.01155711431056261, + 0.020812636241316795, + 0.008114496245980263, + 0.049976930022239685, + 0.010825930163264275, + 0.004950546659529209, + 0.009184337221086025, + 0.006824086885899305, + 0.04346929118037224, + 0.0085123460739851, + 0.005344124510884285, + 0.030845385044813156, + 0.016673441976308823, + 0.004185484256595373, + 0.00326779973693192, + 0.004272162448614836, + 0.008804578334093094, + 0.014066656120121479, + 0.015557531267404556, + 0.014919314533472061, + 0.011231950484216213, + 0.007404661737382412, + 0.005854858551174402, + 0.009564428590238094, + 0.0058138808235526085, + 0.009706538170576096, + 0.009681890718638897, + 0.007427569013088942, + 0.007423199713230133, + 0.004321194253861904, + 0.0027035772800445557, + 0.0044669886119663715, + 0.003961453214287758, + 0.03011346235871315, + 0.01485533732920885, + 0.03923993557691574, + 0.009493207558989525, + 0.010096152313053608, + 0.005779503844678402, + 0.008608948439359665, + 0.007861215621232986, + 0.013102402910590172, + 0.016612913459539413, + 0.015345231629908085, + 0.006673569791018963, + 0.008651288226246834, + 0.0075918156653642654, + 0.006095947232097387, + 0.008404375053942204, + 0.013620661571621895, + 0.020181426778435707, + 0.01237278338521719, + 0.00706661120057106, + 0.009275837801396847, + 0.006952216383069754, + 0.006303604692220688, + 0.005540891550481319, + 0.005630361381918192, + 0.008130352944135666, + 0.006522731389850378, + 0.006226499564945698, + 0.008896645158529282, + 0.007861100137233734, + 0.00786577444523573, + 0.004203717224299908, + 0.01012647058814764, + 0.012624439783394337, + 0.012870858423411846, + 0.008395570330321789, + 0.015091229230165482, + 0.01793782413005829, + 0.010807485319674015, + 0.0070889610797166824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.002812479855492711 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot begins from a slight overhead angle, then steadily pulls back, ultimately culminating in a magnificent long shot. The camera initially focuses on a small, meticulously folded paper boat, crafted from an acceptance letter. The thick, off-white surface of the paper reveals faintly blurred, black Song-style calligraphy, slightly dampened by moisture. Clear creases outline its small, yet resolute outline. Steadily and slowly, it drifts along the gentle ripples of the water. Under the warm, soft golden light of the setting sun, the boat is gilded with a dreamy golden edge. As the camera pulls out, the vast expanse of water within it is gradually revealed. The tranquil, clear, and deep waters reflect the vibrant sunset glow, tinged in warm orange and soft pink-purple. The ripples shimmer where the water meets the sky. Ultimately, the tiny paper boat appears both small and powerful within a vast, vibrant world. The warm tones of the entire scene create a poetic atmosphere of tranquility, hope, and a touch of loneliness.", + "seed": 20086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_420_seed_20086_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": -0.0100768041331321, + "vlm_logprob": -0.0100768041331321, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.0100768041331321, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.03587677329778671 + ], + "opposite_logprob_tokens": [ + -3.410876750946045 + ], + "logprob_reward": -0.03587677329778671, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03587677329778671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4760449528694153, + 0.1258917599916458, + 0.045542310923337936, + 0.06901663541793823, + 0.03689216449856758, + 0.03832510486245155, + 0.020365724340081215, + 0.05624740570783615, + 0.09173506498336792, + 0.039938874542713165 + ], + "credit_map": [ + 0.06580250710248947, + 0.043127816170454025, + 0.01951785944402218, + 0.05518265441060066, + 0.09265416860580444, + 0.041484687477350235, + 0.12580321729183197, + 0.03247205168008804, + 0.024301130324602127, + 0.018931657075881958, + 0.010488845407962799, + 0.012572615407407284, + 0.012477402575314045, + 0.012441829778254032, + 0.015862952917814255, + 0.018815329298377037, + 0.005562600214034319, + 0.0033936025574803352, + 0.00469207065179944, + 0.004544815979897976, + 0.005013857036828995, + 0.01192039530724287, + 0.004339330364018679, + 0.006075641140341759, + 0.006985090207308531, + 0.009166461415588856, + 0.012082078494131565, + 0.008436918258666992, + 0.008887672796845436, + 0.007497431710362434, + 0.007529775146394968, + 0.00843120738863945, + 0.005272830370813608, + 0.004001763183623552, + 0.005588860716670752, + 0.004259121138602495, + 0.0029261650051921606, + 0.005408233497291803, + 0.0032314062118530273, + 0.006203782744705677, + 0.005338944494724274, + 0.00575920008122921, + 0.003934221342206001, + 0.003961765207350254, + 0.005464792717248201, + 0.007014356553554535, + 0.003535850439220667, + 0.0033159744925796986, + 0.002168637467548251, + 0.0025041562039405107, + 0.0020664751064032316, + 0.00320047908462584, + 0.0028725306037813425, + 0.002253521466627717, + 0.0029103991109877825, + 0.0023895243648439646, + 0.005467322655022144, + 0.00918966718018055, + 0.007314865943044424, + 0.005311853718012571, + 0.007444312330335379, + 0.010222443379461765, + 0.005228362511843443, + 0.006068579386919737, + 0.0072244214825332165, + 0.01542387343943119, + 0.014256940223276615, + 0.009232008829712868, + 0.00976323802024126, + 0.011015968397259712, + 0.008317103609442711, + 0.01650151237845421, + 0.0031820032745599747, + 0.00460388045758009, + 0.0041330475360155106, + 0.005178689956665039, + 0.006633027456700802, + 0.0057641626335680485, + 0.00623560044914484, + 0.004208460915833712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03587677329778671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.008650568313896656 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650568313896656, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650568313896656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20818288624286652, + 0.07428314536809921, + 0.04048111289739609, + 0.07622786611318588, + 0.08799368888139725, + 0.05599295720458031, + 0.3139529526233673, + 0.04818703234195709, + 0.060636602342128754, + 0.03406178206205368 + ], + "credit_map": [ + 0.026276271790266037, + 0.02385888434946537, + 0.013361903838813305, + 0.02362161874771118, + 0.03182649612426758, + 0.024968018755316734, + 0.04867081344127655, + 0.01559887733310461, + 0.011826752685010433, + 0.009556337259709835, + 0.006283682305365801, + 0.007896781899034977, + 0.009412342682480812, + 0.010396251454949379, + 0.00980046670883894, + 0.009110529907047749, + 0.005756484344601631, + 0.0028733774088323116, + 0.004460237920284271, + 0.002118278993293643, + 0.003768747206777334, + 0.011763394810259342, + 0.0028938089963048697, + 0.006846781820058823, + 0.008872298523783684, + 0.005578744690865278, + 0.017818525433540344, + 0.007595484144985676, + 0.006230595055967569, + 0.009710234589874744, + 0.007330699823796749, + 0.013091282919049263, + 0.009962869808077812, + 0.009624040685594082, + 0.012573466636240482, + 0.009945254772901535, + 0.008630265481770039, + 0.02029302343726158, + 0.007757219020277262, + 0.00920755323022604, + 0.006747695617377758, + 0.007408573757857084, + 0.006105409935116768, + 0.003593391738831997, + 0.00477589201182127, + 0.017208555713295937, + 0.003993259742856026, + 0.00616017309948802, + 0.0505777969956398, + 0.11955445259809494, + 0.03516273945569992, + 0.011279811151325703, + 0.0337764210999012, + 0.03820598125457764, + 0.008970629423856735, + 0.016425129026174545, + 0.004091149196028709, + 0.006712491158396006, + 0.005755553022027016, + 0.005602159071713686, + 0.005954717751592398, + 0.008798723109066486, + 0.004740857519209385, + 0.006531380582600832, + 0.004911050200462341, + 0.012048978358507156, + 0.008491398766636848, + 0.006428940687328577, + 0.008089277893304825, + 0.007618323899805546, + 0.006138338707387447, + 0.0069102942943573, + 0.0017801537178456783, + 0.003960443660616875, + 0.0038979463279247284, + 0.0052948324009776115, + 0.003502815030515194, + 0.00532506313174963, + 0.0066124736331403255, + 0.003688053460791707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.008650568313896656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.0013366345083341002 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013366345083341002, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013366345083341002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15040478110313416, + 0.12079504877328873, + 0.08579611778259277, + 0.11404112726449966, + 0.113505519926548, + 0.039257559925317764, + 0.21728797256946564, + 0.05904493108391762, + 0.058492474257946014, + 0.041374485939741135 + ], + "credit_map": [ + 0.02038564160466194, + 0.01607915572822094, + 0.01385731901973486, + 0.01533451210707426, + 0.026287727057933807, + 0.019557705149054527, + 0.023680856451392174, + 0.015221849083900452, + 0.023802654817700386, + 0.018411317840218544, + 0.010884889401495457, + 0.009357915259897709, + 0.018331054598093033, + 0.009837586432695389, + 0.012348305433988571, + 0.017821311950683594, + 0.010353648103773594, + 0.005906012374907732, + 0.012214701622724533, + 0.005293852649629116, + 0.006763961631804705, + 0.021390674635767937, + 0.006909974385052919, + 0.016963282600045204, + 0.017288776114583015, + 0.014484815299510956, + 0.018544819205999374, + 0.01080613024532795, + 0.011563348583877087, + 0.018604053184390068, + 0.007484285160899162, + 0.015264887362718582, + 0.01220740843564272, + 0.010974915698170662, + 0.019249584525823593, + 0.009823910892009735, + 0.010957257822155952, + 0.033699486404657364, + 0.006141074933111668, + 0.01045187283307314, + 0.0052510639652609825, + 0.004472771659493446, + 0.005631832405924797, + 0.003930621314793825, + 0.00319076725281775, + 0.00760855246335268, + 0.003960615489631891, + 0.005211330950260162, + 0.03537965938448906, + 0.08184567838907242, + 0.02262483909726143, + 0.008350799791514874, + 0.023366037756204605, + 0.027061304077506065, + 0.00716755073517561, + 0.011492098681628704, + 0.004313244950026274, + 0.006773678120225668, + 0.00776818348094821, + 0.007716469932347536, + 0.006930231116712093, + 0.008711359463632107, + 0.007630351465195417, + 0.009201406501233578, + 0.004841488786041737, + 0.009557166136801243, + 0.00799118634313345, + 0.007035126443952322, + 0.007730429992079735, + 0.007301024626940489, + 0.0064626578241586685, + 0.007573393639177084, + 0.0025313394144177437, + 0.004540612921118736, + 0.005326968617737293, + 0.005955682601779699, + 0.004888924770057201, + 0.006391990464180708, + 0.006155297625809908, + 0.0055836644023656845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0013366345083341002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.003187933936715126 + ], + "opposite_logprob_tokens": [ + -5.753188133239746 + ], + "logprob_reward": -0.003187933936715126, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003187933936715126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06841223686933517, + 0.0469958633184433, + 0.10278830677270889, + 0.048187658190727234, + 0.03468035161495209, + 0.05135633051395416, + 0.3292965292930603, + 0.10710199922323227, + 0.15860570967197418, + 0.0525750070810318 + ], + "credit_map": [ + 0.008761988021433353, + 0.008222612552344799, + 0.005006310064345598, + 0.006819591857492924, + 0.01384226605296135, + 0.008693549782037735, + 0.007595114875584841, + 0.009470809251070023, + 0.012240266427397728, + 0.007220826577395201, + 0.004217484965920448, + 0.0054508415050804615, + 0.0038136423099786043, + 0.004472054541110992, + 0.0036223598290234804, + 0.00595839275047183, + 0.009056603536009789, + 0.007651573047041893, + 0.01747848466038704, + 0.011156068183481693, + 0.00881971139460802, + 0.021312406286597252, + 0.009506833739578724, + 0.017806634306907654, + 0.004749916028231382, + 0.005251928698271513, + 0.0081888847053051, + 0.005215357523411512, + 0.004922970198094845, + 0.00568342674523592, + 0.005985980853438377, + 0.008189194835722446, + 0.005023086443543434, + 0.004072160925716162, + 0.00554804177954793, + 0.0036880946718156338, + 0.0031064448412507772, + 0.003996453247964382, + 0.0035447499249130487, + 0.005701323971152306, + 0.007187975570559502, + 0.0055190264247357845, + 0.005808396264910698, + 0.004068527836352587, + 0.005459167528897524, + 0.012513041496276855, + 0.005409865640103817, + 0.005390331149101257, + 0.05222242325544357, + 0.12556006014347076, + 0.03260980546474457, + 0.013797248713672161, + 0.035849686712026596, + 0.040723253041505814, + 0.012108703143894672, + 0.0164253618568182, + 0.008311031386256218, + 0.015250854194164276, + 0.014291482046246529, + 0.012758718803524971, + 0.011972431093454361, + 0.014853333123028278, + 0.01327612902969122, + 0.01638801582157612, + 0.013358743861317635, + 0.027398381382226944, + 0.023026257753372192, + 0.016647284850478172, + 0.019320370629429817, + 0.020816469565033913, + 0.022028790786862373, + 0.01600942388176918, + 0.005191317293792963, + 0.0065725697204470634, + 0.00545471953228116, + 0.006523853167891502, + 0.009694529697299004, + 0.006006288342177868, + 0.008398989215493202, + 0.004732746630907059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.003187933936715126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.0013321106089279056 + ], + "opposite_logprob_tokens": [ + -6.6263322830200195 + ], + "logprob_reward": -0.0013321106089279056, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013321106089279056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060754645615816116, + 0.030179357156157494, + 0.048118770122528076, + 0.04569803178310394, + 0.0815771147608757, + 0.05802493914961815, + 0.44963496923446655, + 0.061722926795482635, + 0.10680832713842392, + 0.05748095363378525 + ], + "credit_map": [ + 0.011223064735531807, + 0.007933755405247211, + 0.0062757693231105804, + 0.00621286453679204, + 0.009711162187159061, + 0.006940301973372698, + 0.00716208154335618, + 0.0052956449799239635, + 0.003878935007378459, + 0.003367892699316144, + 0.0029440177604556084, + 0.0031030590180307627, + 0.005195458885282278, + 0.004121758043766022, + 0.004407308530062437, + 0.0031609267462044954, + 0.0068545034155249596, + 0.003429494798183441, + 0.005700964014977217, + 0.005361649673432112, + 0.0034266759175807238, + 0.013413611799478531, + 0.004100934136658907, + 0.0058309342712163925, + 0.007372885476797819, + 0.004057153128087521, + 0.008997444063425064, + 0.005116415675729513, + 0.002929728478193283, + 0.00685094902291894, + 0.003198804333806038, + 0.007174651138484478, + 0.010977726429700851, + 0.00789810810238123, + 0.012025848962366581, + 0.007884172722697258, + 0.007431011646986008, + 0.020888520404696465, + 0.006627651862800121, + 0.007844074629247189, + 0.013087088242173195, + 0.006968236528337002, + 0.00732410466298461, + 0.004101949278265238, + 0.0054597314447164536, + 0.009382842108607292, + 0.00440780958160758, + 0.007293176371604204, + 0.07454798370599747, + 0.17254753410816193, + 0.04484730213880539, + 0.016523176804184914, + 0.04889272525906563, + 0.05528426542878151, + 0.01319566834717989, + 0.023796336725354195, + 0.004396992269903421, + 0.007191049866378307, + 0.00953393429517746, + 0.007092336192727089, + 0.00734809460118413, + 0.009528706781566143, + 0.0070071108639240265, + 0.009624703787267208, + 0.008274747990071774, + 0.017745250836014748, + 0.016106566414237022, + 0.01109338365495205, + 0.011354818940162659, + 0.011204003356397152, + 0.010221372358500957, + 0.020808177068829536, + 0.004941802006214857, + 0.007178942207247019, + 0.0064836968667805195, + 0.0070177651941776276, + 0.011737719178199768, + 0.007874640636146069, + 0.007823297753930092, + 0.004423088394105434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.0013321106089279056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot begins from a slight overhead angle, then steadily pulls back, ultimately culminating in a magnificent long shot. The camera initially focuses on a small, meticulously folded paper boat, crafted from an acceptance letter. The thick, off-white surface of the paper reveals faintly blurred, black Song-style calligraphy, slightly dampened by moisture. Clear creases outline its small, yet resolute outline. Steadily and slowly, it drifts along the gentle ripples of the water. Under the warm, soft golden light of the setting sun, the boat is gilded with a dreamy golden edge. As the camera pulls out, the vast expanse of water within it is gradually revealed. The tranquil, clear, and deep waters reflect the vibrant sunset glow, tinged in warm orange and soft pink-purple. The ripples shimmer where the water meets the sky. Ultimately, the tiny paper boat appears both small and powerful within a vast, vibrant world. The warm tones of the entire scene create a poetic atmosphere of tranquility, hope, and a touch of loneliness.", + "seed": 20087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_420_seed_20087_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": -0.02091991058550775, + "vlm_logprob": -0.02091991058550775, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.02091991058550775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.07274677604436874 + ], + "opposite_logprob_tokens": [ + -2.697746753692627 + ], + "logprob_reward": -0.07274677604436874, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07274677604436874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06945253908634186, + 0.04290158301591873, + 0.596602737903595, + 0.08525967597961426, + 0.055473022162914276, + 0.02121398039162159, + 0.0385180339217186, + 0.04060874134302139, + 0.020192187279462814, + 0.029777532443404198 + ], + "credit_map": [ + 0.012129314243793488, + 0.006692843046039343, + 0.007778518833220005, + 0.005269084591418505, + 0.009786801412701607, + 0.007930771447718143, + 0.008014606311917305, + 0.011850590817630291, + 0.003936754073947668, + 0.006132547277957201, + 0.004895378835499287, + 0.005813658237457275, + 0.008246597833931446, + 0.004477838985621929, + 0.005205770954489708, + 0.00419303635135293, + 0.08536019176244736, + 0.05057575926184654, + 0.05329658091068268, + 0.03612899407744408, + 0.028296275064349174, + 0.152124285697937, + 0.028823407366871834, + 0.1619972288608551, + 0.008511709980666637, + 0.015791909769177437, + 0.012725519947707653, + 0.009831053204834461, + 0.00950415339320898, + 0.012505732476711273, + 0.0056446390226483345, + 0.010744951665401459, + 0.005402406677603722, + 0.007014609407633543, + 0.00760809239000082, + 0.011550286784768105, + 0.00484468461945653, + 0.006634619552642107, + 0.007140996400266886, + 0.0052773188799619675, + 0.0013959137722849846, + 0.002624220447614789, + 0.003362140618264675, + 0.0032774966675788164, + 0.002580881817266345, + 0.002079284517094493, + 0.002489410573616624, + 0.003404628951102495, + 0.004197616595774889, + 0.0037447866052389145, + 0.006851617246866226, + 0.003973303362727165, + 0.00343933142721653, + 0.006118113175034523, + 0.0034537215251475573, + 0.006739542819559574, + 0.004515237640589476, + 0.004927980713546276, + 0.0064860456623137, + 0.005193689838051796, + 0.004757747985422611, + 0.004496182780712843, + 0.005092304199934006, + 0.005139548797160387, + 0.0018517832504585385, + 0.0018050724174827337, + 0.004385278094559908, + 0.002233248669654131, + 0.0019039796898141503, + 0.002224483760073781, + 0.002033619210124016, + 0.003754720790311694, + 0.0031112050637602806, + 0.0042363060638308525, + 0.005590696819126606, + 0.003647289704531431, + 0.0032910662703216076, + 0.003226630389690399, + 0.0033036747481673956, + 0.0033706622198224068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.07274677604436874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.01820322498679161 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.01820322498679161, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820322498679161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06088046729564667, + 0.04752714931964874, + 0.48656612634658813, + 0.09250009804964066, + 0.08533650636672974, + 0.026365667581558228, + 0.03414659947156906, + 0.06607618927955627, + 0.023820487782359123, + 0.07678072154521942 + ], + "credit_map": [ + 0.008604531176388264, + 0.0063505591824650764, + 0.008584658615291119, + 0.006626766175031662, + 0.0074768769554793835, + 0.005897126160562038, + 0.0069679273292422295, + 0.010372022166848183, + 0.0073696039617061615, + 0.0077827745117247105, + 0.0063148257322609425, + 0.005246585234999657, + 0.00511785876005888, + 0.004908769391477108, + 0.0046347277238965034, + 0.006152003537863493, + 0.07190775871276855, + 0.03910776600241661, + 0.04879557341337204, + 0.029265832155942917, + 0.01792220026254654, + 0.12917008996009827, + 0.021545330062508583, + 0.12885156273841858, + 0.007790347095578909, + 0.01116635650396347, + 0.014967653900384903, + 0.00888820644468069, + 0.012033038772642612, + 0.009756550192832947, + 0.006891435012221336, + 0.0210065059363842, + 0.006314837373793125, + 0.0056793708354234695, + 0.008912941440939903, + 0.008792671374976635, + 0.006098968908190727, + 0.022383099421858788, + 0.006235746666789055, + 0.020918872207403183, + 0.003652558894827962, + 0.002507241442799568, + 0.004222317598760128, + 0.0028944469522684813, + 0.0030418771784752607, + 0.0038641903083771467, + 0.0028031545225530863, + 0.00337987975217402, + 0.004253902472555637, + 0.0030452676583081484, + 0.00597204128280282, + 0.0036767919082194567, + 0.0025063869543373585, + 0.005751150194555521, + 0.003727582050487399, + 0.005213476251810789, + 0.006869545206427574, + 0.006687634624540806, + 0.00822482444345951, + 0.006566949188709259, + 0.007341208402067423, + 0.013251279480755329, + 0.006189951207488775, + 0.010944797657430172, + 0.0031325765885412693, + 0.002626413246616721, + 0.00358826806768775, + 0.0028219178784638643, + 0.003032178618013859, + 0.002851772354915738, + 0.0024726251140236855, + 0.003294734051451087, + 0.005076149478554726, + 0.009114006534218788, + 0.014665204100310802, + 0.011166641488671303, + 0.0067391954362392426, + 0.008376519195735455, + 0.010301507078111172, + 0.011341498233377934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.01820322498679161 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.0019396792631596327 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.0019396792631596327, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019396792631596327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09861700236797333, + 0.08169186860322952, + 0.24328051507472992, + 0.1261972188949585, + 0.14541485905647278, + 0.0482068695127964, + 0.08006295561790466, + 0.07375825941562653, + 0.032680146396160126, + 0.07009033113718033 + ], + "credit_map": [ + 0.017109457403421402, + 0.010984115302562714, + 0.016300775110721588, + 0.013543029315769672, + 0.01165030151605606, + 0.006351162679493427, + 0.010957250371575356, + 0.011720898561179638, + 0.008841268718242645, + 0.00621392484754324, + 0.007936857640743256, + 0.006909853313118219, + 0.011023374274373055, + 0.019120197743177414, + 0.0061098444275557995, + 0.015536538325250149, + 0.037344519048929214, + 0.02683797851204872, + 0.03595683351159096, + 0.02847367897629738, + 0.023779571056365967, + 0.03991497680544853, + 0.020666072145104408, + 0.030306881293654442, + 0.010147282853722572, + 0.015458651818335056, + 0.015849754214286804, + 0.016531113535165787, + 0.021357309073209763, + 0.01780620962381363, + 0.01057091448456049, + 0.018475977703928947, + 0.012262866832315922, + 0.009788138791918755, + 0.015023344196379185, + 0.010302064009010792, + 0.010875655338168144, + 0.03873619809746742, + 0.00930237676948309, + 0.03912420570850372, + 0.006790450774133205, + 0.004128080327063799, + 0.008818932808935642, + 0.0040450687520205975, + 0.004549078643321991, + 0.008671555668115616, + 0.00443761982023716, + 0.006766078062355518, + 0.009811649098992348, + 0.0064644962549209595, + 0.011958609335124493, + 0.008302969858050346, + 0.005767319817095995, + 0.016345392912626266, + 0.007207752671092749, + 0.014204762876033783, + 0.007279580924659967, + 0.007330880034714937, + 0.009744033217430115, + 0.007622098084539175, + 0.00808109249919653, + 0.01432686485350132, + 0.006542474031448364, + 0.012831226922571659, + 0.003205973422154784, + 0.003934330306947231, + 0.0048946812748909, + 0.003794804448261857, + 0.004599025472998619, + 0.0035428113769739866, + 0.0035758153535425663, + 0.005132703110575676, + 0.005801308900117874, + 0.008265999145805836, + 0.011483520269393921, + 0.008813454769551754, + 0.005764482077211142, + 0.007651943247765303, + 0.009803194552659988, + 0.012506421655416489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.0019396792631596327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.007621851284056902 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.007621851284056902, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621851284056902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04179176315665245, + 0.036370571702718735, + 0.13646568357944489, + 0.06960272043943405, + 0.061160389333963394, + 0.045988503843545914, + 0.05200633406639099, + 0.21779218316078186, + 0.0923612117767334, + 0.24646064639091492 + ], + "credit_map": [ + 0.0049335346557199955, + 0.004834513645619154, + 0.006687932647764683, + 0.00535358814522624, + 0.005989391822367907, + 0.0033478739205747843, + 0.00566629646345973, + 0.004978632554411888, + 0.003192316507920623, + 0.0021228520199656487, + 0.004084079526364803, + 0.003307466162368655, + 0.0028883994091302156, + 0.00827530026435852, + 0.0032112442422658205, + 0.009288913570344448, + 0.010481211356818676, + 0.010659200139343739, + 0.025887709110975266, + 0.02040972001850605, + 0.006077515427023172, + 0.026655694469809532, + 0.014433528296649456, + 0.02186110056936741, + 0.0061821830458939075, + 0.005585220642387867, + 0.011451552622020245, + 0.0038362944032996893, + 0.003573812311515212, + 0.012197863310575485, + 0.003373676910996437, + 0.023402120918035507, + 0.004600317217409611, + 0.0037480047903954983, + 0.008430790156126022, + 0.007696598768234253, + 0.004630398936569691, + 0.014839676208794117, + 0.005324161145836115, + 0.011890438385307789, + 0.005419434048235416, + 0.004452603869140148, + 0.007349818013608456, + 0.004957884084433317, + 0.007019548211246729, + 0.005700577981770039, + 0.0051644109189510345, + 0.0059242271818220615, + 0.00469819363206625, + 0.004535162355750799, + 0.013098637573421001, + 0.0062940060161054134, + 0.004085895139724016, + 0.0063078985549509525, + 0.005538627505302429, + 0.007447913754731417, + 0.02350584790110588, + 0.02440270408987999, + 0.032794687896966934, + 0.024477731436491013, + 0.023777514696121216, + 0.03524147346615791, + 0.023601867258548737, + 0.029990369454026222, + 0.010111878626048565, + 0.015113890171051025, + 0.011568225920200348, + 0.011438235640525818, + 0.013092830777168274, + 0.011258297599852085, + 0.009678623639047146, + 0.010099228471517563, + 0.012397797778248787, + 0.022561535239219666, + 0.03820161148905754, + 0.03580927103757858, + 0.026493161916732788, + 0.0355575755238533, + 0.03612867742776871, + 0.039311014115810394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.007621851284056902 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.004088021349161863 + ], + "opposite_logprob_tokens": [ + -5.504087924957275 + ], + "logprob_reward": -0.004088021349161863, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088021349161863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06695248186588287, + 0.03221028298139572, + 0.16819603741168976, + 0.09559793025255203, + 0.12562566995620728, + 0.10396013408899307, + 0.06999260932207108, + 0.1533229500055313, + 0.07291288673877716, + 0.11122903227806091 + ], + "credit_map": [ + 0.00990024022758007, + 0.007173208519816399, + 0.0059335497207939625, + 0.005706500262022018, + 0.006456120405346155, + 0.007714074570685625, + 0.009709025733172894, + 0.014359759166836739, + 0.004467058461159468, + 0.003784745465964079, + 0.004923587199300528, + 0.003761629108339548, + 0.003590190317481756, + 0.005570234265178442, + 0.0025480156764388084, + 0.0035648259799927473, + 0.021204980090260506, + 0.012041797861456871, + 0.025050457566976547, + 0.022032519802451134, + 0.01010263804346323, + 0.03522443771362305, + 0.013929310254752636, + 0.02860989049077034, + 0.00905571598559618, + 0.00956631824374199, + 0.017064915969967842, + 0.009937194176018238, + 0.007135818246752024, + 0.015497643500566483, + 0.008217200636863708, + 0.019123122096061707, + 0.0104936258867383, + 0.0064881332218647, + 0.02286120317876339, + 0.013208341784775257, + 0.006627364549785852, + 0.027974365279078484, + 0.007047378458082676, + 0.030925258994102478, + 0.014079098589718342, + 0.01034084614366293, + 0.017774417996406555, + 0.01267281360924244, + 0.012525370344519615, + 0.012144984677433968, + 0.010606527328491211, + 0.013816069811582565, + 0.008257599547505379, + 0.006307350005954504, + 0.012994754128158092, + 0.008468949235975742, + 0.005550193600356579, + 0.010425983928143978, + 0.008407192304730415, + 0.009580588899552822, + 0.01848931796848774, + 0.014908591285347939, + 0.018861951306462288, + 0.014729149639606476, + 0.015069611370563507, + 0.03072894737124443, + 0.014880732633173466, + 0.02565464749932289, + 0.010053965263068676, + 0.009552866220474243, + 0.01104351133108139, + 0.007177349179983139, + 0.008926386944949627, + 0.009103089570999146, + 0.006528228987008333, + 0.01052748505026102, + 0.007887816987931728, + 0.015008024871349335, + 0.023966573178768158, + 0.013619118370115757, + 0.01172206737101078, + 0.010181482881307602, + 0.013444573618471622, + 0.015399370342493057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.004088021349161863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot begins from a slight overhead angle, then steadily pulls back, ultimately culminating in a magnificent long shot. The camera initially focuses on a small, meticulously folded paper boat, crafted from an acceptance letter. The thick, off-white surface of the paper reveals faintly blurred, black Song-style calligraphy, slightly dampened by moisture. Clear creases outline its small, yet resolute outline. Steadily and slowly, it drifts along the gentle ripples of the water. Under the warm, soft golden light of the setting sun, the boat is gilded with a dreamy golden edge. As the camera pulls out, the vast expanse of water within it is gradually revealed. The tranquil, clear, and deep waters reflect the vibrant sunset glow, tinged in warm orange and soft pink-purple. The ripples shimmer where the water meets the sky. Ultimately, the tiny paper boat appears both small and powerful within a vast, vibrant world. The warm tones of the entire scene create a poetic atmosphere of tranquility, hope, and a touch of loneliness.", + "seed": 20088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_13/index_420_seed_20088_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": -0.006084358179941773, + "vlm_logprob": -0.006084358179941773, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.006084358179941773, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.008808683604001999 + ], + "opposite_logprob_tokens": [ + -4.883808612823486 + ], + "logprob_reward": -0.008808683604001999, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008808683604001999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6931827068328857, + 0.03287653625011444, + 0.1102859377861023, + 0.05040502920746803, + 0.016066137701272964, + 0.01033165492117405, + 0.02305310033261776, + 0.022011853754520416, + 0.016968952491879463, + 0.024818062782287598 + ], + "credit_map": [ + 0.04374207183718681, + 0.032457221299409866, + 0.0256496649235487, + 0.24945400655269623, + 0.24019873142242432, + 0.04823235422372818, + 0.03158831223845482, + 0.021860405802726746, + 0.0059465812519192696, + 0.006506407167762518, + 0.0022311408538371325, + 0.0034005739726126194, + 0.004777625203132629, + 0.0037847913336008787, + 0.0030222658533602953, + 0.0032071531750261784, + 0.010640215128660202, + 0.020442727953195572, + 0.010649443604052067, + 0.011088989675045013, + 0.02013477310538292, + 0.010141550563275814, + 0.013712174259126186, + 0.013476075604557991, + 0.005622472148388624, + 0.0054680113680660725, + 0.0074895876459777355, + 0.0050033521838486195, + 0.005914733279496431, + 0.004202657379209995, + 0.006223402451723814, + 0.010480820201337337, + 0.0023732467088848352, + 0.0020472060423344374, + 0.0026043264660984278, + 0.001976744271814823, + 0.002106464235112071, + 0.00130697269923985, + 0.002211484359577298, + 0.001439694664441049, + 0.0013409712119027972, + 0.0013027481036260724, + 0.0014626748161390424, + 0.0016697392566129565, + 0.0011797777842730284, + 0.0008839648799039423, + 0.0015043065650388598, + 0.0009874737588688731, + 0.005079030059278011, + 0.0034786781761795282, + 0.004822883289307356, + 0.002417380455881357, + 0.0030699281487613916, + 0.0014182687737047672, + 0.0009239634964615107, + 0.0018429699121043086, + 0.002476239576935768, + 0.0032464598771184683, + 0.0030974294058978558, + 0.002188122132793069, + 0.003560551442205906, + 0.0026883960235863924, + 0.002327599795535207, + 0.002427058294415474, + 0.002634412609040737, + 0.0019484813092276454, + 0.002320478670299053, + 0.0024810591712594032, + 0.00152162357699126, + 0.0019416650757193565, + 0.0019230229081586003, + 0.0021982104517519474, + 0.002376771531999111, + 0.002936301752924919, + 0.0030852477066218853, + 0.0036306981928646564, + 0.0023663321044296026, + 0.0046342466957867146, + 0.0018755479250103235, + 0.003912918735295534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.008808683604001999 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.005957465618848801 + ], + "opposite_logprob_tokens": [ + -5.13095760345459 + ], + "logprob_reward": -0.005957465618848801, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005957465618848801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24462218582630157, + 0.06688368320465088, + 0.14584767818450928, + 0.06571660935878754, + 0.05518073961138725, + 0.04087819531559944, + 0.07532940804958344, + 0.13180916011333466, + 0.02836788445711136, + 0.14536438882350922 + ], + "credit_map": [ + 0.017197690904140472, + 0.019595155492424965, + 0.009775987826287746, + 0.08219107240438461, + 0.07415987551212311, + 0.01740971766412258, + 0.013371355831623077, + 0.01092134602367878, + 0.00866662710905075, + 0.009181712754070759, + 0.00675489567220211, + 0.006184984464198351, + 0.009755857288837433, + 0.008292763493955135, + 0.009321237914264202, + 0.008725605905056, + 0.012853874824941158, + 0.021959969773888588, + 0.01320208702236414, + 0.019496699795126915, + 0.02343803644180298, + 0.023842131718993187, + 0.0174428541213274, + 0.013612036593258381, + 0.006950336508452892, + 0.009510094299912453, + 0.008051527664065361, + 0.005982519127428532, + 0.012797167524695396, + 0.0081408079713583, + 0.0075225625187158585, + 0.006761596538126469, + 0.008513404056429863, + 0.010943729430437088, + 0.00845556054264307, + 0.005287687294185162, + 0.008115706965327263, + 0.0045813326723873615, + 0.004860764369368553, + 0.004422558005899191, + 0.005033929366618395, + 0.006203295662999153, + 0.005221857223659754, + 0.006168094929307699, + 0.005282846745103598, + 0.003951644990593195, + 0.005189585033804178, + 0.003826944390311837, + 0.011323808692395687, + 0.009221135638654232, + 0.015527613461017609, + 0.011730044148862362, + 0.008062245324254036, + 0.010215766727924347, + 0.00416467385366559, + 0.005084121599793434, + 0.015040464699268341, + 0.01863252930343151, + 0.015623152256011963, + 0.019227737560868263, + 0.018629761412739754, + 0.01777666062116623, + 0.013974213041365147, + 0.012904650531709194, + 0.0030466304160654545, + 0.004094248171895742, + 0.003664211602881551, + 0.00467109028249979, + 0.003280175384134054, + 0.0032583822030574083, + 0.002918815240263939, + 0.0034343330189585686, + 0.01369234174489975, + 0.01756437122821808, + 0.01741018146276474, + 0.019713930785655975, + 0.016102738678455353, + 0.02915920503437519, + 0.010591235011816025, + 0.021130386739969254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.005957465618848801 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.0021975908894091845 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.0021975908894091845, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021975908894091845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.106264628469944, + 0.11101923137903214, + 0.16840286552906036, + 0.07289721071720123, + 0.06175185739994049, + 0.04488968476653099, + 0.06095648184418678, + 0.0780869871377945, + 0.09491955488920212, + 0.20081152021884918 + ], + "credit_map": [ + 0.010840333066880703, + 0.012827313505113125, + 0.010733334347605705, + 0.01614578440785408, + 0.012969289906322956, + 0.015446365810930729, + 0.010821748524904251, + 0.0164804570376873, + 0.012492929585278034, + 0.01373932883143425, + 0.013847467489540577, + 0.013514038175344467, + 0.014753943309187889, + 0.016238253563642502, + 0.012546250596642494, + 0.013887024484574795, + 0.015727432444691658, + 0.029476825147867203, + 0.01542613934725523, + 0.016875579953193665, + 0.026620013639330864, + 0.023578330874443054, + 0.022706247866153717, + 0.01799229346215725, + 0.010721071623265743, + 0.00973470788449049, + 0.007724250201135874, + 0.005275801755487919, + 0.009747711941599846, + 0.006722799502313137, + 0.007388248573988676, + 0.015582617372274399, + 0.009516000747680664, + 0.01315361075103283, + 0.009876208379864693, + 0.005825788248330355, + 0.007122453302145004, + 0.005435389932245016, + 0.0051827686838805676, + 0.005639639683067799, + 0.006277170032262802, + 0.008611038327217102, + 0.006025081500411034, + 0.005798681173473597, + 0.006046107970178127, + 0.004627143032848835, + 0.004761484917253256, + 0.00274297664873302, + 0.00742421904578805, + 0.009771352633833885, + 0.01223151944577694, + 0.009152257815003395, + 0.008315455168485641, + 0.005892450921237469, + 0.004236110486090183, + 0.003933115396648645, + 0.012326287105679512, + 0.01227419264614582, + 0.01079220324754715, + 0.006451157853007317, + 0.012175354175269604, + 0.010162497870624065, + 0.00642738863825798, + 0.007477902807295322, + 0.011689546518027782, + 0.013128752820193768, + 0.009646219201385975, + 0.012003877200186253, + 0.014921520836651325, + 0.01030094362795353, + 0.011700370348989964, + 0.011528325267136097, + 0.021693706512451172, + 0.02406541258096695, + 0.025169115513563156, + 0.02492649108171463, + 0.022296082228422165, + 0.04167558625340462, + 0.01397981122136116, + 0.027005299925804138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.0021975908894091845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.006730505730956793 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730505730956793, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730505730956793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45946499705314636, + 0.04948188737034798, + 0.11659751832485199, + 0.02978566661477089, + 0.07207632809877396, + 0.03864603862166405, + 0.04499126225709915, + 0.07023181021213531, + 0.03889808803796768, + 0.07982637733221054 + ], + "credit_map": [ + 0.028950214385986328, + 0.022795042023062706, + 0.016196493059396744, + 0.1654435396194458, + 0.1602153182029724, + 0.03250521793961525, + 0.019531795755028725, + 0.013827412389218807, + 0.004524226766079664, + 0.006438948214054108, + 0.004659304860979319, + 0.005312526598572731, + 0.010480961762368679, + 0.0048384773544967175, + 0.0061221616342663765, + 0.007105278316885233, + 0.009438241831958294, + 0.01813577115535736, + 0.010711736977100372, + 0.013599785976111889, + 0.02787853218615055, + 0.012017279863357544, + 0.015576726756989956, + 0.009239444509148598, + 0.006272941827774048, + 0.004193081054836512, + 0.0048622856847941875, + 0.0027815857902169228, + 0.0033999381121248007, + 0.002047126181423664, + 0.00300240283831954, + 0.0032263067550957203, + 0.00990913063287735, + 0.009474982507526875, + 0.006755698472261429, + 0.008544518612325191, + 0.016731401905417442, + 0.006044760812073946, + 0.007746484596282244, + 0.006869347300380468, + 0.00544655742123723, + 0.006608121562749147, + 0.006221717223525047, + 0.003924318589270115, + 0.0042432378977537155, + 0.0031877874862402678, + 0.00433812290430069, + 0.004676175303757191, + 0.006659775041043758, + 0.006295513827353716, + 0.007449939846992493, + 0.006882388610392809, + 0.00572200259193778, + 0.004827110096812248, + 0.002631397219374776, + 0.004523135256022215, + 0.010309363715350628, + 0.009359716437757015, + 0.009348293766379356, + 0.008928600698709488, + 0.010255339555442333, + 0.008215075358748436, + 0.006939481012523174, + 0.006875942461192608, + 0.006342316512018442, + 0.005145277827978134, + 0.006627278868108988, + 0.005053607281297445, + 0.004716904833912849, + 0.0041436306200921535, + 0.003332759253680706, + 0.0035363154020160437, + 0.006603196728974581, + 0.01042876671999693, + 0.009696499444544315, + 0.015034640207886696, + 0.010568534955382347, + 0.011547003872692585, + 0.006738464348018169, + 0.009209275245666504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.006730505730956793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small paper boat crafted from an acceptance letter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the paper boat have a thick off-white surface with faintly blurred black Song-style calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the paper boat gilded with a dreamy golden edge under warm soft golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slight overhead close-up and steadily pull back to a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the water environment tranquil, clear, deep, and reflecting warm orange and soft pink-purple sunset 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": [ + -0.00672754505649209 + ], + "opposite_logprob_tokens": [ + -5.006727695465088 + ], + "logprob_reward": -0.00672754505649209, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00672754505649209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.632347047328949, + 0.02579251118004322, + 0.012860247865319252, + 0.012677479535341263, + 0.015874598175287247, + 0.012530714273452759, + 0.04535077139735222, + 0.06431735306978226, + 0.04180876538157463, + 0.13644057512283325 + ], + "credit_map": [ + 0.03947422653436661, + 0.029157225042581558, + 0.023501163348555565, + 0.22587265074253082, + 0.22258704900741577, + 0.043873537331819534, + 0.027338774874806404, + 0.020542379468679428, + 0.0030645441729575396, + 0.0032372381538152695, + 0.0029482347890734673, + 0.00211539538577199, + 0.004052620846778154, + 0.0036393352784216404, + 0.0030516651459038258, + 0.003683476708829403, + 0.0008719395846128464, + 0.0017837703926488757, + 0.0016606072895228863, + 0.0016107636038213968, + 0.0020863120444118977, + 0.0015864348970353603, + 0.0017069936729967594, + 0.001553426613099873, + 0.0014855420449748635, + 0.001586021389812231, + 0.0016654651844874024, + 0.0013254729565232992, + 0.002295912243425846, + 0.0013977722264826298, + 0.001185391447506845, + 0.0017359014600515366, + 0.0029864972457289696, + 0.00220789504237473, + 0.002537876134738326, + 0.0018779647070914507, + 0.002121787518262863, + 0.00143423059489578, + 0.0014023585245013237, + 0.0013059882912784815, + 0.0016810254892334342, + 0.002086218213662505, + 0.0018652486614882946, + 0.0017969079781323671, + 0.0014754440635442734, + 0.0012591074919328094, + 0.0016014693537726998, + 0.0007652936037629843, + 0.00564338406547904, + 0.009704207070171833, + 0.007250318769365549, + 0.005332732107490301, + 0.00584192480891943, + 0.004078035708516836, + 0.002305320929735899, + 0.005194853059947491, + 0.010600769892334938, + 0.008703948929905891, + 0.011570434086024761, + 0.0064031039364635944, + 0.007856857031583786, + 0.006959659047424793, + 0.005488417111337185, + 0.006734167691320181, + 0.005402355454862118, + 0.005243306048214436, + 0.005090672522783279, + 0.006247390527278185, + 0.00484779616817832, + 0.0043208785355091095, + 0.004327358677983284, + 0.006329005118459463, + 0.010941965505480766, + 0.014133973978459835, + 0.016496814787387848, + 0.01596972905099392, + 0.011212469078600407, + 0.026582350954413414, + 0.01097745168954134, + 0.030125828459858894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.00672754505649209 + } + ], + "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/0000155/group_14/index_tensor([7196, 7196])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_14/index_tensor([7196, 7196])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dbff622c4aecbe82e16a92b8edec628705fa4481 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_14/index_tensor([7196, 7196])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In one unbroken shot, a dynamic tracking camera fluidly follows a pair of high-top roller skates as they perform a frantic dance across an infinite, canvas-like void. The skates themselves are rendered in explosive cracks of vibrant color, their smooth, chameleon-like surfaces shifting between neon fuchsia, electric cyan, and acid yellow with each movement, capturing unseen light with sharp, specular highlights. They swirl and spin with chaotic precision, one moment gliding smoothly backward, the next executing sharp, aggressive turns that slice into darkness. With each turn and spin, the skates leave behind thick, vibrant, board-like marks of paint in colors that directly correspond to their movements—fiery orange streaks from sudden acceleration, deep purple splashes from stomping landings. The camera remains in close contact with their movements, capturing their chaotic ballet of pivots and turns as they continually construct intricate three-dimensional action paintings within the desolate, matte void, creating an atmosphere of pure, unfettered kinetic energy and abstract expression.", + "seed": 29492, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_14/index_7196_seed_29492_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": -0.02649210626259446, + "vlm_logprob": -0.02649210626259446, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.02649210626259446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0077259535901248455 + ], + "opposite_logprob_tokens": [ + -5.507725715637207 + ], + "logprob_reward": -0.0077259535901248455, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0077259535901248455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3248050808906555, + 0.05341923236846924, + 0.03688930347561836, + 0.03964989259839058, + 0.05652681738138199, + 0.03590770065784454, + 0.10331141203641891, + 0.10783194750547409, + 0.16597265005111694, + 0.07568597048521042 + ], + "credit_map": [ + 0.029106972739100456, + 0.027245251461863518, + 0.045753639191389084, + 0.09378673881292343, + 0.016243351623415947, + 0.045767612755298615, + 0.03987869247794151, + 0.027022814378142357, + 0.00689323153346777, + 0.00561210373416543, + 0.007924459874629974, + 0.013069878332316875, + 0.005645939614623785, + 0.0036605270579457283, + 0.005409257020801306, + 0.00520382821559906, + 0.005066787824034691, + 0.0044636414386332035, + 0.005097486078739166, + 0.0031855711713433266, + 0.0038074306212365627, + 0.005309512373059988, + 0.004258049186319113, + 0.005700820591300726, + 0.005302491132169962, + 0.005194403696805239, + 0.004042778629809618, + 0.004565115086734295, + 0.0068086180835962296, + 0.004877079278230667, + 0.00511737959459424, + 0.003742021508514881, + 0.008864880539476871, + 0.00822409987449646, + 0.006897533778101206, + 0.007590269669890404, + 0.007222537416964769, + 0.006692094728350639, + 0.004972741473466158, + 0.006062655244022608, + 0.005662307608872652, + 0.004505069926381111, + 0.004933045245707035, + 0.0043715424835681915, + 0.003920759540051222, + 0.005966431926935911, + 0.0039136833511292934, + 0.0026348591782152653, + 0.01764283888041973, + 0.011210450902581215, + 0.013764713890850544, + 0.0076738023199141026, + 0.008095807395875454, + 0.016630100086331367, + 0.012786316685378551, + 0.015507380478084087, + 0.009871447458863258, + 0.007776047568768263, + 0.009254046715795994, + 0.009109091944992542, + 0.010266225785017014, + 0.02992112562060356, + 0.009762747213244438, + 0.021871216595172882, + 0.027027443051338196, + 0.04115261137485504, + 0.013531649485230446, + 0.011220033280551434, + 0.029683683067560196, + 0.017355244606733322, + 0.010396069847047329, + 0.015605907887220383, + 0.011694492772221565, + 0.004980912897735834, + 0.008669855073094368, + 0.020119355991482735, + 0.007607247680425644, + 0.00748821534216404, + 0.009849107824265957, + 0.0052767787128686905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0077259535901248455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002511444501578808 + ], + "opposite_logprob_tokens": [ + -6.002511501312256 + ], + "logprob_reward": -0.002511444501578808, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002511444501578808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2788197100162506, + 0.22298286855220795, + 0.04497886076569557, + 0.04299295321106911, + 0.06982415914535522, + 0.03703109547495842, + 0.08286441117525101, + 0.05484636500477791, + 0.08263067156076431, + 0.08302900195121765 + ], + "credit_map": [ + 0.040647659450769424, + 0.02250875160098076, + 0.060525812208652496, + 0.038943100720644, + 0.016993045806884766, + 0.04700743034482002, + 0.023761926218867302, + 0.028431955724954605, + 0.03652065247297287, + 0.01404222846031189, + 0.0277904961258173, + 0.08003757894039154, + 0.01081414707005024, + 0.010263433679938316, + 0.028173163533210754, + 0.01534112449735403, + 0.005228893831372261, + 0.005157840438187122, + 0.008599996566772461, + 0.005546099040657282, + 0.004671724513173103, + 0.005316590424627066, + 0.004554414190351963, + 0.005903297103941441, + 0.006533870939165354, + 0.006475017871707678, + 0.003522278042510152, + 0.0048766965046525, + 0.008236437104642391, + 0.005231386516243219, + 0.004361281171441078, + 0.003755975514650345, + 0.011637859046459198, + 0.006594469770789146, + 0.008575202897191048, + 0.009731107391417027, + 0.007176760118454695, + 0.009497701190412045, + 0.006917511112987995, + 0.00969353411346674, + 0.006366881541907787, + 0.005365878343582153, + 0.003009908366948366, + 0.003498593345284462, + 0.004987908061593771, + 0.006755535025149584, + 0.00413211015984416, + 0.0029142762068659067, + 0.01550484262406826, + 0.005625829100608826, + 0.012883510440587997, + 0.005040932912379503, + 0.006903460714966059, + 0.01949368230998516, + 0.008108575828373432, + 0.009303560480475426, + 0.005903629120439291, + 0.005124358460307121, + 0.006552920211106539, + 0.008519009687006474, + 0.004566579591482878, + 0.009270692244172096, + 0.005834001116454601, + 0.009075162000954151, + 0.013994001783430576, + 0.01046899612993002, + 0.007766009774059057, + 0.0057082464918494225, + 0.009810633957386017, + 0.010152989067137241, + 0.006655077915638685, + 0.018074702471494675, + 0.012198342941701412, + 0.007615880575031042, + 0.021999726071953773, + 0.008914506994187832, + 0.00887184776365757, + 0.007002930622547865, + 0.0077087669633328915, + 0.00871698185801506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.002511444501578808 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463957479223609 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.00463957479223609, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463957479223609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27951180934906006, + 0.1240258514881134, + 0.16378426551818848, + 0.04993082955479622, + 0.022086000069975853, + 0.08064766973257065, + 0.05384673550724983, + 0.0848543643951416, + 0.0807155966758728, + 0.06059690937399864 + ], + "credit_map": [ + 0.01964273490011692, + 0.029489243403077126, + 0.025075964629650116, + 0.1028834730386734, + 0.013663262128829956, + 0.033107832074165344, + 0.03664819151163101, + 0.019001105800271034, + 0.019334984943270683, + 0.005316717550158501, + 0.0183674618601799, + 0.04525258019566536, + 0.006356962025165558, + 0.006911035161465406, + 0.014960736967623234, + 0.007525362074375153, + 0.018380867317318916, + 0.015146155841648579, + 0.039242230355739594, + 0.020716669037938118, + 0.012193435803055763, + 0.022670956328511238, + 0.01383216679096222, + 0.021601779386401176, + 0.006301813758909702, + 0.006234288215637207, + 0.00708893733099103, + 0.00883475411683321, + 0.005276810377836227, + 0.0057491762563586235, + 0.005902159959077835, + 0.004542885813862085, + 0.0034674759954214096, + 0.002964748302474618, + 0.0024960297159850597, + 0.0023760185576975346, + 0.003923342097550631, + 0.0030969923827797174, + 0.0016534331953153014, + 0.0021079587750136852, + 0.01411584671586752, + 0.009968497790396214, + 0.014732136391103268, + 0.01081022247672081, + 0.006531939841806889, + 0.009307857602834702, + 0.00700982753187418, + 0.008171334862709045, + 0.006822243332862854, + 0.005839130841195583, + 0.009127170778810978, + 0.00406211894005537, + 0.005758057348430157, + 0.006217873655259609, + 0.006890085060149431, + 0.009130052290856838, + 0.006501384079456329, + 0.005637066438794136, + 0.007702309638261795, + 0.007969409227371216, + 0.007936163805425167, + 0.022984638810157776, + 0.009983941912651062, + 0.0161394402384758, + 0.010541830211877823, + 0.019188348203897476, + 0.008197891525924206, + 0.007021493744105101, + 0.012594204396009445, + 0.007715938612818718, + 0.006014616694301367, + 0.009441272355616093, + 0.008186525665223598, + 0.007277522701770067, + 0.010660815052688122, + 0.005484566558152437, + 0.008583893068134785, + 0.007241986691951752, + 0.007367722224444151, + 0.005793875083327293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00463957479223609 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294054239988327 + ], + "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": [ + 0.3993959426879883, + 0.12290915101766586, + 0.033523764461278915, + 0.02365337871015072, + 0.02820518985390663, + 0.03260942921042442, + 0.044551678001880646, + 0.12325236201286316, + 0.06998611986637115, + 0.12191300094127655 + ], + "credit_map": [ + 0.04450129717588425, + 0.03471013903617859, + 0.0783018171787262, + 0.09070555120706558, + 0.022679511457681656, + 0.0487138107419014, + 0.0420699417591095, + 0.03771386668086052, + 0.021109072491526604, + 0.006118285935372114, + 0.015130494721233845, + 0.046463076025247574, + 0.006180988159030676, + 0.007775077596306801, + 0.01276683434844017, + 0.007365318480879068, + 0.0052035534754395485, + 0.0035193325020372868, + 0.004185376223176718, + 0.0027015586383640766, + 0.004646063782274723, + 0.004023287445306778, + 0.00450352905318141, + 0.004741064738482237, + 0.0037237165961414576, + 0.0034348273184150457, + 0.0028252853080630302, + 0.0023656960111111403, + 0.0035242028534412384, + 0.002512785606086254, + 0.002403263933956623, + 0.002863600617274642, + 0.004288048949092627, + 0.0034177859779447317, + 0.003770938143134117, + 0.004945436958223581, + 0.0029718796722590923, + 0.0031165392138063908, + 0.0025137756019830704, + 0.0031807846389710903, + 0.005579935386776924, + 0.0039795394986867905, + 0.005451641511172056, + 0.0029250802472233772, + 0.004490640480071306, + 0.00485951965674758, + 0.002920031314715743, + 0.0024030401837080717, + 0.00809505209326744, + 0.0055044288747012615, + 0.005535355303436518, + 0.0030569350346922874, + 0.003988404292613268, + 0.009670108556747437, + 0.004817563574761152, + 0.0038838298059999943, + 0.01276096235960722, + 0.01363228540867567, + 0.012289811857044697, + 0.01568596623837948, + 0.012358666397631168, + 0.022381674498319626, + 0.015891851857304573, + 0.018251145258545876, + 0.008736253716051579, + 0.01667279377579689, + 0.005873559508472681, + 0.005597032606601715, + 0.01199211273342371, + 0.007808671798557043, + 0.004821219481527805, + 0.008484479039907455, + 0.016162630170583725, + 0.014629241079092026, + 0.011477279476821423, + 0.009840903803706169, + 0.022418051958084106, + 0.018900901079177856, + 0.01440869178622961, + 0.014075295999646187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.11294054239988327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643016029149294 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643016029149294, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643016029149294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39828377962112427, + 0.08739697933197021, + 0.051454365253448486, + 0.03772798553109169, + 0.03018769808113575, + 0.04884805902838707, + 0.06749050319194794, + 0.1441359966993332, + 0.04436463117599487, + 0.09011001139879227 + ], + "credit_map": [ + 0.04176951199769974, + 0.03482937067747116, + 0.049629613757133484, + 0.10854025930166245, + 0.022696010768413544, + 0.0612279549241066, + 0.04327178746461868, + 0.03631928563117981, + 0.016503233462572098, + 0.00467331800609827, + 0.011601419188082218, + 0.03189615532755852, + 0.004557004198431969, + 0.004905371461063623, + 0.00843964982777834, + 0.004820825532078743, + 0.006079971790313721, + 0.004201904870569706, + 0.011587792076170444, + 0.005694533232599497, + 0.004689788445830345, + 0.006404089275747538, + 0.007978925481438637, + 0.004817361943423748, + 0.006041788030415773, + 0.005002325400710106, + 0.004249198362231255, + 0.004113924223929644, + 0.005974294152110815, + 0.004817029926925898, + 0.00370940868742764, + 0.0038200158160179853, + 0.004262829199433327, + 0.0031487552914768457, + 0.004417816177010536, + 0.0046184784732759, + 0.003508714260533452, + 0.003491629147902131, + 0.0027457433752715588, + 0.003993731923401356, + 0.007951987907290459, + 0.006989429704844952, + 0.006796605885028839, + 0.00500937644392252, + 0.0052589778788387775, + 0.007151730824261904, + 0.006360623519867659, + 0.003329330123960972, + 0.01594870537519455, + 0.005713487043976784, + 0.005525726359337568, + 0.0038652506191283464, + 0.0060827103443443775, + 0.017793947830796242, + 0.006104579195380211, + 0.006456098984926939, + 0.010686589404940605, + 0.011533875949680805, + 0.010287315584719181, + 0.01601482927799225, + 0.017848161980509758, + 0.030805010348558426, + 0.02235456369817257, + 0.02460564486682415, + 0.0067905038595199585, + 0.0067851054482162, + 0.005517640616744757, + 0.003989989869296551, + 0.00587843032553792, + 0.0048327576369047165, + 0.0031213616020977497, + 0.0074488394893705845, + 0.012079747393727303, + 0.009489384479820728, + 0.015430993400514126, + 0.008832471445202827, + 0.01350341271609068, + 0.012354770675301552, + 0.01198915671557188, + 0.006430074572563171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.004643016029149294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one unbroken shot, a dynamic tracking camera fluidly follows a pair of high-top roller skates as they perform a frantic dance across an infinite, canvas-like void. The skates themselves are rendered in explosive cracks of vibrant color, their smooth, chameleon-like surfaces shifting between neon fuchsia, electric cyan, and acid yellow with each movement, capturing unseen light with sharp, specular highlights. They swirl and spin with chaotic precision, one moment gliding smoothly backward, the next executing sharp, aggressive turns that slice into darkness. With each turn and spin, the skates leave behind thick, vibrant, board-like marks of paint in colors that directly correspond to their movements—fiery orange streaks from sudden acceleration, deep purple splashes from stomping landings. The camera remains in close contact with their movements, capturing their chaotic ballet of pivots and turns as they continually construct intricate three-dimensional action paintings within the desolate, matte void, creating an atmosphere of pure, unfettered kinetic energy and abstract expression.", + "seed": 29493, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_14/index_7196_seed_29493_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": -0.021616617497056723, + "vlm_logprob": -0.021616617497056723, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.021616617497056723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020173491910099983 + ], + "opposite_logprob_tokens": [ + -4.1451735496521 + ], + "logprob_reward": -0.020173491910099983, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020173491910099983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13314412534236908, + 0.10592901706695557, + 0.044311508536338806, + 0.14875082671642303, + 0.06224573403596878, + 0.08051586896181107, + 0.18262237310409546, + 0.06715468317270279, + 0.10331408679485321, + 0.0720118060708046 + ], + "credit_map": [ + 0.012839695438742638, + 0.012701514177024364, + 0.015426477417349815, + 0.02519417554140091, + 0.01048852689564228, + 0.02152380906045437, + 0.018724350258708, + 0.016245586797595024, + 0.015599106438457966, + 0.01721578650176525, + 0.009133021347224712, + 0.01153572741895914, + 0.01698894426226616, + 0.01189073733985424, + 0.01547479908913374, + 0.008090894669294357, + 0.006110338494181633, + 0.00718211755156517, + 0.005599659867584705, + 0.005289812106639147, + 0.00660072872415185, + 0.004162447527050972, + 0.004584725946187973, + 0.004781678318977356, + 0.027344126254320145, + 0.025857077911496162, + 0.009562131948769093, + 0.016138438135385513, + 0.018412897363305092, + 0.019978677853941917, + 0.01999756693840027, + 0.011459908448159695, + 0.011008605360984802, + 0.011937892995774746, + 0.004101566504687071, + 0.012119589373469353, + 0.005255283322185278, + 0.0048575433902442455, + 0.007624674588441849, + 0.005340583622455597, + 0.015424602665007114, + 0.006432713475078344, + 0.006157057825475931, + 0.008386733941733837, + 0.01036880910396576, + 0.01933733932673931, + 0.007097934372723103, + 0.00731067918241024, + 0.024163635447621346, + 0.033649470657110214, + 0.020148715004324913, + 0.02198920212686062, + 0.024777086451649666, + 0.02471064031124115, + 0.013382107950747013, + 0.019801514223217964, + 0.008713907562196255, + 0.010925990529358387, + 0.004667578265070915, + 0.009612257592380047, + 0.011714780703186989, + 0.009968696162104607, + 0.0064816949889063835, + 0.005069781560450792, + 0.004779384937137365, + 0.004816704895347357, + 0.011332144029438496, + 0.014339147135615349, + 0.005957702174782753, + 0.020759020000696182, + 0.03382539749145508, + 0.007504593580961227, + 0.021390587091445923, + 0.004007033538073301, + 0.0070289247669279575, + 0.003313021268695593, + 0.003855181625112891, + 0.025754578411579132, + 0.0027717917691916227, + 0.0038906836416572332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.020173491910099983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036366537678986788 + ], + "opposite_logprob_tokens": [ + -5.628636837005615 + ], + "logprob_reward": -0.0036366537678986788, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036366537678986788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14603379368782043, + 0.15623782575130463, + 0.079006627202034, + 0.1419406533241272, + 0.03708367794752121, + 0.06479410082101822, + 0.07462138682603836, + 0.12524232268333435, + 0.1301843225955963, + 0.04485531523823738 + ], + "credit_map": [ + 0.017959237098693848, + 0.019781390205025673, + 0.015151922591030598, + 0.024623243138194084, + 0.013485509902238846, + 0.02215556800365448, + 0.016866842284798622, + 0.016010073944926262, + 0.014867384918034077, + 0.01814504712820053, + 0.0225484948605299, + 0.027083920314908028, + 0.013657758943736553, + 0.018126100301742554, + 0.02017165906727314, + 0.021637458354234695, + 0.009007977321743965, + 0.006627335213124752, + 0.014532094821333885, + 0.009083305485546589, + 0.008609307929873466, + 0.009063812904059887, + 0.009067495353519917, + 0.013015298172831535, + 0.02236967533826828, + 0.021522613242268562, + 0.015586545690894127, + 0.01952567882835865, + 0.016381805762648582, + 0.02297367714345455, + 0.013696899637579918, + 0.009883753024041653, + 0.005400605034083128, + 0.005051053129136562, + 0.0030388240702450275, + 0.005202730651944876, + 0.0060226269997656345, + 0.003209206974133849, + 0.004481903742998838, + 0.00467672711238265, + 0.010821129195392132, + 0.004426060244441032, + 0.008922887966036797, + 0.007076122798025608, + 0.007084765005856752, + 0.011747312732040882, + 0.006427139509469271, + 0.00828868430107832, + 0.009294557385146618, + 0.01382995955646038, + 0.007649414241313934, + 0.009256389923393726, + 0.009979535825550556, + 0.00967530719935894, + 0.006801596377044916, + 0.008134628646075726, + 0.011643034406006336, + 0.011998644098639488, + 0.014761053957045078, + 0.02286667190492153, + 0.018153229728341103, + 0.018193788826465607, + 0.016631094738841057, + 0.010994811542332172, + 0.00610918877646327, + 0.00622661504894495, + 0.015032444149255753, + 0.018197380006313324, + 0.007551954593509436, + 0.02589123509824276, + 0.041775722056627274, + 0.00939977914094925, + 0.006089142523705959, + 0.00406293012201786, + 0.004734313581138849, + 0.004788949619978666, + 0.00407700939103961, + 0.007520108018070459, + 0.004709939938038588, + 0.008872919715940952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0036366537678986788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189597511664033 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.003189597511664033, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189597511664033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1031729206442833, + 0.1608954668045044, + 0.040235210210084915, + 0.06679171323776245, + 0.06776729226112366, + 0.04773900657892227, + 0.11705338209867477, + 0.2694516181945801, + 0.0557529479265213, + 0.07114040851593018 + ], + "credit_map": [ + 0.010997284203767776, + 0.012914217077195644, + 0.01131567731499672, + 0.014653023332357407, + 0.009675418958067894, + 0.013721007853746414, + 0.013220946304500103, + 0.016675347462296486, + 0.014493468217551708, + 0.017575562000274658, + 0.033689383417367935, + 0.02394784428179264, + 0.0143246715888381, + 0.019222375005483627, + 0.01659463159739971, + 0.02104753442108631, + 0.005822375416755676, + 0.0059382254257798195, + 0.005547736305743456, + 0.0052033038809895515, + 0.005685869138687849, + 0.004393361508846283, + 0.004690271336585283, + 0.0029540685936808586, + 0.010751359164714813, + 0.009878204204142094, + 0.00759335421025753, + 0.009338422678411007, + 0.007195629645138979, + 0.009780030697584152, + 0.007591173984110355, + 0.004663538653403521, + 0.011624639853835106, + 0.011368238367140293, + 0.006476524751633406, + 0.013063915073871613, + 0.005834121722728014, + 0.0054259831085801125, + 0.0059271096251904964, + 0.00804676953703165, + 0.008783446624875069, + 0.003889034502208233, + 0.005639313720166683, + 0.007516073528677225, + 0.004758953582495451, + 0.006086492910981178, + 0.004647997207939625, + 0.006417699158191681, + 0.009508057497441769, + 0.013592150062322617, + 0.015409515239298344, + 0.02150522917509079, + 0.015499898232519627, + 0.014545460231602192, + 0.011646748520433903, + 0.015346335247159004, + 0.02669936791062355, + 0.03191377595067024, + 0.03150421753525734, + 0.043612368404865265, + 0.034317098557949066, + 0.032581545412540436, + 0.04079396277666092, + 0.028029287233948708, + 0.004182161297649145, + 0.0041711279191076756, + 0.0054327333346009254, + 0.006234132684767246, + 0.0043349070474505424, + 0.010520260781049728, + 0.01689111441373825, + 0.00398651510477066, + 0.01927907206118107, + 0.004028382711112499, + 0.00754018547013402, + 0.005370515864342451, + 0.004060712177306414, + 0.02177605964243412, + 0.0037427376955747604, + 0.005342749413102865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.003189597511664033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000836730003357 + ], + "opposite_logprob_tokens": [ + -2.6950082778930664 + ], + "logprob_reward": -0.07000836730003357, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000836730003357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07457894831895828, + 0.11082536727190018, + 0.0665985718369484, + 0.19617292284965515, + 0.0968606024980545, + 0.030207965523004532, + 0.047624651342630386, + 0.06665655225515366, + 0.053593557327985764, + 0.25688090920448303 + ], + "credit_map": [ + 0.00936619471758604, + 0.010000173933804035, + 0.009412622079253197, + 0.008286500349640846, + 0.009174411185085773, + 0.011634237132966518, + 0.006345190107822418, + 0.010359616950154305, + 0.012449708767235279, + 0.010062006302177906, + 0.019981293007731438, + 0.016086455434560776, + 0.009595253504812717, + 0.012985844165086746, + 0.015127439051866531, + 0.014537367038428783, + 0.008217797614634037, + 0.005671005696058273, + 0.010039740242064, + 0.009320953860878944, + 0.006934211123734713, + 0.006902772001922131, + 0.011427780613303185, + 0.008084310218691826, + 0.02964097447693348, + 0.025965869426727295, + 0.02857554703950882, + 0.04395752027630806, + 0.013644920662045479, + 0.01519985031336546, + 0.024292662739753723, + 0.01489558257162571, + 0.01489016879349947, + 0.014966816641390324, + 0.01010864693671465, + 0.010350657626986504, + 0.012275158427655697, + 0.01014383602887392, + 0.009969955310225487, + 0.014155357144773006, + 0.005743037909269333, + 0.002785841003060341, + 0.0019717279355973005, + 0.0035732698161154985, + 0.0040872166864573956, + 0.006903170607984066, + 0.002589237177744508, + 0.002554468344897032, + 0.005691112019121647, + 0.010148966684937477, + 0.004576525185257196, + 0.007368108257651329, + 0.006717696785926819, + 0.0056888870894908905, + 0.003490321571007371, + 0.0039430339820683, + 0.006741973105818033, + 0.0055292216129601, + 0.008066285401582718, + 0.008035186678171158, + 0.008388402871787548, + 0.008744899183511734, + 0.011421329341828823, + 0.009729249402880669, + 0.0038437217008322477, + 0.00351010262966156, + 0.005010535940527916, + 0.007923070341348648, + 0.004251325037330389, + 0.009730491787195206, + 0.015180287882685661, + 0.004144025035202503, + 0.08091351389884949, + 0.014125693589448929, + 0.01931932382285595, + 0.01020576898008585, + 0.011084316298365593, + 0.09929515421390533, + 0.008599000051617622, + 0.013338126242160797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.07000836730003357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074976995587349 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011074976995587349, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074976995587349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1243286281824112, + 0.08585149049758911, + 0.045050449669361115, + 0.03831662982702255, + 0.02502332627773285, + 0.031168073415756226, + 0.04123344644904137, + 0.04942736402153969, + 0.04669616371393204, + 0.5129044651985168 + ], + "credit_map": [ + 0.011537568643689156, + 0.012612167745828629, + 0.015244919806718826, + 0.02016262523829937, + 0.012147861532866955, + 0.017856499180197716, + 0.017814110964536667, + 0.016952862963080406, + 0.011136027984321117, + 0.013053767383098602, + 0.01059835683554411, + 0.010573837906122208, + 0.010997220873832703, + 0.009837196208536625, + 0.011310182511806488, + 0.008344890549778938, + 0.0055204289965331554, + 0.004512206185609102, + 0.00492530083283782, + 0.007493764627724886, + 0.0038915073964744806, + 0.004485694691538811, + 0.008850717917084694, + 0.0053708236664533615, + 0.006148122251033783, + 0.005451119504868984, + 0.005052882246673107, + 0.007695441134274006, + 0.002369414083659649, + 0.003554640570655465, + 0.004663058556616306, + 0.0033819477539509535, + 0.004294422455132008, + 0.004319753497838974, + 0.0024381980765610933, + 0.003985368646681309, + 0.0031570293940603733, + 0.002106365514919162, + 0.002036891644820571, + 0.0026852963492274284, + 0.0054170433431863785, + 0.0021780910901725292, + 0.0024621221236884594, + 0.0038122383411973715, + 0.004126413259655237, + 0.007221694570034742, + 0.002611675299704075, + 0.003338794456794858, + 0.004483157303184271, + 0.01015614066272974, + 0.0034352014772593975, + 0.005847030319273472, + 0.005218448583036661, + 0.00513516366481781, + 0.003378933062776923, + 0.003579367185011506, + 0.005194965749979019, + 0.005054146517068148, + 0.004923091270029545, + 0.005810936447232962, + 0.006862020120024681, + 0.006920831277966499, + 0.0077659375965595245, + 0.006895429454743862, + 0.003501859726384282, + 0.004240093287080526, + 0.005244810599833727, + 0.007425607647746801, + 0.0028805548790842295, + 0.008482624776661396, + 0.011941722594201565, + 0.0029788874089717865, + 0.1571676880121231, + 0.02834671549499035, + 0.03968820348381996, + 0.02081281505525112, + 0.02191563881933689, + 0.195870041847229, + 0.018849803134799004, + 0.030253473669290543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.011074976995587349 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one unbroken shot, a dynamic tracking camera fluidly follows a pair of high-top roller skates as they perform a frantic dance across an infinite, canvas-like void. The skates themselves are rendered in explosive cracks of vibrant color, their smooth, chameleon-like surfaces shifting between neon fuchsia, electric cyan, and acid yellow with each movement, capturing unseen light with sharp, specular highlights. They swirl and spin with chaotic precision, one moment gliding smoothly backward, the next executing sharp, aggressive turns that slice into darkness. With each turn and spin, the skates leave behind thick, vibrant, board-like marks of paint in colors that directly correspond to their movements—fiery orange streaks from sudden acceleration, deep purple splashes from stomping landings. The camera remains in close contact with their movements, capturing their chaotic ballet of pivots and turns as they continually construct intricate three-dimensional action paintings within the desolate, matte void, creating an atmosphere of pure, unfettered kinetic energy and abstract expression.", + "seed": 29494, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_14/index_7196_seed_29494_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": -0.03814115822315216, + "vlm_logprob": -0.03814115822315216, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03814115822315216, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012999314814805984 + ], + "opposite_logprob_tokens": [ + -4.637999534606934 + ], + "logprob_reward": -0.012999314814805984, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012999314814805984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05946403741836548, + 0.0398966483771801, + 0.062421765178442, + 0.2052600234746933, + 0.02842828445136547, + 0.15558196604251862, + 0.04614147171378136, + 0.037081144750118256, + 0.020912116393446922, + 0.3448124825954437 + ], + "credit_map": [ + 0.008012463338673115, + 0.006534335669130087, + 0.00940806046128273, + 0.007314688991755247, + 0.006783744320273399, + 0.006764756049960852, + 0.00635924469679594, + 0.008286743424832821, + 0.005954751744866371, + 0.007145680952817202, + 0.00573357567191124, + 0.0037730280309915543, + 0.005757859442383051, + 0.0043372358195483685, + 0.0046495841816067696, + 0.0025449348613619804, + 0.010376478545367718, + 0.003977064974606037, + 0.005298292264342308, + 0.005426160059869289, + 0.004758047871291637, + 0.0032839325722306967, + 0.01969349943101406, + 0.009608292952179909, + 0.032315611839294434, + 0.028568871319293976, + 0.007677769288420677, + 0.01851469650864601, + 0.051914606243371964, + 0.024752916768193245, + 0.02473309263586998, + 0.016782473772764206, + 0.004274406470358372, + 0.0032234303653240204, + 0.0035549963358789682, + 0.004780118819326162, + 0.003457839135080576, + 0.0025693129282444715, + 0.003885875456035137, + 0.00268230470828712, + 0.01585829071700573, + 0.01283975224941969, + 0.02789539098739624, + 0.019944168627262115, + 0.01431216113269329, + 0.0181843563914299, + 0.024556763470172882, + 0.021991070359945297, + 0.010204432532191277, + 0.005231102928519249, + 0.0032052299939095974, + 0.0027112169191241264, + 0.005697536747902632, + 0.013653281144797802, + 0.00256459042429924, + 0.0028740838170051575, + 0.005541199818253517, + 0.006103075109422207, + 0.003311749082058668, + 0.004862009547650814, + 0.00411522202193737, + 0.005185718182474375, + 0.003695682156831026, + 0.004266493022441864, + 0.0028172002639621496, + 0.00254121795296669, + 0.0038458863273262978, + 0.0017175533575937152, + 0.0026004004757851362, + 0.0032382889185100794, + 0.002287243725731969, + 0.0018643276998773217, + 0.030853234231472015, + 0.041170306503772736, + 0.03597274422645569, + 0.056794341653585434, + 0.04294886440038681, + 0.048075441271066666, + 0.05603301525115967, + 0.0329645611345768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.012999314814805984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00466105155646801 + ], + "opposite_logprob_tokens": [ + -5.3796610832214355 + ], + "logprob_reward": -0.00466105155646801, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00466105155646801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050970710813999176, + 0.028391152620315552, + 0.023013632744550705, + 0.02678823657333851, + 0.0842193216085434, + 0.5312877893447876, + 0.07398848980665207, + 0.05346833914518356, + 0.04055783897638321, + 0.08731439709663391 + ], + "credit_map": [ + 0.00986802764236927, + 0.005777331534773111, + 0.005448135081678629, + 0.006486773956567049, + 0.006693175528198481, + 0.0056367735378444195, + 0.0043619535863399506, + 0.006698539946228266, + 0.0030520372092723846, + 0.0039589423686265945, + 0.003480108454823494, + 0.003809628775343299, + 0.0036812948528677225, + 0.0031093505676835775, + 0.004209895618259907, + 0.003089894773438573, + 0.0036476070526987314, + 0.002233295701444149, + 0.003381571965292096, + 0.0014712520642206073, + 0.0029007347766309977, + 0.0021723562385886908, + 0.0038501182571053505, + 0.003356697503477335, + 0.0037345788441598415, + 0.0027400823310017586, + 0.003050358034670353, + 0.0034555874299257994, + 0.003346429206430912, + 0.0027622494380921125, + 0.004607971291989088, + 0.0030909802298992872, + 0.011984019540250301, + 0.007138142827898264, + 0.010835938155651093, + 0.021210817620158195, + 0.0060782828368246555, + 0.007493337616324425, + 0.014278342016041279, + 0.005200439598411322, + 0.05291874334216118, + 0.038551971316337585, + 0.10463322699069977, + 0.08173566311597824, + 0.04784366860985756, + 0.055015288293361664, + 0.08272862434387207, + 0.06786061078310013, + 0.0097348066046834, + 0.012914127670228481, + 0.005506372544914484, + 0.0066184052266180515, + 0.01800382323563099, + 0.009361299686133862, + 0.006619899533689022, + 0.0052297539077699184, + 0.005627987906336784, + 0.004316144157201052, + 0.005734039470553398, + 0.005472647491842508, + 0.005088163539767265, + 0.00587842334061861, + 0.007280486635863781, + 0.014070449396967888, + 0.006445351522415876, + 0.0034681609831750393, + 0.005955366417765617, + 0.0031260165851563215, + 0.007084030192345381, + 0.006609675940126181, + 0.004538361448794603, + 0.003330877050757408, + 0.0073385899886488914, + 0.009822347201406956, + 0.010081402957439423, + 0.015004943124949932, + 0.009401612915098667, + 0.012345761060714722, + 0.014088250696659088, + 0.00923149287700653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.00466105155646801 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949051585048437 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949051585048437, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949051585048437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056202277541160583, + 0.018610671162605286, + 0.0513857863843441, + 0.024671468883752823, + 0.060067564249038696, + 0.6507452726364136, + 0.03753719851374626, + 0.03910665586590767, + 0.015832414850592613, + 0.045840684324502945 + ], + "credit_map": [ + 0.006608061492443085, + 0.006013878155499697, + 0.009389174170792103, + 0.006044395267963409, + 0.005711262580007315, + 0.006827876903116703, + 0.00707125011831522, + 0.008536387234926224, + 0.0025809588842093945, + 0.0031427154317498207, + 0.0023892640601843596, + 0.00169849069789052, + 0.0034690832253545523, + 0.0015247241826727986, + 0.0021701306104660034, + 0.0016353074461221695, + 0.008373884484171867, + 0.003873310284689069, + 0.004834762308746576, + 0.004536428954452276, + 0.004527372773736715, + 0.00465982872992754, + 0.012390688061714172, + 0.008189518935978413, + 0.0034301881678402424, + 0.002510946476832032, + 0.0021446109749376774, + 0.0035491485614329576, + 0.0037428659852594137, + 0.00271477154456079, + 0.00403959583491087, + 0.0025393443647772074, + 0.007769661024212837, + 0.005329822655767202, + 0.006528932135552168, + 0.01601864956319332, + 0.004698059055954218, + 0.00491721834987402, + 0.011248406022787094, + 0.0035568219609558582, + 0.06281550973653793, + 0.04427633807063103, + 0.13984636962413788, + 0.11134804785251617, + 0.05484562739729881, + 0.06435051560401917, + 0.09608376026153564, + 0.07717917859554291, + 0.009670039638876915, + 0.0032398200128227472, + 0.0022791107185184956, + 0.0021589873358607292, + 0.004653323907405138, + 0.010970585979521275, + 0.0024328986182808876, + 0.0021324355620890856, + 0.003056424669921398, + 0.002559436485171318, + 0.0042038667015731335, + 0.004681152291595936, + 0.003955209162086248, + 0.004098151810467243, + 0.0049841864965856075, + 0.01156823430210352, + 0.002560090972110629, + 0.0016415683785453439, + 0.003072881605476141, + 0.0013206284493207932, + 0.0022793705575168133, + 0.001887288293801248, + 0.0014313004212453961, + 0.001639287220314145, + 0.004430253058671951, + 0.00677734287455678, + 0.003977295942604542, + 0.00616675429046154, + 0.006992195267230272, + 0.006291053723543882, + 0.006329063326120377, + 0.004876735154539347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.005949051585048437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035307943820953 + ], + "opposite_logprob_tokens": [ + -1.9103530645370483 + ], + "logprob_reward": -0.16035307943820953, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035307943820953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12302692234516144, + 0.11328138411045074, + 0.12126535177230835, + 0.09666042029857635, + 0.04541720077395439, + 0.11375685781240463, + 0.20323817431926727, + 0.0460253544151783, + 0.028124487027525902, + 0.10920383781194687 + ], + "credit_map": [ + 0.013886132277548313, + 0.0106179378926754, + 0.018863530829548836, + 0.01099330186843872, + 0.01705046184360981, + 0.015007654204964638, + 0.012571373023092747, + 0.024036534130573273, + 0.012903283350169659, + 0.016692887991666794, + 0.016881344839930534, + 0.023502236232161522, + 0.014414272271096706, + 0.008269675076007843, + 0.01423435378819704, + 0.006383331026881933, + 0.01984776370227337, + 0.015961047261953354, + 0.013279377482831478, + 0.013956394977867603, + 0.014019579626619816, + 0.016386087983846664, + 0.016012517735362053, + 0.011802585795521736, + 0.013983521610498428, + 0.013029840774834156, + 0.006472561974078417, + 0.01235599722713232, + 0.01610064134001732, + 0.010771236382424831, + 0.016538552939891815, + 0.0074080596677958965, + 0.006376758683472872, + 0.005541327875107527, + 0.004860507789999247, + 0.00949066411703825, + 0.00460021710023284, + 0.0040193055756390095, + 0.00730164023116231, + 0.0032267789356410503, + 0.012663743458688259, + 0.009601019322872162, + 0.020941123366355896, + 0.02140679769217968, + 0.011777334846556187, + 0.010845648124814034, + 0.014899395406246185, + 0.011621806770563126, + 0.022089140489697456, + 0.033555157482624054, + 0.022579239681363106, + 0.018782036378979683, + 0.041063904762268066, + 0.026769500225782394, + 0.018350675702095032, + 0.020048512145876884, + 0.004276466555893421, + 0.005252589005976915, + 0.004099649842828512, + 0.00744702760130167, + 0.004649728536605835, + 0.0051227291114628315, + 0.0052580926567316055, + 0.00991907436400652, + 0.0037290204782038927, + 0.0035954948980361223, + 0.005153738893568516, + 0.0023858859203755856, + 0.003751014359295368, + 0.00415074871852994, + 0.002759129274636507, + 0.002599454717710614, + 0.00942470133304596, + 0.014007137157022953, + 0.00928664393723011, + 0.014082922600209713, + 0.015018478967249393, + 0.015274200588464737, + 0.021755430847406387, + 0.010354327037930489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.16035307943820953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743293721228838 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743293721228838, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743293721228838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1290283203125, + 0.06291250884532928, + 0.11316226422786713, + 0.07642177492380142, + 0.09556324779987335, + 0.15088345110416412, + 0.15503861010074615, + 0.059456340968608856, + 0.053470637649297714, + 0.10406280308961868 + ], + "credit_map": [ + 0.01724535971879959, + 0.017727568745613098, + 0.015434598550200462, + 0.01589752733707428, + 0.01730991341173649, + 0.015067697502672672, + 0.011340978555381298, + 0.01900468021631241, + 0.0066980705596506596, + 0.008270432241261005, + 0.007648624945431948, + 0.005172687117010355, + 0.0071245646104216576, + 0.008077285252511501, + 0.01127739530056715, + 0.00864345021545887, + 0.01997467130422592, + 0.007784835994243622, + 0.013666381128132343, + 0.01192924752831459, + 0.008524682372808456, + 0.010103448294103146, + 0.022819295525550842, + 0.018359705805778503, + 0.012216398492455482, + 0.009099396876990795, + 0.007085528690367937, + 0.008648231625556946, + 0.012926646508276463, + 0.007897583767771721, + 0.010307352058589458, + 0.008240633644163609, + 0.013274293392896652, + 0.00911711249500513, + 0.015913451090455055, + 0.022033225744962692, + 0.00685897609218955, + 0.00842325296252966, + 0.012446541339159012, + 0.00749640166759491, + 0.014650403521955013, + 0.01320470031350851, + 0.0292897280305624, + 0.03247341141104698, + 0.013613645918667316, + 0.014050024561583996, + 0.018866516649723053, + 0.014735009521245956, + 0.04057791084051132, + 0.01332488190382719, + 0.0055910139344632626, + 0.008184613659977913, + 0.019710306078195572, + 0.051750294864177704, + 0.008532932959496975, + 0.0073666563257575035, + 0.008360620588064194, + 0.01080753281712532, + 0.005604923237115145, + 0.007171525154262781, + 0.006748075131326914, + 0.007725829724222422, + 0.00658367108553648, + 0.006454164162278175, + 0.007794225122779608, + 0.005610367748886347, + 0.008620262145996094, + 0.0038168609607964754, + 0.0088957529515028, + 0.008933259174227715, + 0.005645052529871464, + 0.004154853522777557, + 0.010858752764761448, + 0.01683141104876995, + 0.011127801612019539, + 0.0142480144277215, + 0.014877069741487503, + 0.014768058434128761, + 0.011018333956599236, + 0.010333363898098469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.006743293721228838 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In one unbroken shot, a dynamic tracking camera fluidly follows a pair of high-top roller skates as they perform a frantic dance across an infinite, canvas-like void. The skates themselves are rendered in explosive cracks of vibrant color, their smooth, chameleon-like surfaces shifting between neon fuchsia, electric cyan, and acid yellow with each movement, capturing unseen light with sharp, specular highlights. They swirl and spin with chaotic precision, one moment gliding smoothly backward, the next executing sharp, aggressive turns that slice into darkness. With each turn and spin, the skates leave behind thick, vibrant, board-like marks of paint in colors that directly correspond to their movements—fiery orange streaks from sudden acceleration, deep purple splashes from stomping landings. The camera remains in close contact with their movements, capturing their chaotic ballet of pivots and turns as they continually construct intricate three-dimensional action paintings within the desolate, matte void, creating an atmosphere of pure, unfettered kinetic energy and abstract expression.", + "seed": 29495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_14/index_7196_seed_29495_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": -0.01809754124842584, + "vlm_logprob": -0.01809754124842584, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.01809754124842584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006922428961843252 + ], + "opposite_logprob_tokens": [ + -5.881922245025635 + ], + "logprob_reward": -0.006922428961843252, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006922428961843252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05910206958651543, + 0.11110150814056396, + 0.20529115200042725, + 0.15268367528915405, + 0.0832749456167221, + 0.07043251395225525, + 0.04953521117568016, + 0.07494655251502991, + 0.04773921146988869, + 0.1458931267261505 + ], + "credit_map": [ + 0.008142679929733276, + 0.007235497701913118, + 0.00782477855682373, + 0.008260486647486687, + 0.007095964625477791, + 0.006314896512776613, + 0.006171587388962507, + 0.008056181482970715, + 0.016656536608934402, + 0.00896828155964613, + 0.01240617223083973, + 0.015619432553648949, + 0.01041930727660656, + 0.007839343510568142, + 0.02312968485057354, + 0.01606275886297226, + 0.031341154128313065, + 0.036199308931827545, + 0.021098533645272255, + 0.022411540150642395, + 0.025148121640086174, + 0.03255085647106171, + 0.021682756021618843, + 0.014858883805572987, + 0.03078118897974491, + 0.020450444892048836, + 0.01812971755862236, + 0.015444137156009674, + 0.017112035304307938, + 0.03410535678267479, + 0.009202994406223297, + 0.007457810454070568, + 0.01769391819834709, + 0.009889354929327965, + 0.010833349078893661, + 0.007329583168029785, + 0.008147384971380234, + 0.013705290853977203, + 0.007954154163599014, + 0.007721918169409037, + 0.010464414022862911, + 0.010551649145781994, + 0.011020887643098831, + 0.0068537453189492226, + 0.007174689322710037, + 0.008066251873970032, + 0.009335664100944996, + 0.006965214386582375, + 0.003071753541007638, + 0.005022034980356693, + 0.0036586918868124485, + 0.004131317604333162, + 0.004095891024917364, + 0.009339063428342342, + 0.0129417534917593, + 0.007274708244949579, + 0.020738964900374413, + 0.00582526670768857, + 0.00566203985363245, + 0.006045591086149216, + 0.006076542194932699, + 0.007608396001160145, + 0.006662776228040457, + 0.016326982527971268, + 0.005280074197798967, + 0.0071134064346551895, + 0.006327814422547817, + 0.007054124493151903, + 0.006084475200623274, + 0.005430776625871658, + 0.0053705195896327496, + 0.0050780288875103, + 0.011084179393947124, + 0.008598417043685913, + 0.009382806718349457, + 0.00846362579613924, + 0.01164038386195898, + 0.034341778606176376, + 0.018314208835363388, + 0.04406774416565895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.006922428961843252 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003633565502241254 + ], + "opposite_logprob_tokens": [ + -5.628633499145508 + ], + "logprob_reward": -0.003633565502241254, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003633565502241254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05709484592080116, + 0.051990192383527756, + 0.0642997995018959, + 0.036985427141189575, + 0.1946856677532196, + 0.04169075936079025, + 0.08788356930017471, + 0.23475204408168793, + 0.05852729454636574, + 0.17209039628505707 + ], + "credit_map": [ + 0.007745910435914993, + 0.006379297003149986, + 0.00904845166951418, + 0.00501901563256979, + 0.005039895419031382, + 0.007986621931195259, + 0.005040911957621574, + 0.010834743268787861, + 0.006819914560765028, + 0.006242178846150637, + 0.006532147526741028, + 0.006806062068790197, + 0.004250082187354565, + 0.0052337246015667915, + 0.008594037033617496, + 0.007512048352509737, + 0.009835383854806423, + 0.01176208071410656, + 0.007023259066045284, + 0.006221081595867872, + 0.007165118586272001, + 0.010724978521466255, + 0.005838687531650066, + 0.005729211959987879, + 0.0058370307087898254, + 0.004920669831335545, + 0.006548842880874872, + 0.003721554996445775, + 0.005166495218873024, + 0.005262547638267279, + 0.003701156936585903, + 0.0018271295120939612, + 0.03178166225552559, + 0.025437699630856514, + 0.01780008338391781, + 0.016049785539507866, + 0.025281023234128952, + 0.030590295791625977, + 0.023052062839269638, + 0.02469305694103241, + 0.005238272249698639, + 0.006446319632232189, + 0.006049459800124168, + 0.004362430423498154, + 0.004169152583926916, + 0.0058615063317120075, + 0.005444319918751717, + 0.004119296092540026, + 0.004375956021249294, + 0.005063669756054878, + 0.00513690710067749, + 0.006105373613536358, + 0.005777627695351839, + 0.0151428347453475, + 0.029292572289705276, + 0.01698862388730049, + 0.07345334440469742, + 0.018069883808493614, + 0.01804070733487606, + 0.01114600244909525, + 0.01576877571642399, + 0.020272688940167427, + 0.01641562208533287, + 0.061585020273923874, + 0.005140396766364574, + 0.009654472582042217, + 0.0069063520058989525, + 0.0053968848660588264, + 0.0066582215949893, + 0.010203547775745392, + 0.006311407312750816, + 0.008256010711193085, + 0.011891611851751804, + 0.009976152330636978, + 0.0105453385040164, + 0.010110857896506786, + 0.0138581907376647, + 0.04451220855116844, + 0.020287448540329933, + 0.05090859904885292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.003633565502241254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733347196131945 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733347196131945, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733347196131945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10337560623884201, + 0.06286893784999847, + 0.13000676035881042, + 0.09247581660747528, + 0.16732211410999298, + 0.051856424659490585, + 0.13115857541561127, + 0.0546538308262825, + 0.05142219737172127, + 0.1548597365617752 + ], + "credit_map": [ + 0.013734797015786171, + 0.010433108545839787, + 0.014882237650454044, + 0.007497283164411783, + 0.009386944584548473, + 0.016155289486050606, + 0.00983215868473053, + 0.02145378664135933, + 0.008127395994961262, + 0.007133473176509142, + 0.009621242992579937, + 0.008412377908825874, + 0.005567425396293402, + 0.005136333871632814, + 0.009527112357318401, + 0.009343573823571205, + 0.01961427554488182, + 0.025333968922495842, + 0.01300887018442154, + 0.01109028048813343, + 0.013874433934688568, + 0.025777386501431465, + 0.010920765809714794, + 0.01038677990436554, + 0.014557112008333206, + 0.011613130569458008, + 0.010555275715887547, + 0.012001529335975647, + 0.012413985095918179, + 0.014923076145350933, + 0.010755539871752262, + 0.005656166933476925, + 0.01690004952251911, + 0.016704419627785683, + 0.012563222087919712, + 0.013242465443909168, + 0.019086429849267006, + 0.026948310434818268, + 0.020601412281394005, + 0.041275810450315475, + 0.006143678445369005, + 0.00803130678832531, + 0.008145214058458805, + 0.005845969542860985, + 0.005820911377668381, + 0.007013610098510981, + 0.006078095640987158, + 0.004777641035616398, + 0.006399640813469887, + 0.006440341472625732, + 0.00637019844725728, + 0.00911243911832571, + 0.008895386010408401, + 0.024689998477697372, + 0.04420837014913559, + 0.02504218928515911, + 0.009946076199412346, + 0.007090540137141943, + 0.0038683153688907623, + 0.004930037073791027, + 0.007782910950481892, + 0.007361854892224073, + 0.005170274991542101, + 0.008503822609782219, + 0.003956535365432501, + 0.0046201734803617, + 0.006111496593803167, + 0.005559365730732679, + 0.006437193136662245, + 0.0074173687025904655, + 0.00752885639667511, + 0.009791208431124687, + 0.007041879929602146, + 0.004380802623927593, + 0.006132734939455986, + 0.002915923250839114, + 0.007070127874612808, + 0.04223989322781563, + 0.02578914165496826, + 0.05928922817111015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.006733347196131945 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000092417001724 + ], + "opposite_logprob_tokens": [ + -2.6950008869171143 + ], + "logprob_reward": -0.07000092417001724, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000092417001724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08347802609205246, + 0.02541232667863369, + 0.05354541167616844, + 0.03356403112411499, + 0.05366875231266022, + 0.03825611248612404, + 0.05848093330860138, + 0.24168388545513153, + 0.042576421052217484, + 0.36933401226997375 + ], + "credit_map": [ + 0.011312306858599186, + 0.006867964752018452, + 0.010516410693526268, + 0.006744543556123972, + 0.008731155656278133, + 0.014457996003329754, + 0.006766021717339754, + 0.018081625923514366, + 0.004201218485832214, + 0.0035301728639751673, + 0.003287527011707425, + 0.0036525900941342115, + 0.0025394235271960497, + 0.002692407462745905, + 0.003331512911245227, + 0.002177475020289421, + 0.008192446082830429, + 0.009459294378757477, + 0.005550019908696413, + 0.004912194330245256, + 0.00606111902743578, + 0.01093400176614523, + 0.004522252362221479, + 0.0039140889421105385, + 0.005019516684114933, + 0.004283929709345102, + 0.0058632162399590015, + 0.003728533396497369, + 0.004079876467585564, + 0.004899225197732449, + 0.0028335703536868095, + 0.0028561654035001993, + 0.0062944646924734116, + 0.005787518806755543, + 0.004634509328752756, + 0.0045170169323682785, + 0.005092514678835869, + 0.007839205674827099, + 0.007944818586111069, + 0.011558707803487778, + 0.004075835924595594, + 0.005713983438909054, + 0.005472326651215553, + 0.004833301994949579, + 0.004246329888701439, + 0.004712729249149561, + 0.006135657429695129, + 0.0030659514013677835, + 0.0029218727722764015, + 0.0034632012248039246, + 0.003171777818351984, + 0.003675369778648019, + 0.0044355266727507114, + 0.010402314364910126, + 0.019101759418845177, + 0.011309115216135979, + 0.0696231946349144, + 0.02171134389936924, + 0.019859876483678818, + 0.016725249588489532, + 0.017352720722556114, + 0.020372547209262848, + 0.01772943139076233, + 0.05830954760313034, + 0.00361196743324399, + 0.009488837793469429, + 0.004928302019834518, + 0.003930588718503714, + 0.006597852800041437, + 0.0053448062390089035, + 0.004471533000469208, + 0.004202539566904306, + 0.016114210709929466, + 0.00931767001748085, + 0.015542506240308285, + 0.009001892991364002, + 0.0180924404412508, + 0.10339242219924927, + 0.05868891626596451, + 0.13918398320674896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.07000092417001724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 high-top roller skates the primary subject visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the skates display shifting colors including neon fuchsia, electric cyan, and acid yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the skates leave behind thick, vibrant paint marks on the ground that correspond to their movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background an infinite, canvas-like void with a matte finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot unbroken and does it track fluidly close to the skates' movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031974404118955135 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.0031974404118955135, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031974404118955135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05004440248012543, + 0.051284000277519226, + 0.1246519535779953, + 0.214249849319458, + 0.0676649883389473, + 0.07592645287513733, + 0.053071506321430206, + 0.10273177176713943, + 0.0697016641497612, + 0.19067339599132538 + ], + "credit_map": [ + 0.0067449212074279785, + 0.006434946786612272, + 0.007609658874571323, + 0.004523840732872486, + 0.006372374948114157, + 0.008063536137342453, + 0.004380304832011461, + 0.0059148212894797325, + 0.006152988411486149, + 0.003942721989005804, + 0.007584700360894203, + 0.00717992102727294, + 0.0035324906930327415, + 0.005043113138526678, + 0.009142160415649414, + 0.008705907501280308, + 0.019328966736793518, + 0.023363322019577026, + 0.012041240930557251, + 0.011769280768930912, + 0.013123653829097748, + 0.025977689772844315, + 0.009685557335615158, + 0.009362248703837395, + 0.03930896148085594, + 0.028196927160024643, + 0.03261616453528404, + 0.03149157762527466, + 0.017251525074243546, + 0.036332279443740845, + 0.0171112772077322, + 0.0119411526247859, + 0.00888662226498127, + 0.009671386331319809, + 0.006732902955263853, + 0.00796812865883112, + 0.00975795928388834, + 0.010381330735981464, + 0.008146611973643303, + 0.0061200507916510105, + 0.008911574259400368, + 0.01240029837936163, + 0.014903240837156773, + 0.007205547299236059, + 0.007858501747250557, + 0.0105363167822361, + 0.008315523155033588, + 0.005795461591333151, + 0.004002354107797146, + 0.004996472038328648, + 0.0051246630027890205, + 0.003961009439080954, + 0.007061568088829517, + 0.009607933461666107, + 0.01128393318504095, + 0.007033579982817173, + 0.02241336926817894, + 0.013552646152675152, + 0.008824999444186687, + 0.008245124481618404, + 0.011400151997804642, + 0.00918656773865223, + 0.008453873917460442, + 0.02065505087375641, + 0.0067312815226614475, + 0.013630655594170094, + 0.008905608206987381, + 0.00928929541260004, + 0.00826379656791687, + 0.007310030050575733, + 0.007069211453199387, + 0.008501793257892132, + 0.011235264129936695, + 0.006879087071865797, + 0.009592772461473942, + 0.00630929134786129, + 0.011358690448105335, + 0.05265006422996521, + 0.02639981172978878, + 0.06624843180179596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.0031974404118955135 + } + ], + "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/0000155/group_15/index_tensor([4308, 4308])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_15/index_tensor([4308, 4308])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..04e00a82fd456822627873498b1885cb95e8e4cc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_15/index_tensor([4308, 4308])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows eight majestic horses, rendered in thick, fluid ink brushstrokes, as they gallop with unbridled force. Each horse's muscular lines are imbued with explosive power, as if shaped by the finest calligraphy. Their manes and tails swirl like wisps of ink in the gusting wind, their gestures distinct but their purpose unified. They charge into an ancient village ablaze. The raging flames, the only source of light, paint the sky an ominous orange-red. The village's wooden buildings are reduced to swaying black silhouettes in the flames, while thick black smoke billows and rises like spilled ink. The blazing light illuminates the horses' sleek, inky bodies, casting shadows on the ground beneath their feet that are elongated, sharp, and wildly dance with the flames. The horses' hooves stir up sparks and dust, blending seamlessly with the ink's ethereal and ethereal effects. The entire picture constitutes a dynamic ink painting full of dramatic interweaving of light and shadow, coexistence of power and destruction, creating an epic, tense and magnificent atmosphere.", + "seed": 1149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_15/index_4308_seed_1149_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": -0.06758048441261053, + "vlm_logprob": -0.06758048441261053, + "vlm_yes_no_margin": 3.699999976158142, + "vlm_token_credit": -0.06758048441261053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14798445999622345 + ], + "opposite_logprob_tokens": [ + -2.022984504699707 + ], + "logprob_reward": -0.14798445999622345, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14798445999622345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2734503149986267, + 0.05645567178726196, + 0.06286457926034927, + 0.04025525599718094, + 0.19402346014976501, + 0.060179583728313446, + 0.03822998329997063, + 0.028536956757307053, + 0.06055798381567001, + 0.18544629216194153 + ], + "credit_map": [ + 0.04236190766096115, + 0.03246724605560303, + 0.03188326954841614, + 0.02920953556895256, + 0.031882382929325104, + 0.03614702448248863, + 0.05108116567134857, + 0.01841774582862854, + 0.005606671329587698, + 0.004703650251030922, + 0.011711115948855877, + 0.008387298323214054, + 0.007146250922232866, + 0.005440218839794397, + 0.006601595785468817, + 0.006858863867819309, + 0.010327127762138844, + 0.012376751750707626, + 0.010267929174005985, + 0.008427919819951057, + 0.006258258130401373, + 0.004584714770317078, + 0.005251619033515453, + 0.005370257422327995, + 0.005370554979890585, + 0.004745726007968187, + 0.007021213416010141, + 0.006327239330857992, + 0.005370877217501402, + 0.004049973096698523, + 0.003700826084241271, + 0.003668841440230608, + 0.01956869475543499, + 0.03125660493969917, + 0.01777779310941696, + 0.024752216413617134, + 0.04032633826136589, + 0.022436710074543953, + 0.022555291652679443, + 0.01534978486597538, + 0.007383351679891348, + 0.006647299509495497, + 0.008429719135165215, + 0.006187723949551582, + 0.006851310841739178, + 0.007273597177118063, + 0.011801037937402725, + 0.005605535581707954, + 0.0038364699576050043, + 0.00618286058306694, + 0.005322960205376148, + 0.0025976900942623615, + 0.004650535993278027, + 0.006001168861985207, + 0.005092936102300882, + 0.004545360803604126, + 0.0038332168478518724, + 0.004171416163444519, + 0.0031892466358840466, + 0.003012744477018714, + 0.004848124459385872, + 0.003804432461038232, + 0.0032844110392034054, + 0.0023933639749884605, + 0.006760685704648495, + 0.009478191845119, + 0.012650816701352596, + 0.006527555640786886, + 0.00875302404165268, + 0.006540835369378328, + 0.0050901551730930805, + 0.0047567118890583515, + 0.012121525593101978, + 0.015934394672513008, + 0.04002745822072029, + 0.015019205398857594, + 0.013226035051047802, + 0.010204107500612736, + 0.0653817355632782, + 0.013531796634197235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.14798445999622345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067558446899056435 + ], + "opposite_logprob_tokens": [ + -5.006755828857422 + ], + "logprob_reward": -0.0067558446899056435, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067558446899056435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17418444156646729, + 0.07940622419118881, + 0.052663654088974, + 0.06905431300401688, + 0.18938413262367249, + 0.042071111500263214, + 0.035741835832595825, + 0.07117830961942673, + 0.12691931426525116, + 0.15939661860466003 + ], + "credit_map": [ + 0.01935633458197117, + 0.023733267560601234, + 0.023118840530514717, + 0.02760506421327591, + 0.020598258823156357, + 0.02050151117146015, + 0.02747812494635582, + 0.011793028563261032, + 0.010192951187491417, + 0.00721671711653471, + 0.010475898161530495, + 0.00864423718303442, + 0.01060521136969328, + 0.00957147404551506, + 0.01538616418838501, + 0.007313569542020559, + 0.007307152263820171, + 0.006946031469851732, + 0.0068720136769115925, + 0.007173505146056414, + 0.004323685076087713, + 0.005349166691303253, + 0.00643053138628602, + 0.008261569775640965, + 0.0070413160137832165, + 0.009206820279359818, + 0.009586933068931103, + 0.007979688234627247, + 0.00814108271151781, + 0.009375904686748981, + 0.010136579163372517, + 0.007585989776998758, + 0.028120875358581543, + 0.025430867448449135, + 0.02037334442138672, + 0.023095425218343735, + 0.03511497750878334, + 0.019642183557152748, + 0.0235923882573843, + 0.014014077372848988, + 0.00638155872002244, + 0.004561265930533409, + 0.005947360768914223, + 0.00601985352113843, + 0.004961029160767794, + 0.005846422631293535, + 0.0051627326756715775, + 0.003190889023244381, + 0.004400982521474361, + 0.0047289528883993626, + 0.004731154069304466, + 0.0052200001664459705, + 0.003921505995094776, + 0.004813789390027523, + 0.0033650854602456093, + 0.004560364875942469, + 0.006957252509891987, + 0.008934813551604748, + 0.00969903077930212, + 0.008613640442490578, + 0.012384631671011448, + 0.011367080733180046, + 0.007586210500448942, + 0.005635646637529135, + 0.01211341843008995, + 0.020227691158652306, + 0.022280141711235046, + 0.014485240913927555, + 0.014126825146377087, + 0.015320457518100739, + 0.018945325165987015, + 0.009420225396752357, + 0.016451088711619377, + 0.034523483365774155, + 0.01893816888332367, + 0.022255856543779373, + 0.013588830828666687, + 0.02053188718855381, + 0.01899023912847042, + 0.014117074199020863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0067558446899056435 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495170131325722 + ], + "opposite_logprob_tokens": [ + -2.9299516677856445 + ], + "logprob_reward": -0.05495170131325722, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495170131325722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34715735912323, + 0.06611668318510056, + 0.043181076645851135, + 0.03835190087556839, + 0.2568892240524292, + 0.03892942890524864, + 0.07855118066072464, + 0.034455716609954834, + 0.04060357064008713, + 0.0557638444006443 + ], + "credit_map": [ + 0.04976499080657959, + 0.04000174626708031, + 0.04253124073147774, + 0.04311540350317955, + 0.04640791565179825, + 0.0402468666434288, + 0.06529615819454193, + 0.019793037325143814, + 0.007037763018161058, + 0.0050096274353563786, + 0.0076962681487202644, + 0.005811755079776049, + 0.012388864532113075, + 0.005827696993947029, + 0.014085457660257816, + 0.008259258233010769, + 0.006951085291802883, + 0.007100674789398909, + 0.0058041224256157875, + 0.007083759177476168, + 0.00517271226271987, + 0.00342422048561275, + 0.0039406390860676765, + 0.0037038661539554596, + 0.003855833550915122, + 0.004447316285222769, + 0.005302722565829754, + 0.004776799585670233, + 0.003919095732271671, + 0.003565557999536395, + 0.0072513362392783165, + 0.0052332403138279915, + 0.02625967562198639, + 0.03704489767551422, + 0.02134702540934086, + 0.03338608145713806, + 0.05213078111410141, + 0.029180439189076424, + 0.03498998284339905, + 0.022550353780388832, + 0.003981037065386772, + 0.004887188319116831, + 0.005245447624474764, + 0.006049280520528555, + 0.004317236598581076, + 0.00545860081911087, + 0.00534959277138114, + 0.0036410505417734385, + 0.008435661904513836, + 0.017726518213748932, + 0.014170453883707523, + 0.009483284316956997, + 0.007243369705975056, + 0.007887563668191433, + 0.00528540275990963, + 0.008318930864334106, + 0.003564645303413272, + 0.00388729153200984, + 0.0037970715202391148, + 0.005425203125923872, + 0.006316250655800104, + 0.004334826488047838, + 0.004443487152457237, + 0.0026869431603699923, + 0.0049665202386677265, + 0.005184988025575876, + 0.004165899474173784, + 0.004452799446880817, + 0.0074699558317661285, + 0.007487019989639521, + 0.003298693336546421, + 0.0035776933655142784, + 0.0051680514588952065, + 0.00704379566013813, + 0.012642220593988895, + 0.0049920217134058475, + 0.004376129247248173, + 0.004626949783414602, + 0.012800004333257675, + 0.004114674869924784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.05495170131325722 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178523525595665 + ], + "opposite_logprob_tokens": [ + -6.751178741455078 + ], + "logprob_reward": -0.001178523525595665, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178523525595665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2234562784433365, + 0.09699451178312302, + 0.04086742922663689, + 0.056622423231601715, + 0.256692111492157, + 0.03385947272181511, + 0.05008677765727043, + 0.042442359030246735, + 0.08305852860212326, + 0.11592011898756027 + ], + "credit_map": [ + 0.0408581979572773, + 0.027137665078043938, + 0.021582452580332756, + 0.016903063282370567, + 0.024287045001983643, + 0.030020758509635925, + 0.050653330981731415, + 0.012013737112283707, + 0.010984471999108791, + 0.008838611654937267, + 0.011999786831438541, + 0.011656754650175571, + 0.015002165921032429, + 0.010385360568761826, + 0.017539437860250473, + 0.010587922297418118, + 0.008201372809708118, + 0.006445732899010181, + 0.005820485297590494, + 0.005831086076796055, + 0.0038439794443547726, + 0.003214264055714011, + 0.004158133640885353, + 0.0033523740712553263, + 0.004885545000433922, + 0.005519449710845947, + 0.007283949293196201, + 0.007200438063591719, + 0.006416346412152052, + 0.00510599697008729, + 0.01106764655560255, + 0.009143046103417873, + 0.0275218915194273, + 0.037171948701143265, + 0.020981663838028908, + 0.03499250113964081, + 0.05372905731201172, + 0.029977571219205856, + 0.03247423470020294, + 0.019843226298689842, + 0.003597547532990575, + 0.004244150593876839, + 0.003503193147480488, + 0.00487024150788784, + 0.0046759978868067265, + 0.004698951728641987, + 0.004966018721461296, + 0.003303365781903267, + 0.005183489061892033, + 0.008057263679802418, + 0.007989672012627125, + 0.007711692713201046, + 0.004377094563096762, + 0.00505983317270875, + 0.003685549832880497, + 0.008022179827094078, + 0.005097344517707825, + 0.0063569266349077225, + 0.006131857167929411, + 0.004876402206718922, + 0.00603838125243783, + 0.005554017145186663, + 0.004388221073895693, + 0.003999204374849796, + 0.006644922308623791, + 0.007958076894283295, + 0.02264544740319252, + 0.0054635037668049335, + 0.009890872053802013, + 0.006111051421612501, + 0.019398158416152, + 0.004946491681039333, + 0.007932931184768677, + 0.00773311173543334, + 0.02802439220249653, + 0.0063841091468930244, + 0.00899035669863224, + 0.007208413910120726, + 0.04202421382069588, + 0.0076225921511650085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.001178523525595665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703189253807068 + ], + "opposite_logprob_tokens": [ + -2.1270318031311035 + ], + "logprob_reward": -0.12703189253807068, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703189253807068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24910245835781097, + 0.09670758247375488, + 0.03157700225710869, + 0.04544490575790405, + 0.18640390038490295, + 0.06801930069923401, + 0.041945185512304306, + 0.08818291872739792, + 0.09302117675542831, + 0.0995955839753151 + ], + "credit_map": [ + 0.03903751075267792, + 0.026012012735009193, + 0.02330661192536354, + 0.025713693350553513, + 0.025231076404452324, + 0.034912336617708206, + 0.05184154212474823, + 0.0230476725846529, + 0.009020189754664898, + 0.00915735587477684, + 0.015215987339615822, + 0.014070413075387478, + 0.00940711423754692, + 0.012949461117386818, + 0.016347331926226616, + 0.010539732873439789, + 0.005960473325103521, + 0.004152974113821983, + 0.004685998894274235, + 0.003827781416475773, + 0.003644578391686082, + 0.0033965306356549263, + 0.002909000962972641, + 0.002999665215611458, + 0.0054297964088618755, + 0.0053742616437375546, + 0.00532192550599575, + 0.006852584891021252, + 0.006128263659775257, + 0.004265001509338617, + 0.006465997081249952, + 0.005607075989246368, + 0.01935933157801628, + 0.022547107189893723, + 0.016148243099451065, + 0.0254710353910923, + 0.029911871999502182, + 0.021930823102593422, + 0.028139684349298477, + 0.022895807400345802, + 0.006802501622587442, + 0.006967016030102968, + 0.014399426057934761, + 0.01223988737910986, + 0.010464034974575043, + 0.006817776244133711, + 0.005454936996102333, + 0.0048737251199781895, + 0.0057610757648944855, + 0.0060461354441940784, + 0.005183284170925617, + 0.003916075453162193, + 0.004753881599754095, + 0.005416611209511757, + 0.00526798889040947, + 0.005600132513791323, + 0.00903733167797327, + 0.010731304064393044, + 0.010155532509088516, + 0.014162075705826283, + 0.015083982609212399, + 0.011637832969427109, + 0.010553322732448578, + 0.006821541581302881, + 0.012183536775410175, + 0.014503853395581245, + 0.010593772865831852, + 0.012121106497943401, + 0.012168416753411293, + 0.010270807892084122, + 0.009946051053702831, + 0.011233632452785969, + 0.008385032415390015, + 0.018776120617985725, + 0.008577282540500164, + 0.015278156846761703, + 0.008353271521627903, + 0.01089995726943016, + 0.01829318143427372, + 0.011032583191990852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.12703189253807068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows eight majestic horses, rendered in thick, fluid ink brushstrokes, as they gallop with unbridled force. Each horse's muscular lines are imbued with explosive power, as if shaped by the finest calligraphy. Their manes and tails swirl like wisps of ink in the gusting wind, their gestures distinct but their purpose unified. They charge into an ancient village ablaze. The raging flames, the only source of light, paint the sky an ominous orange-red. The village's wooden buildings are reduced to swaying black silhouettes in the flames, while thick black smoke billows and rises like spilled ink. The blazing light illuminates the horses' sleek, inky bodies, casting shadows on the ground beneath their feet that are elongated, sharp, and wildly dance with the flames. The horses' hooves stir up sparks and dust, blending seamlessly with the ink's ethereal and ethereal effects. The entire picture constitutes a dynamic ink painting full of dramatic interweaving of light and shadow, coexistence of power and destruction, creating an epic, tense and magnificent atmosphere.", + "seed": 1150, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_15/index_4308_seed_1150_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": -0.05054443465778604, + "vlm_logprob": -0.05054443465778604, + "vlm_yes_no_margin": 3.999999976158142, + "vlm_token_credit": -0.05054443465778604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14855583012104034 + ], + "opposite_logprob_tokens": [ + -2.0235557556152344 + ], + "logprob_reward": -0.14855583012104034, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14855583012104034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10833480209112167, + 0.11631016433238983, + 0.09073762595653534, + 0.08034723997116089, + 0.08035164326429367, + 0.07667391002178192, + 0.09872101992368698, + 0.11429667472839355, + 0.20746572315692902, + 0.026761217042803764 + ], + "credit_map": [ + 0.01274429727345705, + 0.010622209869325161, + 0.0207921601831913, + 0.012085702270269394, + 0.011728341691195965, + 0.0106388283893466, + 0.015086449682712555, + 0.0146368108689785, + 0.013840248808264732, + 0.018713384866714478, + 0.012596379034221172, + 0.012160548940300941, + 0.021688763052225113, + 0.012344061397016048, + 0.013358251191675663, + 0.011608527973294258, + 0.007570294663310051, + 0.01540627796202898, + 0.016330275684595108, + 0.010738932527601719, + 0.010278292000293732, + 0.01030611339956522, + 0.009977662935853004, + 0.010129778645932674, + 0.009375912137329578, + 0.017078129574656487, + 0.005652768071740866, + 0.005637612659484148, + 0.020740779116749763, + 0.009793231263756752, + 0.006086613517254591, + 0.005982189439237118, + 0.007407312281429768, + 0.008656054735183716, + 0.01022348739206791, + 0.012194786220788956, + 0.009222672320902348, + 0.010183765552937984, + 0.014430715702474117, + 0.008032849058508873, + 0.00531264953315258, + 0.02295638434588909, + 0.0048017119988799095, + 0.006267320830374956, + 0.022793876007199287, + 0.005774882156401873, + 0.0039127906784415245, + 0.00485429260879755, + 0.00875952746719122, + 0.01059326808899641, + 0.0032848920673131943, + 0.002756024245172739, + 0.040732383728027344, + 0.007456890773028135, + 0.0037644824478775263, + 0.021373553201556206, + 0.01006655115634203, + 0.00753199914470315, + 0.008858161978423595, + 0.012048273347318172, + 0.021289097145199776, + 0.012770585715770721, + 0.020008932799100876, + 0.021723072975873947, + 0.022060059010982513, + 0.02465832233428955, + 0.04050173982977867, + 0.020404132083058357, + 0.015313096344470978, + 0.021713707596063614, + 0.022054901346564293, + 0.04075976088643074, + 0.0024647072423249483, + 0.0032568201422691345, + 0.0036692905705422163, + 0.0039501311257481575, + 0.0031041705515235662, + 0.003044083248823881, + 0.0034191494341939688, + 0.0038528665900230408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.14855583012104034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964931100606918 + ], + "opposite_logprob_tokens": [ + -5.130964756011963 + ], + "logprob_reward": -0.005964931100606918, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964931100606918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14542871713638306, + 0.30806562304496765, + 0.09272090345621109, + 0.10543930530548096, + 0.04707876592874527, + 0.03129011392593384, + 0.11390290409326553, + 0.10063206404447556, + 0.038131676614284515, + 0.017309902235865593 + ], + "credit_map": [ + 0.014958986081182957, + 0.019082484766840935, + 0.023380132392048836, + 0.016957849264144897, + 0.017067570239305496, + 0.014315875247120857, + 0.017190704122185707, + 0.0224751103669405, + 0.02085818536579609, + 0.04059428721666336, + 0.03478590399026871, + 0.041291482746601105, + 0.031399041414260864, + 0.03857511281967163, + 0.07261625677347183, + 0.027945369482040405, + 0.01656576804816723, + 0.018851762637495995, + 0.011097395792603493, + 0.007112185936421156, + 0.009002395905554295, + 0.014962765388190746, + 0.008438643999397755, + 0.006689990870654583, + 0.013639996759593487, + 0.017360616475343704, + 0.008728762157261372, + 0.009438962675631046, + 0.021880989894270897, + 0.01246601715683937, + 0.009500457905232906, + 0.0124235013499856, + 0.005244632251560688, + 0.0038885835092514753, + 0.007218200713396072, + 0.006172436289489269, + 0.005222272593528032, + 0.006283618044108152, + 0.008064079098403454, + 0.0049849459901452065, + 0.0032239726278930902, + 0.00936700776219368, + 0.002296942286193371, + 0.0021322232205420732, + 0.007834037765860558, + 0.0026587615720927715, + 0.0017427430720999837, + 0.0020344266667962074, + 0.011898169293999672, + 0.012873685918748379, + 0.003991649020463228, + 0.0038432825822383165, + 0.04437798261642456, + 0.009758341126143932, + 0.00696610426530242, + 0.020193688571453094, + 0.007194393314421177, + 0.006696760188788176, + 0.008645524270832539, + 0.011192056350409985, + 0.018566757440567017, + 0.01199648343026638, + 0.01712384819984436, + 0.019216232001781464, + 0.002846749732270837, + 0.00784892775118351, + 0.0036041231360286474, + 0.0036164356861263514, + 0.003866987070068717, + 0.004812732804566622, + 0.0036078256089240313, + 0.007927893660962582, + 0.0018035771790891886, + 0.0016051633283495903, + 0.002779794856905937, + 0.0021956535056233406, + 0.0016415984136983752, + 0.002380941528826952, + 0.001982246758416295, + 0.0029209256172180176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.005964931100606918 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486447848379612 + ], + "opposite_logprob_tokens": [ + -3.048644781112671 + ], + "logprob_reward": -0.0486447848379612, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486447848379612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12846533954143524, + 0.09275821596384048, + 0.2146158218383789, + 0.04020598903298378, + 0.05905850604176521, + 0.05403124913573265, + 0.12871615588665009, + 0.15200664103031158, + 0.06427914649248123, + 0.0658629760146141 + ], + "credit_map": [ + 0.013104305602610111, + 0.012918391264975071, + 0.019741469994187355, + 0.011519112624228, + 0.01735641621053219, + 0.014682536013424397, + 0.017746487632393837, + 0.02139660343527794, + 0.009265927597880363, + 0.01042716484516859, + 0.012256044894456863, + 0.011492977850139141, + 0.017597021535038948, + 0.009214776568114758, + 0.013116374611854553, + 0.009387927129864693, + 0.03508540615439415, + 0.022869059816002846, + 0.044835347682237625, + 0.018748199567198753, + 0.022079093381762505, + 0.02690666913986206, + 0.02526390552520752, + 0.01882810704410076, + 0.005688222125172615, + 0.005677198059856892, + 0.0035519502125680447, + 0.0037539913319051266, + 0.006839239504188299, + 0.006320503074675798, + 0.004046360030770302, + 0.004328523296862841, + 0.007274406496435404, + 0.01029130443930626, + 0.006852286867797375, + 0.0053520300425589085, + 0.00868160929530859, + 0.007707640528678894, + 0.006109423004090786, + 0.006789799779653549, + 0.003580583492293954, + 0.015518894419074059, + 0.003458102000877261, + 0.004311473108828068, + 0.015317522920668125, + 0.004932687617838383, + 0.0032718295697122812, + 0.003640152048319578, + 0.010205890983343124, + 0.013224144466221333, + 0.0042022159323096275, + 0.0029511533211916685, + 0.052649371325969696, + 0.008878613822162151, + 0.005619420204311609, + 0.030985334888100624, + 0.01897391490638256, + 0.010702227242290974, + 0.012293970212340355, + 0.014057465828955173, + 0.02493424154818058, + 0.017511136829853058, + 0.024808308109641075, + 0.02872537262737751, + 0.005871250759810209, + 0.009181356057524681, + 0.006652223411947489, + 0.00614203279837966, + 0.00816496554762125, + 0.010679473169147968, + 0.008343013934791088, + 0.009244824759662151, + 0.010054947808384895, + 0.008753182366490364, + 0.008766748942434788, + 0.008455164730548859, + 0.006796366535127163, + 0.0081718098372221, + 0.007552258670330048, + 0.007312493864446878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.0486447848379612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009180859779007733 + ], + "opposite_logprob_tokens": [ + -7.000917911529541 + ], + "logprob_reward": -0.0009180859779007733, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009180859779007733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13235624134540558, + 0.07340161502361298, + 0.10516318678855896, + 0.09230691939592361, + 0.17382369935512543, + 0.07809504866600037, + 0.1115383431315422, + 0.1397494524717331, + 0.047034818679094315, + 0.04653071239590645 + ], + "credit_map": [ + 0.013617766089737415, + 0.01470937766134739, + 0.020543057471513748, + 0.014036889187991619, + 0.016761839389801025, + 0.016056742519140244, + 0.01776900514960289, + 0.018861547112464905, + 0.007380868308246136, + 0.009578493423759937, + 0.007161264307796955, + 0.008023344911634922, + 0.010640319436788559, + 0.009407171979546547, + 0.01202031597495079, + 0.009189832024276257, + 0.00999588705599308, + 0.015787484124302864, + 0.021315403282642365, + 0.008666844107210636, + 0.010436643846333027, + 0.0181281641125679, + 0.010928691364824772, + 0.009904062375426292, + 0.011548412963747978, + 0.01444490347057581, + 0.007839739322662354, + 0.007394199725240469, + 0.023266328498721123, + 0.011066768318414688, + 0.008251214399933815, + 0.008495347574353218, + 0.0173899307847023, + 0.04748481139540672, + 0.013531460426747799, + 0.010999517515301704, + 0.023570161312818527, + 0.03626590594649315, + 0.012160489335656166, + 0.012421426363289356, + 0.004865952301770449, + 0.02374153584241867, + 0.002610902301967144, + 0.004005568102002144, + 0.03125040978193283, + 0.006123645231127739, + 0.002826416166499257, + 0.0026706161443144083, + 0.009615810588002205, + 0.012230810709297657, + 0.005833988543599844, + 0.004506472032517195, + 0.041657961905002594, + 0.00968874804675579, + 0.0074662151746451855, + 0.020538337528705597, + 0.01732240431010723, + 0.009402300231158733, + 0.01173462625592947, + 0.014707897789776325, + 0.021148281171917915, + 0.01569327525794506, + 0.023635366931557655, + 0.026105310767889023, + 0.005150133743882179, + 0.0053847008384764194, + 0.007909434847533703, + 0.004859219305217266, + 0.0052007585763931274, + 0.00592071283608675, + 0.004622072447091341, + 0.007987788878381252, + 0.0036749443970620632, + 0.005158828105777502, + 0.008024548180401325, + 0.006385984364897013, + 0.003921933937817812, + 0.004979185294359922, + 0.006034758407622576, + 0.008350527845323086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.0009180859779007733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048638541251420975 + ], + "opposite_logprob_tokens": [ + -3.0486385822296143 + ], + "logprob_reward": -0.048638541251420975, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048638541251420975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06283976137638092, + 0.2519124448299408, + 0.052325427532196045, + 0.05135836452245712, + 0.2744579315185547, + 0.07468409836292267, + 0.058538250625133514, + 0.07393188029527664, + 0.08219880610704422, + 0.017753073945641518 + ], + "credit_map": [ + 0.006809119135141373, + 0.006515301298350096, + 0.011640621349215508, + 0.00716799171641469, + 0.0069872853346168995, + 0.006691174115985632, + 0.009321024641394615, + 0.007707235869020224, + 0.020217295736074448, + 0.035025134682655334, + 0.027519477531313896, + 0.03217865899205208, + 0.02234390564262867, + 0.03320877254009247, + 0.05597763881087303, + 0.025441540405154228, + 0.0045450651086866856, + 0.00696324510499835, + 0.010552062653005123, + 0.005894629750400782, + 0.0047890059649944305, + 0.008057888597249985, + 0.0042039272375404835, + 0.007319597527384758, + 0.007699082605540752, + 0.010385704226791859, + 0.0042367433197796345, + 0.0032089892774820328, + 0.008200341835618019, + 0.010104247368872166, + 0.004014537204056978, + 0.0035087161231786013, + 0.025934047996997833, + 0.016839951276779175, + 0.034787435084581375, + 0.046444784849882126, + 0.025586498901247978, + 0.05269981920719147, + 0.04281272739171982, + 0.02935265377163887, + 0.004476830828934908, + 0.026292474940419197, + 0.0021319661755114794, + 0.0032894292380660772, + 0.027288902550935745, + 0.0060348715633153915, + 0.002318090759217739, + 0.0028515232261270285, + 0.004766572266817093, + 0.006164167076349258, + 0.002757556037977338, + 0.002911714371293783, + 0.021718842908740044, + 0.0050353072583675385, + 0.003789664478972554, + 0.011394423432648182, + 0.010641845874488354, + 0.005124910268932581, + 0.006397542543709278, + 0.006586826406419277, + 0.01125421840697527, + 0.00977399479597807, + 0.011196608655154705, + 0.01295593474060297, + 0.0065956758335232735, + 0.006195481866598129, + 0.010386410169303417, + 0.010803751647472382, + 0.010617162100970745, + 0.011903410777449608, + 0.013555203564465046, + 0.01214171014726162, + 0.0017021544044837356, + 0.002113083377480507, + 0.002753742504864931, + 0.0032188836485147476, + 0.0015887789195403457, + 0.0020921151153743267, + 0.001786532229743898, + 0.002497783163562417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.048638541251420975 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows eight majestic horses, rendered in thick, fluid ink brushstrokes, as they gallop with unbridled force. Each horse's muscular lines are imbued with explosive power, as if shaped by the finest calligraphy. Their manes and tails swirl like wisps of ink in the gusting wind, their gestures distinct but their purpose unified. They charge into an ancient village ablaze. The raging flames, the only source of light, paint the sky an ominous orange-red. The village's wooden buildings are reduced to swaying black silhouettes in the flames, while thick black smoke billows and rises like spilled ink. The blazing light illuminates the horses' sleek, inky bodies, casting shadows on the ground beneath their feet that are elongated, sharp, and wildly dance with the flames. The horses' hooves stir up sparks and dust, blending seamlessly with the ink's ethereal and ethereal effects. The entire picture constitutes a dynamic ink painting full of dramatic interweaving of light and shadow, coexistence of power and destruction, creating an epic, tense and magnificent atmosphere.", + "seed": 1151, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_15/index_4308_seed_1151_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": -0.06615554962772877, + "vlm_logprob": -0.06615554962772877, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.06615554962772877, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13448071479797363 + ], + "opposite_logprob_tokens": [ + -2.1344807147979736 + ], + "logprob_reward": -0.13448071479797363, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13448071479797363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1884913444519043, + 0.06651797890663147, + 0.05650225281715393, + 0.09580184519290924, + 0.24078308045864105, + 0.034576840698719025, + 0.09413066506385803, + 0.057817332446575165, + 0.0840705931186676, + 0.08130805939435959 + ], + "credit_map": [ + 0.022409088909626007, + 0.029999004676938057, + 0.016714608296751976, + 0.022501589730381966, + 0.019613303244113922, + 0.029507804661989212, + 0.022607848048210144, + 0.025138095021247864, + 0.007827947847545147, + 0.005956242326647043, + 0.009045940823853016, + 0.007338276132941246, + 0.010286848992109299, + 0.009557188488543034, + 0.008579196408390999, + 0.00792633555829525, + 0.004653383977711201, + 0.010919144377112389, + 0.0065474859438836575, + 0.006921548396348953, + 0.0061226049438118935, + 0.007136829663068056, + 0.007625579833984375, + 0.0065756747499108315, + 0.009427373297512531, + 0.011910553090274334, + 0.011534268967807293, + 0.014127491973340511, + 0.01867033541202545, + 0.009846637025475502, + 0.010282911360263824, + 0.010002274997532368, + 0.029844971373677254, + 0.016880128532648087, + 0.019530611112713814, + 0.053986694663763046, + 0.05504254624247551, + 0.033844590187072754, + 0.015504587441682816, + 0.01614895462989807, + 0.005046680569648743, + 0.003998036962002516, + 0.0034006244968622923, + 0.004216366447508335, + 0.00552435452118516, + 0.004527321085333824, + 0.0029369427356868982, + 0.004926512949168682, + 0.013621377758681774, + 0.008654871955513954, + 0.009107351303100586, + 0.014892643317580223, + 0.011236376129090786, + 0.01644039899110794, + 0.009561688639223576, + 0.01061596255749464, + 0.004695381037890911, + 0.00579182431101799, + 0.00689993891865015, + 0.00811643898487091, + 0.005874160211533308, + 0.007490368094295263, + 0.01099406648427248, + 0.007955153472721577, + 0.0067613618448376656, + 0.009657805785536766, + 0.011030593886971474, + 0.013311698101460934, + 0.009558149613440037, + 0.011248492635786533, + 0.012199743650853634, + 0.010302749462425709, + 0.006911286152899265, + 0.008466250263154507, + 0.006704532541334629, + 0.01026998646557331, + 0.010798740200698376, + 0.010971336625516415, + 0.0156925730407238, + 0.011493354104459286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.13448071479797363 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270515102893114 + ], + "opposite_logprob_tokens": [ + -5.255270481109619 + ], + "logprob_reward": -0.005270515102893114, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270515102893114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16509732604026794, + 0.04427829757332802, + 0.027575144544243813, + 0.03176015242934227, + 0.49980154633522034, + 0.03413596376776695, + 0.026212628930807114, + 0.03614940866827965, + 0.03873981907963753, + 0.09624969959259033 + ], + "credit_map": [ + 0.034574948251247406, + 0.061210647225379944, + 0.00870424509048462, + 0.011670668609440327, + 0.012352874502539635, + 0.015254884958267212, + 0.01064230501651764, + 0.010686743073165417, + 0.006593126803636551, + 0.004445821978151798, + 0.006126969121396542, + 0.005483637563884258, + 0.005340166389942169, + 0.005343963857740164, + 0.005962076596915722, + 0.004982536658644676, + 0.0029383674263954163, + 0.004070884082466364, + 0.003490705508738756, + 0.003700475674122572, + 0.0038190404884517193, + 0.002821017289534211, + 0.0026164983864873648, + 0.004118154291063547, + 0.004658397752791643, + 0.004146832972764969, + 0.0040376028046011925, + 0.004083639942109585, + 0.0037303133867681026, + 0.004082139115780592, + 0.003338918322697282, + 0.003682309528812766, + 0.0669274851679802, + 0.03693096339702606, + 0.039391595870256424, + 0.10699746012687683, + 0.14030137658119202, + 0.05849459767341614, + 0.026848910376429558, + 0.023909160867333412, + 0.004741308744996786, + 0.0033380913082510233, + 0.0029686689376831055, + 0.0029595920350402594, + 0.00923825241625309, + 0.004109629429876804, + 0.002773141022771597, + 0.004007278475910425, + 0.003989106509834528, + 0.0019971616566181183, + 0.0022349704522639513, + 0.0022668750025331974, + 0.0030240328051149845, + 0.008820783346891403, + 0.0018579643219709396, + 0.0020217339042574167, + 0.005032053682953119, + 0.005341825541108847, + 0.0038965214043855667, + 0.004334050230681896, + 0.0055905855260789394, + 0.005355592351406813, + 0.003729115938767791, + 0.002869663294404745, + 0.0028900345787405968, + 0.003210460999980569, + 0.005439383909106255, + 0.005481967236846685, + 0.005344982258975506, + 0.00532586919143796, + 0.006474512163549662, + 0.004572609439492226, + 0.009618045762181282, + 0.012785904109477997, + 0.007695174776017666, + 0.010638099163770676, + 0.00955633819103241, + 0.011952069588005543, + 0.017425628378987312, + 0.016578434035182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.005270515102893114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029608756303787 + ], + "opposite_logprob_tokens": [ + -2.3502960205078125 + ], + "logprob_reward": -0.10029608756303787, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029608756303787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2285844087600708, + 0.05702444538474083, + 0.04840746894478798, + 0.09499209374189377, + 0.1470867246389389, + 0.06228943169116974, + 0.06284774094820023, + 0.05495704337954521, + 0.17551884055137634, + 0.0682918131351471 + ], + "credit_map": [ + 0.04174334183335304, + 0.08223039656877518, + 0.012549937702715397, + 0.015132730826735497, + 0.012385117821395397, + 0.03241134434938431, + 0.01369635108858347, + 0.01843518763780594, + 0.007526771165430546, + 0.0057083433493971825, + 0.008762559853494167, + 0.008587993681430817, + 0.006195827852934599, + 0.006457454990595579, + 0.007757063489407301, + 0.006028429605066776, + 0.004954198841005564, + 0.0057425289414823055, + 0.00493233185261488, + 0.005700209178030491, + 0.004809575155377388, + 0.0057653882540762424, + 0.0069953324273228645, + 0.009507901035249233, + 0.013731296174228191, + 0.017057565972208977, + 0.008937017060816288, + 0.01179245114326477, + 0.017997868359088898, + 0.009891078807413578, + 0.007439218927174807, + 0.008145599626004696, + 0.05047502741217613, + 0.011867888271808624, + 0.008667275309562683, + 0.008988505229353905, + 0.0439167246222496, + 0.010261022485792637, + 0.006863126531243324, + 0.006047157105058432, + 0.010526045225560665, + 0.00625963881611824, + 0.004324383568018675, + 0.00414804695174098, + 0.016970915719866753, + 0.0075933728367090225, + 0.004178817383944988, + 0.008288211189210415, + 0.012085366062819958, + 0.004448229447007179, + 0.003788005094975233, + 0.004000543151050806, + 0.007715573534369469, + 0.019291404634714127, + 0.006451904773712158, + 0.005066707730293274, + 0.007761191111057997, + 0.008239885792136192, + 0.005879500415176153, + 0.005280245095491409, + 0.007514645345509052, + 0.008617863059043884, + 0.0056524574756622314, + 0.006011251360177994, + 0.015908412635326385, + 0.02063889056444168, + 0.023263778537511826, + 0.027503034099936485, + 0.020397858694195747, + 0.023223280906677246, + 0.0243656188249588, + 0.020217953249812126, + 0.005173380021005869, + 0.004660532344132662, + 0.008611319586634636, + 0.01255526952445507, + 0.004596714861690998, + 0.007691626902669668, + 0.01704065315425396, + 0.007962314411997795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.10029608756303787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017116909148171544 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017116909148171544, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017116909148171544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13980621099472046, + 0.10927659273147583, + 0.06436949968338013, + 0.06227551028132439, + 0.297347754240036, + 0.04657286778092384, + 0.047093458473682404, + 0.053194500505924225, + 0.12581117451190948, + 0.054252494126558304 + ], + "credit_map": [ + 0.023793336004018784, + 0.017728284001350403, + 0.010971044190227985, + 0.013098005205392838, + 0.014157703146338463, + 0.018866343423724174, + 0.016398204490542412, + 0.024793269112706184, + 0.02235799841582775, + 0.01128996443003416, + 0.016653381288051605, + 0.011973811313509941, + 0.010499216616153717, + 0.013617077842354774, + 0.012467017397284508, + 0.010418107733130455, + 0.005440657492727041, + 0.007607656996697187, + 0.007412774953991175, + 0.007035264745354652, + 0.008426829241216183, + 0.008347458206117153, + 0.010262537747621536, + 0.00983631331473589, + 0.008940039202570915, + 0.0078104836866259575, + 0.008414902724325657, + 0.006062345579266548, + 0.00864003598690033, + 0.00731613440439105, + 0.007732229772955179, + 0.0073593370616436005, + 0.10317834466695786, + 0.023952627554535866, + 0.016850920394062996, + 0.016081972047686577, + 0.091490238904953, + 0.020690618082880974, + 0.013348512351512909, + 0.011754492297768593, + 0.00815780833363533, + 0.00547670666128397, + 0.004204897675663233, + 0.005148106254637241, + 0.009442394599318504, + 0.0071010226383805275, + 0.002888708608224988, + 0.004153216723352671, + 0.0056550996378064156, + 0.0058090221136808395, + 0.005194745026528835, + 0.005606816150248051, + 0.006529155187308788, + 0.009732572361826897, + 0.004850216209888458, + 0.003715833881869912, + 0.007339757401496172, + 0.006235572509467602, + 0.006458819843828678, + 0.006238441914319992, + 0.005811075679957867, + 0.007747108582407236, + 0.007427624426782131, + 0.005936093162745237, + 0.017410220578312874, + 0.015910862013697624, + 0.01310050766915083, + 0.019270118325948715, + 0.017304273322224617, + 0.014710526913404465, + 0.016055351123213768, + 0.012049310840666294, + 0.004023138899356127, + 0.008209871128201485, + 0.005875893868505955, + 0.0070796143263578415, + 0.005707156844437122, + 0.006180895492434502, + 0.011102288030087948, + 0.006073631811887026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0017116909148171544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901873975992203 + ], + "opposite_logprob_tokens": [ + -2.4640188217163086 + ], + "logprob_reward": -0.08901873975992203, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901873975992203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22966356575489044, + 0.0782516747713089, + 0.06767741590738297, + 0.12793390452861786, + 0.1446676105260849, + 0.04824966937303543, + 0.07926423847675323, + 0.020966077223420143, + 0.08849725872278214, + 0.11482865363359451 + ], + "credit_map": [ + 0.02817084826529026, + 0.07370606064796448, + 0.01499631255865097, + 0.01638823188841343, + 0.014029919169843197, + 0.054981671273708344, + 0.011564657092094421, + 0.015825873240828514, + 0.012653348967432976, + 0.007688693702220917, + 0.009978330694139004, + 0.010552097111940384, + 0.010929158888757229, + 0.009277681820094585, + 0.009530496783554554, + 0.007641867268830538, + 0.0088327182456851, + 0.012205006554722786, + 0.006232726387679577, + 0.008476262912154198, + 0.008585932664573193, + 0.006863447371870279, + 0.008040732704102993, + 0.008440590463578701, + 0.01726793870329857, + 0.018915845081210136, + 0.0127381831407547, + 0.01222732849419117, + 0.02498345635831356, + 0.013899752870202065, + 0.013655845075845718, + 0.014245557598769665, + 0.02818343974649906, + 0.01113185565918684, + 0.016811946406960487, + 0.017201818525791168, + 0.03647896647453308, + 0.012162672355771065, + 0.011221328750252724, + 0.011475572362542152, + 0.010931262746453285, + 0.004767599981278181, + 0.003218103665858507, + 0.004328643903136253, + 0.007074730470776558, + 0.006581691559404135, + 0.0033067557960748672, + 0.008040882647037506, + 0.014141072519123554, + 0.005315415095537901, + 0.004708702210336924, + 0.006781787611544132, + 0.008526960387825966, + 0.027646971866488457, + 0.006988083012402058, + 0.005155239254236221, + 0.0032338667660951614, + 0.0019442230695858598, + 0.002433700254186988, + 0.0020149629563093185, + 0.00244545447640121, + 0.0032631263602524996, + 0.003328928956761956, + 0.0023018133360892534, + 0.010069342330098152, + 0.009605949744582176, + 0.012577262707054615, + 0.009688868187367916, + 0.015133613720536232, + 0.010449334979057312, + 0.01255930494517088, + 0.008413584902882576, + 0.008455758914351463, + 0.010510765016078949, + 0.010250816121697426, + 0.018649030476808548, + 0.013009104877710342, + 0.014274660497903824, + 0.024885766208171844, + 0.014792757108807564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.08901873975992203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows eight majestic horses, rendered in thick, fluid ink brushstrokes, as they gallop with unbridled force. Each horse's muscular lines are imbued with explosive power, as if shaped by the finest calligraphy. Their manes and tails swirl like wisps of ink in the gusting wind, their gestures distinct but their purpose unified. They charge into an ancient village ablaze. The raging flames, the only source of light, paint the sky an ominous orange-red. The village's wooden buildings are reduced to swaying black silhouettes in the flames, while thick black smoke billows and rises like spilled ink. The blazing light illuminates the horses' sleek, inky bodies, casting shadows on the ground beneath their feet that are elongated, sharp, and wildly dance with the flames. The horses' hooves stir up sparks and dust, blending seamlessly with the ink's ethereal and ethereal effects. The entire picture constitutes a dynamic ink painting full of dramatic interweaving of light and shadow, coexistence of power and destruction, creating an epic, tense and magnificent atmosphere.", + "seed": 1152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_15/index_4308_seed_1152_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": -0.06275531116407365, + "vlm_logprob": -0.06275531116407365, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.06275531116407365, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18498100340366364 + ], + "opposite_logprob_tokens": [ + -1.8099809885025024 + ], + "logprob_reward": -0.18498100340366364, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18498100340366364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43184030055999756, + 0.011878639459609985, + 0.05925239622592926, + 0.02261005900800228, + 0.2830881178379059, + 0.03806550055742264, + 0.04499387368559837, + 0.043886035680770874, + 0.023221630603075027, + 0.04116346314549446 + ], + "credit_map": [ + 0.0861947312951088, + 0.04485437646508217, + 0.03855229914188385, + 0.02059232071042061, + 0.06119956448674202, + 0.029782971367239952, + 0.03198743611574173, + 0.11867658793926239, + 0.0012576711596921086, + 0.0015504540642723441, + 0.001761482679285109, + 0.0012977393344044685, + 0.001818701159209013, + 0.0013828439405187964, + 0.0015107447979971766, + 0.0012990026734769344, + 0.014452597126364708, + 0.004486319608986378, + 0.005413379054516554, + 0.004558082204312086, + 0.01576954685151577, + 0.006923249457031488, + 0.004127573687583208, + 0.0035216466058045626, + 0.0031314147636294365, + 0.00260135973803699, + 0.0036417029332369566, + 0.0033104203175753355, + 0.002933540614321828, + 0.0018668973352760077, + 0.0027729433495551348, + 0.0023517811205238104, + 0.0714237168431282, + 0.011468113400042057, + 0.06994973868131638, + 0.012401396408677101, + 0.02681129239499569, + 0.0625406876206398, + 0.010270928964018822, + 0.018222257494926453, + 0.003853554604575038, + 0.004648954141885042, + 0.004422082100063562, + 0.0033964444883167744, + 0.007739816792309284, + 0.006954075302928686, + 0.004349912516772747, + 0.002700659679248929, + 0.00275298859924078, + 0.0022558351047337055, + 0.011933311820030212, + 0.01697775535285473, + 0.0015074016992002726, + 0.004681620746850967, + 0.0026565329171717167, + 0.0022284246515482664, + 0.0026063735131174326, + 0.0024356450885534286, + 0.003312750719487667, + 0.0038668031338602304, + 0.0028405135963112116, + 0.011303640902042389, + 0.01438602339476347, + 0.0031342857982963324, + 0.0034258542582392693, + 0.0040464093908667564, + 0.002537198830395937, + 0.00194948585703969, + 0.002731931395828724, + 0.0024314147885888815, + 0.0022045851219445467, + 0.003894749330356717, + 0.005108625162392855, + 0.003870415734127164, + 0.006190163549035788, + 0.005072610452771187, + 0.005640705116093159, + 0.003410617122426629, + 0.0064450702629983425, + 0.0054252538830041885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.18498100340366364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052668387070298195 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.0052668387070298195, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052668387070298195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39418187737464905, + 0.06151833385229111, + 0.12631180882453918, + 0.05654867738485336, + 0.15685459971427917, + 0.05544087290763855, + 0.07160373777151108, + 0.019419919699430466, + 0.01741034910082817, + 0.040709782391786575 + ], + "credit_map": [ + 0.07738549262285233, + 0.03761691227555275, + 0.0349716916680336, + 0.02066236175596714, + 0.05955751985311508, + 0.02468245103955269, + 0.03007553517818451, + 0.1092299222946167, + 0.007875707000494003, + 0.009289192035794258, + 0.017763890326023102, + 0.005394565407186747, + 0.006435642950236797, + 0.0046515450812876225, + 0.005451972596347332, + 0.004655825439840555, + 0.029509760439395905, + 0.011470441706478596, + 0.012350657023489475, + 0.011477993801236153, + 0.02708307094871998, + 0.015923766419291496, + 0.010153341107070446, + 0.008342808112502098, + 0.009712676517665386, + 0.008289558812975883, + 0.00760697852820158, + 0.006636884529143572, + 0.006017384584993124, + 0.005164287984371185, + 0.006744466722011566, + 0.006376444362103939, + 0.032274529337882996, + 0.007856405340135098, + 0.015120865777134895, + 0.005445878487080336, + 0.02052498795092106, + 0.057316478341817856, + 0.006511596497148275, + 0.011803875677287579, + 0.0048795356415212154, + 0.006546358577907085, + 0.0063758729957044125, + 0.00780743733048439, + 0.008233598433434963, + 0.0075270445086061954, + 0.008465474471449852, + 0.005605555605143309, + 0.004931634292006493, + 0.004846430849283934, + 0.02713968977332115, + 0.020243462175130844, + 0.0037464534398168325, + 0.002267400501295924, + 0.0034486514050513506, + 0.004980020225048065, + 0.001653182553127408, + 0.0010371741373091936, + 0.002606624038890004, + 0.002011818578466773, + 0.0011423988034948707, + 0.0025967301335185766, + 0.005586980376392603, + 0.0027850137557834387, + 0.002534182509407401, + 0.0015036059776321054, + 0.003007427789270878, + 0.0017805909737944603, + 0.0022363916505128145, + 0.001621217466890812, + 0.0020948874298483133, + 0.002632048912346363, + 0.004885889124125242, + 0.0052245864644646645, + 0.0046862647868692875, + 0.00500840786844492, + 0.005197129212319851, + 0.005174430552870035, + 0.004445320460945368, + 0.00608776044100523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0052668387070298195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.0724618211388588, + 0.021851826459169388, + 0.0201927088201046, + 0.01939838007092476, + 0.7692214250564575, + 0.03316698595881462, + 0.008700553327798843, + 0.029496939852833748, + 0.016240302473306656, + 0.009269092231988907 + ], + "credit_map": [ + 0.011511807329952717, + 0.006959116086363792, + 0.008173794485628605, + 0.005613364744931459, + 0.009689172729849815, + 0.00862592551857233, + 0.0045709130354225636, + 0.01731773279607296, + 0.0026495237834751606, + 0.003965528216212988, + 0.005664218682795763, + 0.0015567796071991324, + 0.0024124733172357082, + 0.0015036900294944644, + 0.0018903423333540559, + 0.002209270605817437, + 0.004399233963340521, + 0.0019296430982649326, + 0.0019980710931122303, + 0.001924990676343441, + 0.0048724631778895855, + 0.0020482700783759356, + 0.001439201063476503, + 0.0015808368334546685, + 0.0033343317918479443, + 0.0030841524712741375, + 0.002473898697644472, + 0.0022278521209955215, + 0.002846544375643134, + 0.0017668548971414566, + 0.001603290787898004, + 0.0020614550448954105, + 0.19660255312919617, + 0.031319510191679, + 0.19257888197898865, + 0.03411806374788284, + 0.0718412846326828, + 0.16633206605911255, + 0.028097648173570633, + 0.048331424593925476, + 0.004126700572669506, + 0.004734447691589594, + 0.0042180554009974, + 0.004012812860310078, + 0.003951724153012037, + 0.003957539331167936, + 0.005250238813459873, + 0.0029154678341001272, + 0.0006558452732861042, + 0.0011184824397787452, + 0.0005843554972670972, + 0.0027057265397161245, + 0.0007957934285514057, + 0.0011591751826927066, + 0.0009706172859296203, + 0.000710557505954057, + 0.001703456393443048, + 0.00159913906827569, + 0.002037661150097847, + 0.002102646278217435, + 0.0019465053919702768, + 0.008304593153297901, + 0.009529715403914452, + 0.0022732203360646963, + 0.003017115406692028, + 0.0019279284169897437, + 0.0025405860505998135, + 0.0012876413529738784, + 0.001463004620745778, + 0.0010798529256135225, + 0.0015453817322850227, + 0.003378791268914938, + 0.0011560390703380108, + 0.0009733042679727077, + 0.0008428737637586892, + 0.0010465987725183368, + 0.0016735042445361614, + 0.000871692318469286, + 0.0014756795717403293, + 0.0012293991167098284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "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\": \"Are eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001510551548562944 + ], + "opposite_logprob_tokens": [ + -6.5015106201171875 + ], + "logprob_reward": -0.001510551548562944, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001510551548562944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4022659659385681, + 0.04362673684954643, + 0.04266011342406273, + 0.03588169440627098, + 0.3117251396179199, + 0.07199136912822723, + 0.02269561029970646, + 0.016016114503145218, + 0.038701411336660385, + 0.014435810036957264 + ], + "credit_map": [ + 0.06627723574638367, + 0.044756434857845306, + 0.030335472896695137, + 0.023091061040759087, + 0.069420725107193, + 0.025610025972127914, + 0.033335983753204346, + 0.10943905264139175, + 0.005044748540967703, + 0.007308395579457283, + 0.011635045520961285, + 0.003368076868355274, + 0.0046311309561133385, + 0.0032865717075765133, + 0.0036487062461674213, + 0.004704063758254051, + 0.008200242184102535, + 0.0032422642689198256, + 0.003117989283055067, + 0.002890851581469178, + 0.015063511207699776, + 0.00464111752808094, + 0.002871379954740405, + 0.002632756484672427, + 0.007248066831380129, + 0.006243220996111631, + 0.004405153449624777, + 0.004919353406876326, + 0.003930194769054651, + 0.0031945500522851944, + 0.002765567973256111, + 0.0031755876261740923, + 0.08210662752389908, + 0.014656773768365383, + 0.130079984664917, + 0.016882840543985367, + 0.013151305727660656, + 0.027871038764715195, + 0.010676796548068523, + 0.016299787908792496, + 0.011960621923208237, + 0.008430907502770424, + 0.014105791226029396, + 0.007633357308804989, + 0.0051157306879758835, + 0.009174860082566738, + 0.005615066736936569, + 0.009955035522580147, + 0.001981504959985614, + 0.002943704603239894, + 0.001491820439696312, + 0.0058235712349414825, + 0.0019621350802481174, + 0.0033043548464775085, + 0.002187218051403761, + 0.0030013008508831263, + 0.001782542560249567, + 0.0014769758563488722, + 0.0016958306077867746, + 0.0018385166767984629, + 0.0012841627467423677, + 0.002664602128788829, + 0.003561189165338874, + 0.0017122956924140453, + 0.007969536818563938, + 0.003264260245487094, + 0.005689847748726606, + 0.003195319790393114, + 0.0038449792191386223, + 0.0036907754838466644, + 0.0038677225820720196, + 0.007178969215601683, + 0.0025434806011617184, + 0.0015150784747675061, + 0.0015711899613961577, + 0.0014109696494415402, + 0.0014469720190390944, + 0.0019994850736111403, + 0.0021727175917476416, + 0.001775916782207787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.001510551548562944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 eight horses visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in thick, fluid ink brushstrokes resembling calligraphy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the horses galloping and charging into an ancient village that is on fire?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by raging flames casting an ominous orange-red light on the sky and black silhouettes of wooden buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot taken from a low angle while tracking the horses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896366715431213 + ], + "opposite_logprob_tokens": [ + -2.5789637565612793 + ], + "logprob_reward": -0.07896366715431213, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896366715431213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10403800010681152, + 0.06553612649440765, + 0.02622927352786064, + 0.08665981143712997, + 0.3486201763153076, + 0.10711176693439484, + 0.11714895069599152, + 0.048708055168390274, + 0.06982962042093277, + 0.026118215173482895 + ], + "credit_map": [ + 0.014370010234415531, + 0.013715578243136406, + 0.00840494129806757, + 0.010702574625611305, + 0.01530695240944624, + 0.011253445409238338, + 0.008591149933636189, + 0.021693339571356773, + 0.008618348278105259, + 0.009516667574644089, + 0.018523067235946655, + 0.005276129115372896, + 0.00622725673019886, + 0.004318621940910816, + 0.006229236256331205, + 0.006826792843639851, + 0.004645830485969782, + 0.003317735157907009, + 0.0034465512726455927, + 0.0026224302127957344, + 0.0030231375712901354, + 0.004224920645356178, + 0.002628194633871317, + 0.0023204695899039507, + 0.013006036169826984, + 0.012090676464140415, + 0.00893542543053627, + 0.012201099656522274, + 0.01173309050500393, + 0.007711054757237434, + 0.011897525750100613, + 0.009084898978471756, + 0.025267986580729485, + 0.01782589964568615, + 0.07551070302724838, + 0.013471366837620735, + 0.041910041123628616, + 0.13931052386760712, + 0.013355698436498642, + 0.02196793630719185, + 0.012184989638626575, + 0.015006407164037228, + 0.008106958121061325, + 0.0101270517334342, + 0.011706491000950336, + 0.0266546830534935, + 0.014764908701181412, + 0.008560274727642536, + 0.007319330237805843, + 0.00474399933591485, + 0.04639941826462746, + 0.040811359882354736, + 0.004738728050142527, + 0.0034109733533114195, + 0.0052916607819497585, + 0.004433467984199524, + 0.0033062442671507597, + 0.0039020124822854996, + 0.006517619360238314, + 0.0054596541449427605, + 0.003990133758634329, + 0.011287880130112171, + 0.009739978238940239, + 0.004504531156271696, + 0.013367914594709873, + 0.006431033834815025, + 0.009284582920372486, + 0.005832263268530369, + 0.006611271761357784, + 0.0068464335054159164, + 0.007253260817378759, + 0.014202854596078396, + 0.0030159621965140104, + 0.0026909494772553444, + 0.0027280685026198626, + 0.0028115566819906235, + 0.0036263198126107454, + 0.002538262167945504, + 0.004737020004540682, + 0.003970074001699686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.07896366715431213 + } + ], + "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/0000155/group_2/index_tensor([1242, 1242])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_tensor([1242, 1242])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..28b2817f8df2be61a82c99787dd55ae379e00941 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_tensor([1242, 1242])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic wide shot pans slowly to the right from high above, sweeping across a steep, winding, mist-shrouded mountain ridge. Under the low, golden sunlight of early morning, a new section of the Great Wall rises from the ground. The figures of thousands of ancient craftsmen dot the steep slopes and the nascent wall, creating a magnificent scene of labor. Most of them are sturdy men, dressed in worn, faded coarse linen shorts. Their bronzed skin, beaded with sweat, gleams in the sun. As the camera pans down and slowly zooms in, we can see details: a team of artisans, working in unison, chanting slogans, painstakingly drag huge, unpolished granite blocks up the slope using thick hemp ropes and crude wooden rollers. Elsewhere, on bamboo scaffolding erected at the edge of the wall, stonemasons carefully hammer and chisel the edges of the stones, forcing them to fit more tightly, their movements focused and rhythmic. Below, scores of people, carrying heavy stones or buckets filled with mud, navigate the narrow and rugged construction site, their steps stumbling under the weight but unwavering. The entire scene is shrouded in a stark palette of gray rock, brown earth, and the variegated colors of the artisans' clothing. The warm light of the rising sun gilds this arduous scene with a golden glow of hope, creating an epic atmosphere of solemnity, resilience, and the boundless creativity of humankind.", + "seed": 31152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_1242_seed_31152_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": -0.01250706221908331, + "vlm_logprob": -0.01250706221908331, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.01250706221908331, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0077477190643548965 + ], + "opposite_logprob_tokens": [ + -6.007747650146484 + ], + "logprob_reward": -0.0077477190643548965, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0077477190643548965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30692893266677856, + 0.15115681290626526, + 0.07993238419294357, + 0.06987503170967102, + 0.07937636226415634, + 0.08802900463342667, + 0.07447607815265656, + 0.043389856815338135, + 0.05309123173356056, + 0.05374421551823616 + ], + "credit_map": [ + 0.03236138075590134, + 0.05336397513747215, + 0.04982834681868553, + 0.03621872514486313, + 0.04641968384385109, + 0.030014963820576668, + 0.03632897883653641, + 0.02239290624856949, + 0.012717229314148426, + 0.028073139488697052, + 0.02395007386803627, + 0.021387837827205658, + 0.012631632387638092, + 0.024004723876714706, + 0.01861940324306488, + 0.009772787801921368, + 0.012689434923231602, + 0.010875186882913113, + 0.010670505464076996, + 0.013149530626833439, + 0.007742694113403559, + 0.009838717058300972, + 0.007536177057772875, + 0.007430139929056168, + 0.008261116221547127, + 0.00566057488322258, + 0.011351740919053555, + 0.007525070570409298, + 0.0075614782981574535, + 0.008926421403884888, + 0.010320267640054226, + 0.010268372483551502, + 0.01567523181438446, + 0.008310169912874699, + 0.013635451905429363, + 0.02053997851908207, + 0.004175976384431124, + 0.003969980403780937, + 0.0059192306362092495, + 0.007150356192141771, + 0.016551339998841286, + 0.010608011856675148, + 0.012510523200035095, + 0.02139853686094284, + 0.006674908567219973, + 0.005998118780553341, + 0.006016182247549295, + 0.008271393366158009, + 0.009763437323272228, + 0.011681979522109032, + 0.010819139890372753, + 0.007059046998620033, + 0.008513624779880047, + 0.00978520791977644, + 0.008152960799634457, + 0.008700686506927013, + 0.007718182168900967, + 0.005651159677654505, + 0.007022690959274769, + 0.004401070531457663, + 0.0040557244792580605, + 0.004089380148798227, + 0.0053116679191589355, + 0.005139985587447882, + 0.009480083361268044, + 0.008499819785356522, + 0.0074925837107002735, + 0.0044974954798817635, + 0.006117256823927164, + 0.007144950795918703, + 0.00423151021823287, + 0.005627536680549383, + 0.009161858819425106, + 0.005544997286051512, + 0.008697406388819218, + 0.006634439341723919, + 0.0061578466556966305, + 0.0051127830520272255, + 0.005560376215726137, + 0.006874510552734137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0077477190643548965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963627714663744 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.005963627714663744, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963627714663744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28931403160095215, + 0.20259256660938263, + 0.07577289640903473, + 0.05225307121872902, + 0.04754459857940674, + 0.14570271968841553, + 0.08253699541091919, + 0.0393034853041172, + 0.032437704503536224, + 0.03254194185137749 + ], + "credit_map": [ + 0.029186604544520378, + 0.04446018114686012, + 0.04650583863258362, + 0.03875435143709183, + 0.037916284054517746, + 0.02891675755381584, + 0.041813746094703674, + 0.021760236471891403, + 0.026567842811346054, + 0.03821706399321556, + 0.032094042748212814, + 0.023470429703593254, + 0.02164757438004017, + 0.02027929574251175, + 0.025956232100725174, + 0.014360083267092705, + 0.007290890906006098, + 0.011730046011507511, + 0.010125440545380116, + 0.009957418777048588, + 0.012479821220040321, + 0.007068842649459839, + 0.009676853194832802, + 0.007443581707775593, + 0.008260496892035007, + 0.004955613054335117, + 0.009085776284337044, + 0.005224349908530712, + 0.006198068615049124, + 0.0065267132595181465, + 0.007365425117313862, + 0.00463662575930357, + 0.006864843424409628, + 0.00785793550312519, + 0.006951797753572464, + 0.005710144992917776, + 0.0061684357933700085, + 0.0037467593792825937, + 0.004843883216381073, + 0.0054007931612432, + 0.03400165215134621, + 0.015461590141057968, + 0.013722460716962814, + 0.048597171902656555, + 0.0070541296154260635, + 0.008817247115075588, + 0.009159295819699764, + 0.0088891526684165, + 0.014488188549876213, + 0.0074641285464167595, + 0.01584489643573761, + 0.013558564707636833, + 0.008413716219365597, + 0.008741547353565693, + 0.005432458594441414, + 0.0085934828966856, + 0.0067062159068882465, + 0.005195285193622112, + 0.006263361312448978, + 0.0038165473379194736, + 0.003492851508781314, + 0.004586376249790192, + 0.004002315923571587, + 0.005240532103925943, + 0.0058320085518062115, + 0.004518031608313322, + 0.005314671900123358, + 0.004847521428018808, + 0.003110890043899417, + 0.0035076600033789873, + 0.002576937200501561, + 0.0027299828361719847, + 0.004462081007659435, + 0.004883801098912954, + 0.004583545494824648, + 0.003764156950637698, + 0.003920396324247122, + 0.0036761893425136805, + 0.003167196176946163, + 0.004084575455635786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.005963627714663744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596321672201157 + ], + "opposite_logprob_tokens": [ + -3.8955962657928467 + ], + "logprob_reward": -0.020596321672201157, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596321672201157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21959923207759857, + 0.07251782715320587, + 0.05450700968503952, + 0.113410584628582, + 0.21366927027702332, + 0.07802284508943558, + 0.07542232424020767, + 0.0629783347249031, + 0.06436266005039215, + 0.04550984874367714 + ], + "credit_map": [ + 0.01692606881260872, + 0.02690529264509678, + 0.03531105816364288, + 0.029355155304074287, + 0.02342298999428749, + 0.015086046420037746, + 0.053678639233112335, + 0.01891399174928665, + 0.005519482307136059, + 0.013309534639120102, + 0.011468472890555859, + 0.010991954244673252, + 0.006710667163133621, + 0.01096878107637167, + 0.00867746863514185, + 0.004871469922363758, + 0.00862468034029007, + 0.006574067287147045, + 0.007294676266610622, + 0.007476978003978729, + 0.005626630038022995, + 0.006612193305045366, + 0.0062921601347625256, + 0.006005627568811178, + 0.01746162213385105, + 0.00976650882512331, + 0.03196476027369499, + 0.013018197380006313, + 0.008416537195444107, + 0.01538790762424469, + 0.008440151810646057, + 0.008954910561442375, + 0.03802642598748207, + 0.027421124279499054, + 0.03442956879734993, + 0.03288893774151802, + 0.022276919335126877, + 0.02170315757393837, + 0.015717584639787674, + 0.021205568686127663, + 0.014553971588611603, + 0.009195886552333832, + 0.012978111393749714, + 0.01799936033785343, + 0.0053439633920788765, + 0.005222508683800697, + 0.006978584453463554, + 0.005750460550189018, + 0.011824595741927624, + 0.008421732112765312, + 0.013268056325614452, + 0.012152285315096378, + 0.006813558284193277, + 0.00799315795302391, + 0.006786262150853872, + 0.008162688463926315, + 0.011280447244644165, + 0.006618184968829155, + 0.010124036110937595, + 0.004522229544818401, + 0.007373468484729528, + 0.009188681840896606, + 0.0063949767500162125, + 0.007476315833628178, + 0.008909701369702816, + 0.006078074686229229, + 0.007571174297481775, + 0.00527120241895318, + 0.00845521129667759, + 0.013655602931976318, + 0.006624131929129362, + 0.007797567173838615, + 0.008517413400113583, + 0.0051314933225512505, + 0.009579510428011417, + 0.005128236021846533, + 0.0037705358117818832, + 0.004861286375671625, + 0.004359534941613674, + 0.0041618444956839085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.020596321672201157 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.28933918476104736, + 0.06500619649887085, + 0.043299511075019836, + 0.07063481211662292, + 0.07106731832027435, + 0.08384241163730621, + 0.155567929148674, + 0.10930616408586502, + 0.06681424379348755, + 0.04512215405702591 + ], + "credit_map": [ + 0.02925759181380272, + 0.0404314324259758, + 0.04357681795954704, + 0.037401530891656876, + 0.03468941152095795, + 0.022872159257531166, + 0.05877288430929184, + 0.022337408736348152, + 0.0075737228617072105, + 0.010791865177452564, + 0.011058658361434937, + 0.008767623454332352, + 0.005353151820600033, + 0.006926325149834156, + 0.008028055541217327, + 0.006506802048534155, + 0.00504396203905344, + 0.006840115413069725, + 0.00556575320661068, + 0.0056086089462041855, + 0.005351233761757612, + 0.005017344374209642, + 0.005330164451152086, + 0.004542330279946327, + 0.00997807364910841, + 0.005782551132142544, + 0.016593795269727707, + 0.008181583136320114, + 0.0058547514490783215, + 0.00752592645585537, + 0.010360938496887684, + 0.006357197184115648, + 0.01037828903645277, + 0.009039669297635555, + 0.012990106828510761, + 0.010602544993162155, + 0.007887660525739193, + 0.007878946140408516, + 0.0056411041878163815, + 0.006648998241871595, + 0.014756767079234123, + 0.01243798527866602, + 0.013018216006457806, + 0.018647558987140656, + 0.007361073978245258, + 0.005633101332932711, + 0.00655581708997488, + 0.0054319025948643684, + 0.021185418590903282, + 0.014249876141548157, + 0.02801728993654251, + 0.025603678077459335, + 0.01646452769637108, + 0.017926273867487907, + 0.012892873957753181, + 0.019227992743253708, + 0.015720976516604424, + 0.01349891722202301, + 0.014289163053035736, + 0.007677498273551464, + 0.015467827208340168, + 0.017935534939169884, + 0.012416000477969646, + 0.012300257571041584, + 0.010428198613226414, + 0.006483602803200483, + 0.015760594978928566, + 0.012095854617655277, + 0.006278099957853556, + 0.005934680812060833, + 0.004380714148283005, + 0.00545250391587615, + 0.005733206868171692, + 0.0055437348783016205, + 0.00592963770031929, + 0.004947373177856207, + 0.006464160978794098, + 0.006271189544349909, + 0.004783921875059605, + 0.005448938813060522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "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 video begin with a wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631788495928049 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631788495928049, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631788495928049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1357869803905487, + 0.05189693719148636, + 0.12362859398126602, + 0.06006468087434769, + 0.09577236324548721, + 0.08375240117311478, + 0.20186002552509308, + 0.11515578627586365, + 0.0594305619597435, + 0.07265166193246841 + ], + "credit_map": [ + 0.01214207336306572, + 0.01694725826382637, + 0.01995617151260376, + 0.019825739786028862, + 0.02151503972709179, + 0.014808298088610172, + 0.018265925347805023, + 0.012326476164162159, + 0.005032398272305727, + 0.006792947184294462, + 0.006842134520411491, + 0.006216271780431271, + 0.006032208912074566, + 0.00996899139136076, + 0.005752649623900652, + 0.005259336903691292, + 0.015053492970764637, + 0.021159544587135315, + 0.01833229511976242, + 0.020816566422581673, + 0.011811000294983387, + 0.013877533376216888, + 0.010553961619734764, + 0.012024199590086937, + 0.007735821884125471, + 0.006221714895218611, + 0.008240834809839725, + 0.006914883386343718, + 0.007438939996063709, + 0.006298457738012075, + 0.009325860999524593, + 0.007888171821832657, + 0.012900294736027718, + 0.01653490588068962, + 0.012679883278906345, + 0.010914100334048271, + 0.013353584334254265, + 0.011085287667810917, + 0.008743228390812874, + 0.009561074897646904, + 0.01774740219116211, + 0.011616918258368969, + 0.012429784052073956, + 0.023577187210321426, + 0.0036507451441138983, + 0.005347253754734993, + 0.004746966529637575, + 0.004636138677597046, + 0.03109169192612171, + 0.01981763355433941, + 0.0426747091114521, + 0.025644412264227867, + 0.019019506871700287, + 0.020425012335181236, + 0.01984689012169838, + 0.023340174928307533, + 0.015935631468892097, + 0.016131851822137833, + 0.022457093000411987, + 0.010241053998470306, + 0.01246437057852745, + 0.010488384403288364, + 0.016409993171691895, + 0.011027405969798565, + 0.010083536617457867, + 0.005415492225438356, + 0.012644830159842968, + 0.00854525901377201, + 0.006430717650800943, + 0.006373247131705284, + 0.004855722654610872, + 0.005081753712147474, + 0.0061430055648088455, + 0.009556114673614502, + 0.007308285217732191, + 0.009559937752783298, + 0.010632744990289211, + 0.008108807727694511, + 0.011378731578588486, + 0.00996403954923153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.007631788495928049 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot pans slowly to the right from high above, sweeping across a steep, winding, mist-shrouded mountain ridge. Under the low, golden sunlight of early morning, a new section of the Great Wall rises from the ground. The figures of thousands of ancient craftsmen dot the steep slopes and the nascent wall, creating a magnificent scene of labor. Most of them are sturdy men, dressed in worn, faded coarse linen shorts. Their bronzed skin, beaded with sweat, gleams in the sun. As the camera pans down and slowly zooms in, we can see details: a team of artisans, working in unison, chanting slogans, painstakingly drag huge, unpolished granite blocks up the slope using thick hemp ropes and crude wooden rollers. Elsewhere, on bamboo scaffolding erected at the edge of the wall, stonemasons carefully hammer and chisel the edges of the stones, forcing them to fit more tightly, their movements focused and rhythmic. Below, scores of people, carrying heavy stones or buckets filled with mud, navigate the narrow and rugged construction site, their steps stumbling under the weight but unwavering. The entire scene is shrouded in a stark palette of gray rock, brown earth, and the variegated colors of the artisans' clothing. The warm light of the rising sun gilds this arduous scene with a golden glow of hope, creating an epic atmosphere of solemnity, resilience, and the boundless creativity of humankind.", + "seed": 31153, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_1242_seed_31153_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": -0.06449622195214033, + "vlm_logprob": -0.06449622195214033, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.06449622195214033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015383869409561157 + ], + "opposite_logprob_tokens": [ + -4.765383720397949 + ], + "logprob_reward": -0.015383869409561157, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015383869409561157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23910310864448547, + 0.08937117457389832, + 0.13636431097984314, + 0.23477892577648163, + 0.0558968223631382, + 0.07065688073635101, + 0.05058736726641655, + 0.047803156077861786, + 0.030502140522003174, + 0.044936127960681915 + ], + "credit_map": [ + 0.02181166596710682, + 0.0297665074467659, + 0.033706262707710266, + 0.034004248678684235, + 0.03540991619229317, + 0.03504370525479317, + 0.027950195595622063, + 0.021410593762993813, + 0.005842078942805529, + 0.00918635819107294, + 0.01046193391084671, + 0.008147542364895344, + 0.024689162150025368, + 0.010137730278074741, + 0.012142025865614414, + 0.008764334954321384, + 0.01862058974802494, + 0.019811565056443214, + 0.01738061010837555, + 0.022835830226540565, + 0.015189682133495808, + 0.014882134273648262, + 0.011994867585599422, + 0.01564902812242508, + 0.0870499312877655, + 0.014711170457303524, + 0.01289467141032219, + 0.011013336479663849, + 0.01311836950480938, + 0.07180178910493851, + 0.011744888499379158, + 0.012444738298654556, + 0.005077809561043978, + 0.0050062634982168674, + 0.006514432840049267, + 0.004406875930726528, + 0.00649683503434062, + 0.010408581234514713, + 0.004727476742118597, + 0.013258542865514755, + 0.007568735629320145, + 0.009380136616528034, + 0.009880107827484608, + 0.00720904441550374, + 0.00981706753373146, + 0.008229650557041168, + 0.011623687110841274, + 0.006948450114578009, + 0.006146366707980633, + 0.0058676982298493385, + 0.005964463111013174, + 0.007610307540744543, + 0.005295733455568552, + 0.006702856160700321, + 0.0067169624380767345, + 0.006282979156821966, + 0.005035193171352148, + 0.010774602182209492, + 0.006689198315143585, + 0.004266382660716772, + 0.0075452132150530815, + 0.005561606027185917, + 0.002879195846617222, + 0.005051766522228718, + 0.0028486293740570545, + 0.002777513349428773, + 0.0051981378346681595, + 0.0022343313321471214, + 0.004347920883446932, + 0.003747672773897648, + 0.005197193939238787, + 0.00415074173361063, + 0.005012507550418377, + 0.004622550215572119, + 0.00835414882749319, + 0.0053568873554468155, + 0.005563780665397644, + 0.004451100714504719, + 0.006223059259355068, + 0.005352089647203684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.015383869409561157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561958283185959 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.012561958283185959, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561958283185959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2863330841064453, + 0.05134471505880356, + 0.12800905108451843, + 0.1632002592086792, + 0.08436905592679977, + 0.11742862313985825, + 0.041331697255373, + 0.07081988453865051, + 0.023266691714525223, + 0.033896997570991516 + ], + "credit_map": [ + 0.03614841029047966, + 0.05283070728182793, + 0.04422247037291527, + 0.03256461024284363, + 0.04260692745447159, + 0.033739253878593445, + 0.025812169536948204, + 0.018408536911010742, + 0.005239200312644243, + 0.008287070319056511, + 0.006922876928001642, + 0.004834774881601334, + 0.008099397644400597, + 0.00715198228135705, + 0.006139142904430628, + 0.004670260939747095, + 0.014228153973817825, + 0.015782572329044342, + 0.01637059822678566, + 0.02601180598139763, + 0.013126214034855366, + 0.017103979364037514, + 0.011668384075164795, + 0.013717328198254108, + 0.061003100126981735, + 0.010247721336781979, + 0.008763042278587818, + 0.007113381288945675, + 0.009505413472652435, + 0.0511493906378746, + 0.007491530850529671, + 0.007926671765744686, + 0.007563663646578789, + 0.0074753426015377045, + 0.0067491913214325905, + 0.006980126723647118, + 0.008629754185676575, + 0.017761768773198128, + 0.005153054371476173, + 0.024056144058704376, + 0.010487530380487442, + 0.01597941294312477, + 0.011953849345445633, + 0.01089232787489891, + 0.017713187262415886, + 0.011587423272430897, + 0.024192804470658302, + 0.014622081071138382, + 0.004138189367949963, + 0.0045125302858650684, + 0.005163562949746847, + 0.00484659755602479, + 0.003911140374839306, + 0.004401550628244877, + 0.005384605377912521, + 0.008973515592515469, + 0.008030642755329609, + 0.009168076328933239, + 0.010759648866951466, + 0.007252558600157499, + 0.011035184375941753, + 0.00865435041487217, + 0.0075039928779006, + 0.008415419608354568, + 0.002664902014657855, + 0.0024789695162326097, + 0.002819407032802701, + 0.0024062448646873236, + 0.004817558452486992, + 0.0020303602796047926, + 0.002974254311993718, + 0.0030749926809221506, + 0.004401677753776312, + 0.0035808547399938107, + 0.006096746772527695, + 0.0027901045978069305, + 0.004065460525453091, + 0.0041078501380980015, + 0.003228641813620925, + 0.005625659599900246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012561958283185959 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25216206908226013 + ], + "opposite_logprob_tokens": [ + -1.5021620988845825 + ], + "logprob_reward": -0.25216206908226013, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25216206908226013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15502196550369263, + 0.12231650948524475, + 0.18659941852092743, + 0.24923750758171082, + 0.0424160473048687, + 0.09698419272899628, + 0.01953202299773693, + 0.06289303302764893, + 0.0226149782538414, + 0.04238436371088028 + ], + "credit_map": [ + 0.015288928523659706, + 0.025614526122808456, + 0.022010765969753265, + 0.0197269469499588, + 0.017767280340194702, + 0.021205581724643707, + 0.02032768353819847, + 0.013080255128443241, + 0.011564035899937153, + 0.011934746988117695, + 0.017913397401571274, + 0.00879356637597084, + 0.037624530494213104, + 0.011507701128721237, + 0.014655496925115585, + 0.008323028683662415, + 0.018819652497768402, + 0.02375071682035923, + 0.02000790275633335, + 0.036976732313632965, + 0.020988348871469498, + 0.02059268020093441, + 0.019561657682061195, + 0.025901708751916885, + 0.0920015200972557, + 0.013751114718616009, + 0.015245331451296806, + 0.011321226134896278, + 0.015320652164518833, + 0.07729365676641464, + 0.01067030057311058, + 0.013633682392537594, + 0.006447738502174616, + 0.0042106714099645615, + 0.006259876303374767, + 0.003353466046974063, + 0.00589726073667407, + 0.005266799125820398, + 0.003506918903440237, + 0.007473311852663755, + 0.010128472931683064, + 0.014911748468875885, + 0.012825924903154373, + 0.01024322584271431, + 0.011447001248598099, + 0.01065622828900814, + 0.010067194700241089, + 0.016704395413398743, + 0.0019069523550570011, + 0.001679680310189724, + 0.002570528769865632, + 0.0013304307358339429, + 0.0016411152901127934, + 0.002913871081545949, + 0.002052182098850608, + 0.005437260959297419, + 0.007598409429192543, + 0.008551839739084244, + 0.01065787859261036, + 0.007959176786243916, + 0.007234940771013498, + 0.007054088171571493, + 0.006903133355081081, + 0.006933566648513079, + 0.002124105580151081, + 0.0021558445878326893, + 0.002792901359498501, + 0.0011608239728957415, + 0.005242265295237303, + 0.003842419246211648, + 0.0017814552411437035, + 0.0035151627380400896, + 0.006090138573199511, + 0.006140525918453932, + 0.005894457921385765, + 0.005435631610453129, + 0.005338881630450487, + 0.005272080656141043, + 0.004463402088731527, + 0.003749244147911668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.25216206908226013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029826322570443153 + ], + "opposite_logprob_tokens": [ + -3.5298264026641846 + ], + "logprob_reward": -0.029826322570443153, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029826322570443153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1288776695728302, + 0.0730719119310379, + 0.10108614712953568, + 0.3798214793205261, + 0.049114786088466644, + 0.12018509209156036, + 0.032493699342012405, + 0.04379274323582649, + 0.031088272109627724, + 0.04046819731593132 + ], + "credit_map": [ + 0.0129048777744174, + 0.01800370216369629, + 0.021956512704491615, + 0.017657404765486717, + 0.014784068800508976, + 0.01581547223031521, + 0.017233138903975487, + 0.010522490367293358, + 0.008041268214583397, + 0.008858127519488335, + 0.012264810502529144, + 0.005070118233561516, + 0.021040040999650955, + 0.0058157616294920444, + 0.007243118714541197, + 0.004738672636449337, + 0.010863034054636955, + 0.014092239551246166, + 0.012391525320708752, + 0.017784930765628815, + 0.012171468697488308, + 0.013799195177853107, + 0.009623113088309765, + 0.010360636748373508, + 0.14534512162208557, + 0.021927334368228912, + 0.021365705877542496, + 0.016558928415179253, + 0.02195708081126213, + 0.1167115643620491, + 0.01712377928197384, + 0.01883193850517273, + 0.006834982894361019, + 0.005196554586291313, + 0.005771954078227282, + 0.004094031639397144, + 0.005811815150082111, + 0.006791853345930576, + 0.0041105374693870544, + 0.010503055527806282, + 0.012840584851801395, + 0.017751522362232208, + 0.017643515020608902, + 0.012030533514916897, + 0.013058377429842949, + 0.012491161935031414, + 0.018275195732712746, + 0.016094205901026726, + 0.003136167535558343, + 0.003965114708989859, + 0.0040700798854231834, + 0.004679861944168806, + 0.0030005997978150845, + 0.004198519047349691, + 0.003910011611878872, + 0.005533344577997923, + 0.0056573329493403435, + 0.00724842818453908, + 0.006147776264697313, + 0.005539418198168278, + 0.005307245068252087, + 0.004212504718452692, + 0.004201216157525778, + 0.005478818900883198, + 0.0026485221460461617, + 0.003004631493240595, + 0.0033247137907892466, + 0.002111930400133133, + 0.009619981981813908, + 0.0035857821349054575, + 0.0027542628813534975, + 0.004038447514176369, + 0.006296419072896242, + 0.006800143979489803, + 0.005703930277377367, + 0.0050140926614403725, + 0.004092725925147533, + 0.004279944114387035, + 0.0038696234114468098, + 0.004411318805068731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.029826322570443153 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546890415251255 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012546890415251255, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546890415251255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1334288865327835, + 0.04603913053870201, + 0.08383391797542572, + 0.504109799861908, + 0.025948278605937958, + 0.059916649013757706, + 0.024998757988214493, + 0.07724601775407791, + 0.023981435224413872, + 0.02049718052148819 + ], + "credit_map": [ + 0.011410369537770748, + 0.01709158904850483, + 0.021133773028850555, + 0.01563730277121067, + 0.020703841000795364, + 0.020630881190299988, + 0.01570814475417137, + 0.011112969368696213, + 0.005042379721999168, + 0.007244801614433527, + 0.006794331595301628, + 0.004870174452662468, + 0.006591087207198143, + 0.005309707019478083, + 0.005120906513184309, + 0.005065735895186663, + 0.008913733996450901, + 0.011079919524490833, + 0.010431874543428421, + 0.01750248856842518, + 0.006979526486247778, + 0.011805855669081211, + 0.008236849680542946, + 0.008883663453161716, + 0.20188488066196442, + 0.030103979632258415, + 0.028122449293732643, + 0.02151934802532196, + 0.03022949956357479, + 0.1518338918685913, + 0.019606605172157288, + 0.02080908790230751, + 0.002733837580308318, + 0.0028562077786773443, + 0.003206577617675066, + 0.0028758603148162365, + 0.0037179591599851847, + 0.0049415468238294125, + 0.002535283798351884, + 0.003081002738326788, + 0.005583478137850761, + 0.006915416102856398, + 0.007401753216981888, + 0.008291429840028286, + 0.006196342874318361, + 0.008173043839633465, + 0.011602520942687988, + 0.005752663128077984, + 0.0023458211217075586, + 0.002705032005906105, + 0.0028621687088161707, + 0.0025640048552304506, + 0.0024483855813741684, + 0.0032045713160187006, + 0.0038468509446829557, + 0.005021920893341303, + 0.0072565218433737755, + 0.013110792264342308, + 0.013256166130304337, + 0.007836892269551754, + 0.012750585563480854, + 0.008958978578448296, + 0.0077951280400156975, + 0.006280948407948017, + 0.0027189282700419426, + 0.002394895302131772, + 0.0030414036009460688, + 0.0019727221224457026, + 0.0051672449335455894, + 0.0030771479941904545, + 0.002434590831398964, + 0.0031745012383908033, + 0.0029249663930386305, + 0.002750387182459235, + 0.002659364603459835, + 0.0024000483099371195, + 0.002250893507152796, + 0.002454562345519662, + 0.0024912783410400152, + 0.0025656772777438164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012546890415251255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot pans slowly to the right from high above, sweeping across a steep, winding, mist-shrouded mountain ridge. Under the low, golden sunlight of early morning, a new section of the Great Wall rises from the ground. The figures of thousands of ancient craftsmen dot the steep slopes and the nascent wall, creating a magnificent scene of labor. Most of them are sturdy men, dressed in worn, faded coarse linen shorts. Their bronzed skin, beaded with sweat, gleams in the sun. As the camera pans down and slowly zooms in, we can see details: a team of artisans, working in unison, chanting slogans, painstakingly drag huge, unpolished granite blocks up the slope using thick hemp ropes and crude wooden rollers. Elsewhere, on bamboo scaffolding erected at the edge of the wall, stonemasons carefully hammer and chisel the edges of the stones, forcing them to fit more tightly, their movements focused and rhythmic. Below, scores of people, carrying heavy stones or buckets filled with mud, navigate the narrow and rugged construction site, their steps stumbling under the weight but unwavering. The entire scene is shrouded in a stark palette of gray rock, brown earth, and the variegated colors of the artisans' clothing. The warm light of the rising sun gilds this arduous scene with a golden glow of hope, creating an epic atmosphere of solemnity, resilience, and the boundless creativity of humankind.", + "seed": 31154, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_1242_seed_31154_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": -0.024788563326001167, + "vlm_logprob": -0.024788563326001167, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.024788563326001167, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025120023638010025 + ], + "opposite_logprob_tokens": [ + -4.400120258331299 + ], + "logprob_reward": -0.025120023638010025, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.025120023638010025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05573652312159538, + 0.025084178894758224, + 0.03741028532385826, + 0.013188277371227741, + 0.715033233165741, + 0.015578042715787888, + 0.044854212552309036, + 0.05661284178495407, + 0.009891044348478317, + 0.026611320674419403 + ], + "credit_map": [ + 0.003697864478453994, + 0.007270465604960918, + 0.0027934301178902388, + 0.005674136336892843, + 0.00458014290779829, + 0.013972015120089054, + 0.004837976768612862, + 0.012910496443510056, + 0.0018245747778564692, + 0.0031645060516893864, + 0.0034173866733908653, + 0.0035207218024879694, + 0.0020502102561295033, + 0.0027624524664133787, + 0.0024181462358683348, + 0.005926182493567467, + 0.00287418719381094, + 0.002283262088894844, + 0.012904541566967964, + 0.0021180370822548866, + 0.0019637160003185272, + 0.001930583268404007, + 0.0018768997397273779, + 0.01145906001329422, + 0.001569850486703217, + 0.0015122605254873633, + 0.0020054993219673634, + 0.001707445946522057, + 0.001664811046794057, + 0.0017673852853477001, + 0.0012397227110341191, + 0.001721303560771048, + 0.03741448000073433, + 0.03255688026547432, + 0.03984025865793228, + 0.2160528302192688, + 0.06168096140027046, + 0.20501300692558289, + 0.06824301183223724, + 0.05423181504011154, + 0.001126650022342801, + 0.0018569192616268992, + 0.002211127197369933, + 0.0027566025964915752, + 0.0018800549441948533, + 0.0019395812414586544, + 0.0018411005148664117, + 0.0019660070538520813, + 0.0019491288112476468, + 0.0013930528657510877, + 0.017481479793787003, + 0.0023315446451306343, + 0.0021665189415216446, + 0.016922175884246826, + 0.0011643976904451847, + 0.001445918227545917, + 0.004889442585408688, + 0.0036927165929228067, + 0.0059158108197152615, + 0.010977915488183498, + 0.0045606293715536594, + 0.008672622963786125, + 0.012759353965520859, + 0.005144353955984116, + 0.0010507708648219705, + 0.0009010839276015759, + 0.0013670126209035516, + 0.0015851317439228296, + 0.0011526021407917142, + 0.0014197805430740118, + 0.0010928870178759098, + 0.001321776071563363, + 0.0024257039185613394, + 0.0024176607839763165, + 0.0032095955684781075, + 0.0052341497503221035, + 0.0029706545174121857, + 0.0031367044430226088, + 0.0036552948877215385, + 0.0035615600645542145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.025120023638010025 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221016973257065 + ], + "opposite_logprob_tokens": [ + -4.018220901489258 + ], + "logprob_reward": -0.018221016973257065, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221016973257065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14066679775714874, + 0.03936367481946945, + 0.018190957605838776, + 0.041267987340688705, + 0.5308312773704529, + 0.010790128260850906, + 0.021560808643698692, + 0.049923285841941833, + 0.11985917389392853, + 0.02754582278430462 + ], + "credit_map": [ + 0.01161988265812397, + 0.023747259750962257, + 0.0114549919962883, + 0.014016566798090935, + 0.0156510341912508, + 0.022357190027832985, + 0.014881668612360954, + 0.026938237249851227, + 0.0024377931840717793, + 0.004336475860327482, + 0.003373952815309167, + 0.002991561545059085, + 0.004786462057381868, + 0.006865162868052721, + 0.002749338513240218, + 0.011822937056422234, + 0.002165416721254587, + 0.0021669284906238317, + 0.0031209627632051706, + 0.0025653240736573935, + 0.0023411265574395657, + 0.002475881949067116, + 0.0016742160078138113, + 0.0016811054665595293, + 0.00307577196508646, + 0.004014888312667608, + 0.004307358991354704, + 0.00571897067129612, + 0.0046034082770347595, + 0.008039630949497223, + 0.0034045991487801075, + 0.008103369735181332, + 0.027501318603754044, + 0.024555180221796036, + 0.029847951605916023, + 0.16073164343833923, + 0.045377466827631, + 0.15125595033168793, + 0.05067608505487442, + 0.040885746479034424, + 0.0009429390192963183, + 0.0015323725529015064, + 0.0017996751703321934, + 0.0014570123748853803, + 0.002027892740443349, + 0.0010778509313240647, + 0.0009360749390907586, + 0.001016312511637807, + 0.0014006163692101836, + 0.001746711670421064, + 0.0045224702917039394, + 0.0031241315882653, + 0.0023761375341564417, + 0.005114797502756119, + 0.0017976617673411965, + 0.0014782864600419998, + 0.0037341807037591934, + 0.004895680584013462, + 0.005052960943430662, + 0.010269738733768463, + 0.004444880876690149, + 0.006528275553137064, + 0.010507439263164997, + 0.004490140825510025, + 0.014846991747617722, + 0.017447134479880333, + 0.012089393101632595, + 0.02163129299879074, + 0.010974143631756306, + 0.0138898566365242, + 0.012801310047507286, + 0.01617908477783203, + 0.0020833169110119343, + 0.001685542636550963, + 0.0032528494484722614, + 0.0034621087834239006, + 0.0020158691331744194, + 0.002561993896961212, + 0.005442803725600243, + 0.007041343953460455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.018221016973257065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814423456788063 + ], + "opposite_logprob_tokens": [ + -3.28814435005188 + ], + "logprob_reward": -0.03814423456788063, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814423456788063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038733500987291336, + 0.04236387088894844, + 0.0481436625123024, + 0.027478141710162163, + 0.6205747127532959, + 0.0066928379237651825, + 0.09113902598619461, + 0.044259849935770035, + 0.04723405838012695, + 0.03338030353188515 + ], + "credit_map": [ + 0.00342071452178061, + 0.0034560493659228086, + 0.0025960372295230627, + 0.004476653877645731, + 0.0029981089755892754, + 0.008986307308077812, + 0.002781307091936469, + 0.010018323548138142, + 0.0030177035368978977, + 0.007276058662682772, + 0.005924390163272619, + 0.005772915203124285, + 0.0035272217355668545, + 0.004460333380848169, + 0.005709164775907993, + 0.006676083896309137, + 0.005272515118122101, + 0.00544663704931736, + 0.012109804898500443, + 0.0041918628849089146, + 0.0031947698444128036, + 0.003472256939858198, + 0.0034931369591504335, + 0.01096267905086279, + 0.0022741486318409443, + 0.0036717781331390142, + 0.0032442945521324873, + 0.004610654432326555, + 0.004075991455465555, + 0.00312441005371511, + 0.0023956429213285446, + 0.0040812245570123196, + 0.0323287658393383, + 0.028988489881157875, + 0.034685295075178146, + 0.18813307583332062, + 0.05287026986479759, + 0.17668692767620087, + 0.059265680611133575, + 0.04761616513133049, + 0.0005007840227335691, + 0.0005999832064844668, + 0.0011260547908023, + 0.0011992406798526645, + 0.0009472395176999271, + 0.0006668457062914968, + 0.0007935771718621254, + 0.0008591128862462938, + 0.0038093712646514177, + 0.0035795264411717653, + 0.032173097133636475, + 0.007130092475563288, + 0.004563459195196629, + 0.03327435255050659, + 0.003425046568736434, + 0.0031840833835303783, + 0.0037331520579755306, + 0.00397926801815629, + 0.005087817087769508, + 0.006832382641732693, + 0.005035106558352709, + 0.008761115372180939, + 0.007374376989901066, + 0.003456630278378725, + 0.005835083778947592, + 0.006523681338876486, + 0.004849323537200689, + 0.00845697708427906, + 0.004342943895608187, + 0.005491096060723066, + 0.005264486651867628, + 0.006470467895269394, + 0.0036324516404420137, + 0.0037671453319489956, + 0.0042212046682834625, + 0.005389300175011158, + 0.0030426380690187216, + 0.0039418721571564674, + 0.0051134745590388775, + 0.004272217396646738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.03814423456788063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026376035064458847 + ], + "opposite_logprob_tokens": [ + -3.6513760089874268 + ], + "logprob_reward": -0.026376035064458847, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026376035064458847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024067388847470284, + 0.02635941095650196, + 0.02836379036307335, + 0.01331597100943327, + 0.6527177691459656, + 0.016632016748189926, + 0.01357201300561428, + 0.03378051146864891, + 0.17243173718452454, + 0.018759386613965034 + ], + "credit_map": [ + 0.0027746958658099174, + 0.002781282179057598, + 0.002690405584871769, + 0.0024882324505597353, + 0.0035851828288286924, + 0.004266435280442238, + 0.0021573177073150873, + 0.0033238353207707405, + 0.0021613938733935356, + 0.004879893735051155, + 0.0038938517682254314, + 0.0037985872477293015, + 0.002482548588886857, + 0.0024372334592044353, + 0.0033527384512126446, + 0.003353159874677658, + 0.002547638490796089, + 0.0028536419849842787, + 0.00854235328733921, + 0.0028310043271631002, + 0.002040415769442916, + 0.002699305536225438, + 0.001912008854560554, + 0.004937417339533567, + 0.0015515059931203723, + 0.0018540043383836746, + 0.0014140564016997814, + 0.00180999340955168, + 0.0018219101475551724, + 0.0017352091381326318, + 0.001146491733379662, + 0.0019827978685498238, + 0.03399481996893883, + 0.030176077038049698, + 0.036582864820957184, + 0.19771818816661835, + 0.05588766559958458, + 0.1860191375017166, + 0.06237221509218216, + 0.049966778606176376, + 0.0011651460081338882, + 0.001333208056166768, + 0.003244468243792653, + 0.0018231028225272894, + 0.004506813827902079, + 0.0018439855193719268, + 0.001249779830686748, + 0.0014655105769634247, + 0.0007527504931204021, + 0.001048521138727665, + 0.002789047546684742, + 0.002561788074672222, + 0.0016472396673634648, + 0.002723491517826915, + 0.0013151898747310042, + 0.0007339832372963428, + 0.0025260592810809612, + 0.0023502495605498552, + 0.0035819541662931442, + 0.007261944934725761, + 0.002516359556466341, + 0.005113872233778238, + 0.007492521777749062, + 0.0029375466983765364, + 0.021291159093379974, + 0.0252018254250288, + 0.017491405829787254, + 0.030849890783429146, + 0.01597714051604271, + 0.019918693229556084, + 0.018379805609583855, + 0.023321814835071564, + 0.0019025023793801665, + 0.0014272305415943265, + 0.002234827959910035, + 0.003111858619377017, + 0.0013820393942296505, + 0.0014881694223731756, + 0.0038620701525360346, + 0.0033506853505969048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.026376035064458847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608150638639927 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.01608150638639927, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608150638639927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03360772877931595, + 0.01738767884671688, + 0.04251750186085701, + 0.006510356441140175, + 0.7535029649734497, + 0.01630723662674427, + 0.021495621651411057, + 0.0321623831987381, + 0.06125415116548538, + 0.015254384838044643 + ], + "credit_map": [ + 0.0027229655534029007, + 0.007086153607815504, + 0.0029777162708342075, + 0.003881806740537286, + 0.0043536894954741, + 0.0047112419269979, + 0.004385188687592745, + 0.0034889630042016506, + 0.001180193037725985, + 0.0017396322218701243, + 0.0030257320031523705, + 0.0019220600370317698, + 0.0013813289115205407, + 0.0018747736467048526, + 0.0015847519971430302, + 0.004679204896092415, + 0.0034213385079056025, + 0.0031848640646785498, + 0.014712867327034473, + 0.004143999423831701, + 0.0021213104482740164, + 0.002550457138568163, + 0.0023349991533905268, + 0.01004765648394823, + 0.0008883190457709134, + 0.0008308914257213473, + 0.0010721618309617043, + 0.0004964214167557657, + 0.0007413753774017096, + 0.0009694844484329224, + 0.0007624217541888356, + 0.000749280268792063, + 0.039405595511198044, + 0.03451477363705635, + 0.042081814259290695, + 0.22805167734622955, + 0.06478995084762573, + 0.21465188264846802, + 0.07230401784181595, + 0.05770322307944298, + 0.001070222700946033, + 0.0013731335056945682, + 0.0035172791685909033, + 0.0013959197094663978, + 0.004763406701385975, + 0.0020188118796795607, + 0.0011384133249521255, + 0.0010300491703674197, + 0.001050080289132893, + 0.0009757744846865535, + 0.006966653745621443, + 0.0022394696716219187, + 0.001334593747742474, + 0.007140714675188065, + 0.0008572617662139237, + 0.0009310721652582288, + 0.0026844919193536043, + 0.0025512201245874166, + 0.0035660695284605026, + 0.006811856757849455, + 0.0025090475101023912, + 0.004664879757910967, + 0.006647706497460604, + 0.002727110404521227, + 0.007410548627376556, + 0.008950779214501381, + 0.006340437568724155, + 0.010623740032315254, + 0.005952269304543734, + 0.006983746308833361, + 0.006463239435106516, + 0.00852938462048769, + 0.001526449341326952, + 0.0012562222545966506, + 0.0019129883730784059, + 0.0028299852274358273, + 0.0015042528975754976, + 0.0016730957431718707, + 0.0022767421323806047, + 0.0022746471222490072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.01608150638639927 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide shot pans slowly to the right from high above, sweeping across a steep, winding, mist-shrouded mountain ridge. Under the low, golden sunlight of early morning, a new section of the Great Wall rises from the ground. The figures of thousands of ancient craftsmen dot the steep slopes and the nascent wall, creating a magnificent scene of labor. Most of them are sturdy men, dressed in worn, faded coarse linen shorts. Their bronzed skin, beaded with sweat, gleams in the sun. As the camera pans down and slowly zooms in, we can see details: a team of artisans, working in unison, chanting slogans, painstakingly drag huge, unpolished granite blocks up the slope using thick hemp ropes and crude wooden rollers. Elsewhere, on bamboo scaffolding erected at the edge of the wall, stonemasons carefully hammer and chisel the edges of the stones, forcing them to fit more tightly, their movements focused and rhythmic. Below, scores of people, carrying heavy stones or buckets filled with mud, navigate the narrow and rugged construction site, their steps stumbling under the weight but unwavering. The entire scene is shrouded in a stark palette of gray rock, brown earth, and the variegated colors of the artisans' clothing. The warm light of the rising sun gilds this arduous scene with a golden glow of hope, creating an epic atmosphere of solemnity, resilience, and the boundless creativity of humankind.", + "seed": 31155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_2/index_1242_seed_31155_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": -0.011874537914991379, + "vlm_logprob": -0.011874537914991379, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.011874537914991379, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012697681784629822 + ], + "opposite_logprob_tokens": [ + -5.012697696685791 + ], + "logprob_reward": -0.012697681784629822, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.012697681784629822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12500989437103271, + 0.10721489042043686, + 0.03984294459223747, + 0.04075844585895538, + 0.02658340521156788, + 0.05176766589283943, + 0.12289951741695404, + 0.4274449646472931, + 0.04022560641169548, + 0.018252672627568245 + ], + "credit_map": [ + 0.012887846678495407, + 0.013998233713209629, + 0.01255751308053732, + 0.01691620983183384, + 0.017853690311312675, + 0.012055900879204273, + 0.01980089209973812, + 0.018939608708024025, + 0.016696564853191376, + 0.011105870828032494, + 0.031223436817526817, + 0.007953844033181667, + 0.0046672816388309, + 0.01021328940987587, + 0.014424926601350307, + 0.010929670184850693, + 0.005554758477956057, + 0.004246924072504044, + 0.004379668273031712, + 0.0028581509832292795, + 0.003746630856767297, + 0.006356481928378344, + 0.0037351229693740606, + 0.008965209126472473, + 0.004981537349522114, + 0.004187860060483217, + 0.006460085045546293, + 0.0048949080519378185, + 0.0032928662840276957, + 0.004176207818090916, + 0.007056301925331354, + 0.005708681419491768, + 0.004168638028204441, + 0.002742309123277664, + 0.0034880824387073517, + 0.002840055851265788, + 0.002880820771679282, + 0.003226092318072915, + 0.0028044418431818485, + 0.004432963207364082, + 0.005252572242170572, + 0.0029209088534116745, + 0.007049156352877617, + 0.0038338182494044304, + 0.0034180658403784037, + 0.012746005319058895, + 0.0033778492361307144, + 0.01316929142922163, + 0.01096002571284771, + 0.00864726584404707, + 0.039838362485170364, + 0.010961201973259449, + 0.006157632917165756, + 0.008280686102807522, + 0.00797341950237751, + 0.03008093126118183, + 0.025517437607049942, + 0.010302922688424587, + 0.17293356359004974, + 0.038908056914806366, + 0.02109161578118801, + 0.07436763495206833, + 0.06508753448724747, + 0.019236231222748756, + 0.006144489627331495, + 0.002992133842781186, + 0.007210372481495142, + 0.0037587485276162624, + 0.005416056606918573, + 0.006024088244885206, + 0.003776558442041278, + 0.004903155844658613, + 0.001997779356315732, + 0.0013631804613396525, + 0.0021855393424630165, + 0.003169099800288677, + 0.002292339690029621, + 0.0022687502205371857, + 0.0024829572066664696, + 0.0024930268991738558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.012697681784629822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646576501429081 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646576501429081, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646576501429081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09963252395391464, + 0.08947613090276718, + 0.0838942602276802, + 0.0404190719127655, + 0.055220965296030045, + 0.1943984478712082, + 0.11917182058095932, + 0.20075425505638123, + 0.08682163804769516, + 0.030210889875888824 + ], + "credit_map": [ + 0.011168955825269222, + 0.015032664872705936, + 0.012739600613713264, + 0.01834229938685894, + 0.010827112942934036, + 0.010247791185975075, + 0.01285643968731165, + 0.008417661301791668, + 0.014634870924055576, + 0.010326983407139778, + 0.02308792620897293, + 0.00963064655661583, + 0.006163233891129494, + 0.0108473040163517, + 0.0067680575884878635, + 0.008017114363610744, + 0.014030808582901955, + 0.006872608792036772, + 0.00744818476960063, + 0.0070415157824754715, + 0.006179737858474255, + 0.01237333845347166, + 0.007552666123956442, + 0.02239539846777916, + 0.0028247430454939604, + 0.0034662671387195587, + 0.0028885158244520426, + 0.00345085421577096, + 0.003477700985968113, + 0.003081843489781022, + 0.011235577054321766, + 0.009993569925427437, + 0.008300021290779114, + 0.007194431032985449, + 0.008036648854613304, + 0.00608322536572814, + 0.005553137976676226, + 0.007255895994603634, + 0.006831491831690073, + 0.005966111086308956, + 0.015912214294075966, + 0.011074945330619812, + 0.022550854831933975, + 0.012339798733592033, + 0.009731580503284931, + 0.06225075572729111, + 0.012997739017009735, + 0.04754055663943291, + 0.009915064089000225, + 0.009337243624031544, + 0.03282146155834198, + 0.016705099493265152, + 0.01042031217366457, + 0.012036626227200031, + 0.011535814963281155, + 0.016400199383497238, + 0.011558513157069683, + 0.004967099986970425, + 0.07988971471786499, + 0.017842601984739304, + 0.00956828985363245, + 0.02673828974366188, + 0.04118546470999718, + 0.00900427345186472, + 0.01710360497236252, + 0.007742380257695913, + 0.014778128825128078, + 0.009423606097698212, + 0.008237520232796669, + 0.010918200947344303, + 0.007872666232287884, + 0.010745528154075146, + 0.0026265366468578577, + 0.0023175696842372417, + 0.005879960488528013, + 0.0037350987549871206, + 0.003568043699488044, + 0.0033854583743959665, + 0.004321436397731304, + 0.004376785829663277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.007646576501429081 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308338597416878 + ], + "opposite_logprob_tokens": [ + -3.773308277130127 + ], + "logprob_reward": -0.023308338597416878, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308338597416878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27186426520347595, + 0.08232031017541885, + 0.034907661378383636, + 0.10699313879013062, + 0.03747836872935295, + 0.14715154469013214, + 0.11029615998268127, + 0.10171924531459808, + 0.08291545510292053, + 0.024353865534067154 + ], + "credit_map": [ + 0.023353450000286102, + 0.04202483221888542, + 0.029711369425058365, + 0.02509179525077343, + 0.04893257096409798, + 0.024484381079673767, + 0.0367467924952507, + 0.041519057005643845, + 0.014503143727779388, + 0.007946297526359558, + 0.022277358919382095, + 0.0053731282241642475, + 0.004414875525981188, + 0.009101923555135727, + 0.011699891649186611, + 0.007003696169704199, + 0.004441004246473312, + 0.0036745076067745686, + 0.005152310710400343, + 0.002855338156223297, + 0.003970201127231121, + 0.00624196557328105, + 0.004323436878621578, + 0.004248900339007378, + 0.011053926311433315, + 0.012491156347095966, + 0.021796172484755516, + 0.009985932148993015, + 0.011889254674315453, + 0.01150585152208805, + 0.012726504355669022, + 0.015544349327683449, + 0.005121635273098946, + 0.004197811242192984, + 0.004076555836945772, + 0.0034236491192132235, + 0.003135314444079995, + 0.00421269191429019, + 0.005205671768635511, + 0.008105040527880192, + 0.010492952540516853, + 0.008990103378891945, + 0.0181506909430027, + 0.010353675112128258, + 0.008508160710334778, + 0.053674980998039246, + 0.010571980848908424, + 0.02640899270772934, + 0.007857182063162327, + 0.007429555524140596, + 0.01939454674720764, + 0.015874488279223442, + 0.009063090197741985, + 0.011103981174528599, + 0.012129411101341248, + 0.027443913742899895, + 0.007493443787097931, + 0.004451303277164698, + 0.029919644817709923, + 0.009949217550456524, + 0.0069598364643752575, + 0.03100034035742283, + 0.006067671347409487, + 0.005877793300896883, + 0.014095200225710869, + 0.009243268519639969, + 0.01206958293914795, + 0.009744257666170597, + 0.007858171127736568, + 0.011374659836292267, + 0.008074169047176838, + 0.010456155054271221, + 0.0025675231590867043, + 0.00297136465087533, + 0.0034993847366422415, + 0.002719048410654068, + 0.003189927199855447, + 0.0028878236189484596, + 0.00299673224799335, + 0.0035220617428421974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.023308338597416878 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03463171795010567, + 0.02592989057302475, + 0.027917152270674706, + 0.020507322624325752, + 0.04550708457827568, + 0.3478768467903137, + 0.3217613995075226, + 0.1293882578611374, + 0.03412128612399101, + 0.012358995154500008 + ], + "credit_map": [ + 0.003629577811807394, + 0.005052021238952875, + 0.003720030188560486, + 0.003384837880730629, + 0.005151547957211733, + 0.003226407105103135, + 0.0033464892767369747, + 0.007120806258171797, + 0.002565833041444421, + 0.0025947701651602983, + 0.0053742434829473495, + 0.001977169420570135, + 0.0021219479385763407, + 0.003329778090119362, + 0.0049523585475981236, + 0.0030137887224555016, + 0.0038266584742814302, + 0.0029736682772636414, + 0.0033889974001795053, + 0.002750675892457366, + 0.002695742528885603, + 0.005857556127011776, + 0.0031419831793755293, + 0.003281873185187578, + 0.0021069145295768976, + 0.0022744780872017145, + 0.0020500710234045982, + 0.002740350319072604, + 0.002086135558784008, + 0.002144664293155074, + 0.0029675611294806004, + 0.00413714861497283, + 0.00657168123871088, + 0.0033068868797272444, + 0.004893432836979628, + 0.004601086489856243, + 0.0030432797502726316, + 0.004218634683638811, + 0.0051931170746684074, + 0.013678964227437973, + 0.029366321861743927, + 0.01965990476310253, + 0.04327050596475601, + 0.022643152624368668, + 0.01797953061759472, + 0.10514236241579056, + 0.02461964637041092, + 0.08519544452428818, + 0.027083981782197952, + 0.02393355965614319, + 0.09995373338460922, + 0.03424739092588425, + 0.02312363125383854, + 0.027425367385149002, + 0.024908917024731636, + 0.06108482927083969, + 0.007368248887360096, + 0.003805427346378565, + 0.05109051614999771, + 0.011987832374870777, + 0.006681510247290134, + 0.03704297915101051, + 0.005267610773444176, + 0.006144135259091854, + 0.004579044878482819, + 0.0025267398450523615, + 0.004956609103828669, + 0.005678902380168438, + 0.0031098483595997095, + 0.004864574875682592, + 0.0041986568830907345, + 0.004206910729408264, + 0.0020890950690954924, + 0.0013338890857994556, + 0.001983118010684848, + 0.0011817673221230507, + 0.0008573899394832551, + 0.0015214115846902132, + 0.0017709825187921524, + 0.0016213421477004886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.011080636642873287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 wide shot panning slowly to the right from a high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a steep, winding mountain ridge shrouded in mist under low golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thousands of ancient craftsmen visible working on a new section of the Great Wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the artisans include men in worn linen shorts dragging granite blocks with hemp ropes and wooden rollers while chanting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan down and zoom in to reveal stonemasons hammering stones on bamboo scaffolding?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639456048607826 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639456048607826, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639456048607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054885510355234146, + 0.061089783906936646, + 0.025851165875792503, + 0.033777035772800446, + 0.02583734318614006, + 0.2393031120300293, + 0.13450589776039124, + 0.31734558939933777, + 0.061423156410455704, + 0.045981407165527344 + ], + "credit_map": [ + 0.006970829330384731, + 0.006602948531508446, + 0.00586096802726388, + 0.00891630258411169, + 0.00685472134500742, + 0.00588179100304842, + 0.008414598181843758, + 0.005383354611694813, + 0.008641204796731472, + 0.0071606081910431385, + 0.007678348571062088, + 0.0046704919077456, + 0.005643453914672136, + 0.012191087938845158, + 0.0072867232374846935, + 0.007817868143320084, + 0.003544776001945138, + 0.0023478635121136904, + 0.0031155094038695097, + 0.002730058506131172, + 0.002786736935377121, + 0.004695675801485777, + 0.003454859834164381, + 0.0031756849493831396, + 0.0030879955738782883, + 0.0032498203217983246, + 0.004842783324420452, + 0.005228154361248016, + 0.0027047882322221994, + 0.003200314473360777, + 0.0060044629499316216, + 0.005458714906126261, + 0.0032568241003900766, + 0.0023702613543719053, + 0.003118367400020361, + 0.0030036482494324446, + 0.0023216509725898504, + 0.002596708480268717, + 0.002997864503413439, + 0.006172017194330692, + 0.02916267327964306, + 0.012136243283748627, + 0.030891025438904762, + 0.015215097926557064, + 0.011686938814818859, + 0.05319223180413246, + 0.018735982477664948, + 0.06828291714191437, + 0.012902043759822845, + 0.010953071527183056, + 0.0418831929564476, + 0.019472042098641396, + 0.009540656581521034, + 0.010739585384726524, + 0.010939995758235455, + 0.018075309693813324, + 0.017890464514493942, + 0.008632230572402477, + 0.12973374128341675, + 0.0280600693076849, + 0.01546338852494955, + 0.05506953224539757, + 0.04853707179427147, + 0.013959084637463093, + 0.007103447336703539, + 0.003087203251197934, + 0.01642807200551033, + 0.007114452309906483, + 0.005862642545253038, + 0.009368734434247017, + 0.0049459077417850494, + 0.007512695621699095, + 0.007988683879375458, + 0.0041886079125106335, + 0.007187835406512022, + 0.004786447621881962, + 0.004141421057283878, + 0.007138388231396675, + 0.004610379692167044, + 0.005939646158367395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.004639456048607826 + } + ], + "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/0000155/group_3/index_tensor([3699, 3699])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_tensor([3699, 3699])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c1890c003e3852b6a14ff9279871dd0cf9cf3805 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_tensor([3699, 3699])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a scene imbued with authentic CGI cinematic realism, a sweeping wide shot unfolds in slow, graceful motion around a massive lion's mane jellyfish suspended in the deep, abyssal azure of the ocean. The creature's massive, domed bell, emitting a vibrant, fiery orange-red hue, pulsates gently and rhythmically, its gelatinous form contracting steadily and deliberately, propelling itself forward with silent, effortless force. A vast, intricate network of countless tiny, thread-like tentacles, some as thick as ribbons, others as delicate as silk, floats behind it in a tangled, fluid mass emitting a soft, ethereal blue-green bioluminescence. In the distance, wisps of brilliant sunlight filter down into shimmering divine light, piercing the darkness and illuminating the jellyfish's translucent body, revealing its intricate internal structure and highlighting tiny grains of marine snow that drift lazily through the water like suspended dust. The camera’s fluid, arcing movements capture the majestic and serene scene from every angle, emphasizing the creature’s breathtaking size and the hypnotic beauty of its slow, graceful dance through the depths.", + "seed": 2792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_3699_seed_2792_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": -0.09898560158908368, + "vlm_logprob": -0.09898560158908368, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.09898560158908368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03727797046303749 + ], + "opposite_logprob_tokens": [ + -3.412277936935425 + ], + "logprob_reward": -0.03727797046303749, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03727797046303749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1695728749036789, + 0.12663888931274414, + 0.14404305815696716, + 0.06245281919836998, + 0.04720817506313324, + 0.14964613318443298, + 0.09749304503202438, + 0.0689578428864479, + 0.0562310516834259, + 0.07775614410638809 + ], + "credit_map": [ + 0.026298610493540764, + 0.011703704483807087, + 0.029927264899015427, + 0.01778678223490715, + 0.02412988804280758, + 0.01170341856777668, + 0.027517778798937798, + 0.02050541527569294, + 0.01901998557150364, + 0.009769409894943237, + 0.023106569424271584, + 0.018526574596762657, + 0.011404749937355518, + 0.01080884039402008, + 0.015141495503485203, + 0.018861252814531326, + 0.018770506605505943, + 0.014832359738647938, + 0.023968132212758064, + 0.01041064690798521, + 0.017303116619586945, + 0.01788312941789627, + 0.01810736209154129, + 0.022767797112464905, + 0.008279731497168541, + 0.006590140052139759, + 0.009055336937308311, + 0.005900268442928791, + 0.010226407088339329, + 0.004954117815941572, + 0.009595823474228382, + 0.007850995287299156, + 0.007220088504254818, + 0.003117005806416273, + 0.006774197332561016, + 0.004180708900094032, + 0.004762241616845131, + 0.004267858806997538, + 0.008191762492060661, + 0.008694309741258621, + 0.015025535598397255, + 0.011584261432290077, + 0.02176065370440483, + 0.015411771833896637, + 0.02110186032950878, + 0.009158237837255001, + 0.031218720600008965, + 0.024385081604123116, + 0.01411577221006155, + 0.00856013223528862, + 0.012363058514893055, + 0.014853869564831257, + 0.012777753174304962, + 0.006394923664629459, + 0.013579480350017548, + 0.014848055317997932, + 0.015241440385580063, + 0.004908604547381401, + 0.012981848791241646, + 0.005887881852686405, + 0.0058104004710912704, + 0.004842990543693304, + 0.00820558425039053, + 0.011079087853431702, + 0.006186259910464287, + 0.006068769842386246, + 0.006872008088976145, + 0.0060943616554141045, + 0.009149872697889805, + 0.004610300529748201, + 0.008221875876188278, + 0.009027602151036263, + 0.011574812233448029, + 0.007539854384958744, + 0.01072725746780634, + 0.007813012227416039, + 0.012060029432177544, + 0.005423518363386393, + 0.011031515896320343, + 0.011586141772568226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.03727797046303749 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212341979146004 + ], + "opposite_logprob_tokens": [ + -4.264212131500244 + ], + "logprob_reward": -0.014212341979146004, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212341979146004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1759188026189804, + 0.19904357194900513, + 0.24430467188358307, + 0.0807940810918808, + 0.04923897981643677, + 0.05611138045787811, + 0.039394330233335495, + 0.08649315685033798, + 0.036567967385053635, + 0.032133106142282486 + ], + "credit_map": [ + 0.030242273584008217, + 0.011822785250842571, + 0.026138976216316223, + 0.023515135049819946, + 0.024690691381692886, + 0.012884399853646755, + 0.027533816173672676, + 0.019090719521045685, + 0.022828154265880585, + 0.024630913510918617, + 0.030304379761219025, + 0.024384047836065292, + 0.027043862268328667, + 0.015800293534994125, + 0.029117345809936523, + 0.024934574961662292, + 0.02845650166273117, + 0.025375064462423325, + 0.03497076779603958, + 0.012078541330993176, + 0.028127169236540794, + 0.04448382556438446, + 0.020972227677702904, + 0.049840573221445084, + 0.008883313275873661, + 0.0077658360823988914, + 0.013442271389067173, + 0.0064206430688500404, + 0.013939104042947292, + 0.005835775751620531, + 0.011825033463537693, + 0.012682104483246803, + 0.00594820873811841, + 0.0025874695274978876, + 0.008020737208425999, + 0.004797460976988077, + 0.004794990178197622, + 0.0034346207976341248, + 0.008012277074158192, + 0.011643216013908386, + 0.006389237008988857, + 0.004378810990601778, + 0.008440270088613033, + 0.006087688263505697, + 0.00760999321937561, + 0.0030817468650639057, + 0.011762301437556744, + 0.008361331187188625, + 0.007423056289553642, + 0.002419034019112587, + 0.006901501677930355, + 0.00557178258895874, + 0.0039473543874919415, + 0.002467963146045804, + 0.005083753727376461, + 0.005579881835728884, + 0.012448922730982304, + 0.00737174553796649, + 0.013797683641314507, + 0.008845874108374119, + 0.012377132661640644, + 0.006002242211252451, + 0.012596678920090199, + 0.013052879832684994, + 0.004057880956679583, + 0.0034966454841196537, + 0.004587982315570116, + 0.005193136632442474, + 0.004899895749986172, + 0.0025863773189485073, + 0.005646116565912962, + 0.006099931430071592, + 0.004338148981332779, + 0.0026156778912991285, + 0.0042077600955963135, + 0.003865279955789447, + 0.0033087292686104774, + 0.0019080493366345763, + 0.004935690201818943, + 0.006953770760446787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014212341979146004 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204032897949219 + ], + "opposite_logprob_tokens": [ + -2.812040328979492 + ], + "logprob_reward": -0.06204032897949219, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204032897949219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13791146874427795, + 0.20198938250541687, + 0.15332496166229248, + 0.05780481547117233, + 0.08367076516151428, + 0.047886766493320465, + 0.07225950807332993, + 0.13295811414718628, + 0.05858093500137329, + 0.05361327901482582 + ], + "credit_map": [ + 0.02711823396384716, + 0.010085330344736576, + 0.024246830493211746, + 0.019089611247181892, + 0.016882775351405144, + 0.0091243339702487, + 0.016656581312417984, + 0.014707772061228752, + 0.037716418504714966, + 0.016693703830242157, + 0.037566136568784714, + 0.023197762668132782, + 0.016405712813138962, + 0.018734905868768692, + 0.023519527167081833, + 0.028155207633972168, + 0.02524181269109249, + 0.017012156546115875, + 0.029950886964797974, + 0.01466342993080616, + 0.018781349062919617, + 0.015050948597490788, + 0.016502385959029198, + 0.016121990978717804, + 0.00908052921295166, + 0.005210463888943195, + 0.009794811718165874, + 0.005433814134448767, + 0.007855745032429695, + 0.004889010917395353, + 0.007983426563441753, + 0.007557016797363758, + 0.014049223624169827, + 0.005529749672859907, + 0.012994393706321716, + 0.009137095883488655, + 0.009681403636932373, + 0.009076987393200397, + 0.012290959246456623, + 0.010910957120358944, + 0.005658055189996958, + 0.004209465347230434, + 0.006972116883844137, + 0.004091464914381504, + 0.008846206590533257, + 0.0032493295148015022, + 0.008325045928359032, + 0.006535081658512354, + 0.00809707771986723, + 0.005103969946503639, + 0.01134073082357645, + 0.012348084710538387, + 0.00677458057180047, + 0.004929041024297476, + 0.012235680595040321, + 0.011430345475673676, + 0.03799137473106384, + 0.009998735971748829, + 0.021587392315268517, + 0.010114820674061775, + 0.0091411666944623, + 0.01051531732082367, + 0.012872268445789814, + 0.020737046375870705, + 0.005560882389545441, + 0.0053041111677885056, + 0.006960728671401739, + 0.008470495231449604, + 0.009113341569900513, + 0.003360540373250842, + 0.007257602643221617, + 0.012553229928016663, + 0.011409864760935307, + 0.0037443845067173243, + 0.007232146803289652, + 0.004343415144830942, + 0.006901158951222897, + 0.0032974339555948973, + 0.006107771769165993, + 0.010577105917036533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.06204032897949219 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015380859375 + ], + "opposite_logprob_tokens": [ + -1.7015380859375 + ], + "logprob_reward": -0.2015380859375, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015380859375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19098463654518127, + 0.22028855979442596, + 0.0772663876414299, + 0.05309288203716278, + 0.10906472057104111, + 0.06184445321559906, + 0.10928872227668762, + 0.09796349704265594, + 0.04720066115260124, + 0.03300553560256958 + ], + "credit_map": [ + 0.029960934072732925, + 0.016194447875022888, + 0.029896946623921394, + 0.021363340318202972, + 0.035840123891830444, + 0.012450417503714561, + 0.030273860320448875, + 0.015004565007984638, + 0.02638275735080242, + 0.02284446731209755, + 0.0331655889749527, + 0.044146399945020676, + 0.02079351805150509, + 0.019424432888627052, + 0.028962748125195503, + 0.02456863969564438, + 0.014485465362668037, + 0.009732766076922417, + 0.01326381042599678, + 0.006674068048596382, + 0.010604227893054485, + 0.006299748085439205, + 0.009965977631509304, + 0.006240320857614279, + 0.007118194829672575, + 0.005583102349191904, + 0.00808040238916874, + 0.005439148750156164, + 0.007038832176476717, + 0.004426890984177589, + 0.009945213794708252, + 0.005461098160594702, + 0.014271102845668793, + 0.00618481682613492, + 0.0195781122893095, + 0.014234309084713459, + 0.009422706440091133, + 0.012686096131801605, + 0.015700791031122208, + 0.016986776143312454, + 0.009703502990305424, + 0.004363679792732, + 0.011210913769900799, + 0.006694090086966753, + 0.008153618313372135, + 0.0031794284004718065, + 0.011690057814121246, + 0.0068491674028337, + 0.014124776236712933, + 0.009241731837391853, + 0.014215263538062572, + 0.01612093858420849, + 0.012458807788789272, + 0.008307256735861301, + 0.017619241029024124, + 0.017200710251927376, + 0.01549108512699604, + 0.00769334752112627, + 0.016519643366336823, + 0.008768715895712376, + 0.011547732166945934, + 0.006286378018558025, + 0.01680653728544712, + 0.014850056730210781, + 0.005321416072547436, + 0.0036347522400319576, + 0.006306598894298077, + 0.0069859810173511505, + 0.004969718866050243, + 0.0033455241937190294, + 0.005668492056429386, + 0.010968174785375595, + 0.006625343579798937, + 0.002043310320004821, + 0.0060729095712304115, + 0.002674301154911518, + 0.0029548495076596737, + 0.0018036712426692247, + 0.005334218963980675, + 0.00549692939966917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.2015380859375 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985928058624268 + ], + "opposite_logprob_tokens": [ + -1.8048592805862427 + ], + "logprob_reward": -0.17985928058624268, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985928058624268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18349327147006989, + 0.09508681297302246, + 0.052229657769203186, + 0.06501371413469315, + 0.07704310864210129, + 0.046780508011579514, + 0.07324559986591339, + 0.276883602142334, + 0.07285264879465103, + 0.057371046394109726 + ], + "credit_map": [ + 0.042180050164461136, + 0.010691690258681774, + 0.044464875012636185, + 0.016078250482678413, + 0.017959609627723694, + 0.010908705182373524, + 0.0238150916993618, + 0.017395000904798508, + 0.018389133736491203, + 0.007675790227949619, + 0.02300594188272953, + 0.010565525852143764, + 0.007833666168153286, + 0.007305474951863289, + 0.009402351453900337, + 0.0109089445322752, + 0.01113998331129551, + 0.004222062416374683, + 0.0077094463631510735, + 0.0037906994111835957, + 0.005352766253054142, + 0.005184876266866922, + 0.005717013962566853, + 0.009112813510000706, + 0.012052737176418304, + 0.004745044745504856, + 0.010945525020360947, + 0.004294070415198803, + 0.008510332554578781, + 0.005551449954509735, + 0.009678993374109268, + 0.009235559031367302, + 0.012059486471116543, + 0.005827011540532112, + 0.012000083923339844, + 0.0079930629581213, + 0.00897299312055111, + 0.006575249135494232, + 0.012317907996475697, + 0.011297324672341347, + 0.007796168327331543, + 0.003696650732308626, + 0.008676685392856598, + 0.004377863835543394, + 0.005022125784307718, + 0.0025741595309227705, + 0.006723486818373203, + 0.007913372479379177, + 0.010293333791196346, + 0.0048772795125842094, + 0.009872648864984512, + 0.007975632324814796, + 0.0065244268625974655, + 0.005647103767842054, + 0.00828166026622057, + 0.01977352425456047, + 0.07360100001096725, + 0.01855536550283432, + 0.04473063722252846, + 0.02864980138838291, + 0.01600133441388607, + 0.01680680178105831, + 0.030564870685338974, + 0.047973815351724625, + 0.008181924931704998, + 0.005810852628201246, + 0.00988076813519001, + 0.010215741582214832, + 0.006291746627539396, + 0.005044696386903524, + 0.00975862331688404, + 0.01766829565167427, + 0.012011411599814892, + 0.0055601089261472225, + 0.009504306130111217, + 0.005425731651484966, + 0.005976652726531029, + 0.0035003896337002516, + 0.005455755162984133, + 0.009936696849763393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.17985928058624268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene imbued with authentic CGI cinematic realism, a sweeping wide shot unfolds in slow, graceful motion around a massive lion's mane jellyfish suspended in the deep, abyssal azure of the ocean. The creature's massive, domed bell, emitting a vibrant, fiery orange-red hue, pulsates gently and rhythmically, its gelatinous form contracting steadily and deliberately, propelling itself forward with silent, effortless force. A vast, intricate network of countless tiny, thread-like tentacles, some as thick as ribbons, others as delicate as silk, floats behind it in a tangled, fluid mass emitting a soft, ethereal blue-green bioluminescence. In the distance, wisps of brilliant sunlight filter down into shimmering divine light, piercing the darkness and illuminating the jellyfish's translucent body, revealing its intricate internal structure and highlighting tiny grains of marine snow that drift lazily through the water like suspended dust. The camera’s fluid, arcing movements capture the majestic and serene scene from every angle, emphasizing the creature’s breathtaking size and the hypnotic beauty of its slow, graceful dance through the depths.", + "seed": 2793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_3699_seed_2793_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": -0.1160164300352335, + "vlm_logprob": -0.1160164300352335, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.1160164300352335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08413136005401611 + ], + "opposite_logprob_tokens": [ + -2.5841312408447266 + ], + "logprob_reward": -0.08413136005401611, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08413136005401611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08135519176721573, + 0.17306683957576752, + 0.049624115228652954, + 0.10022763162851334, + 0.026954445987939835, + 0.21157024800777435, + 0.02468251623213291, + 0.02056865580379963, + 0.05668945237994194, + 0.2552609443664551 + ], + "credit_map": [ + 0.005652994383126497, + 0.016506478190422058, + 0.009999619796872139, + 0.008640630170702934, + 0.01262429915368557, + 0.01097950804978609, + 0.005053043365478516, + 0.011898616328835487, + 0.011618058197200298, + 0.02784043177962303, + 0.02078854665160179, + 0.019488634541630745, + 0.025907263159751892, + 0.022319583222270012, + 0.018755270168185234, + 0.026349050924181938, + 0.006608532276004553, + 0.007258747238665819, + 0.00705401785671711, + 0.006402354687452316, + 0.005834947805851698, + 0.005597664974629879, + 0.005118581932038069, + 0.00574926333501935, + 0.010473743081092834, + 0.010732617229223251, + 0.015564657747745514, + 0.018680786713957787, + 0.010147455148398876, + 0.009750967845320702, + 0.01246744953095913, + 0.012409958988428116, + 0.0020770460832864046, + 0.003975340165197849, + 0.0031916776206344366, + 0.003416914027184248, + 0.003792708972468972, + 0.005308843683451414, + 0.0022535251919180155, + 0.0029383902437984943, + 0.05725284293293953, + 0.027465330436825752, + 0.029961613938212395, + 0.020542211830615997, + 0.0242215134203434, + 0.019299019128084183, + 0.01439602766185999, + 0.01843169517815113, + 0.006283995229750872, + 0.004108766093850136, + 0.002806831616908312, + 0.002220612484961748, + 0.0023208747152239084, + 0.001965089701116085, + 0.0021466873586177826, + 0.0028296601958572865, + 0.0021439662668854, + 0.002446545520797372, + 0.0029378144536167383, + 0.002624192973598838, + 0.0023332794662564993, + 0.002741818316280842, + 0.002419714815914631, + 0.0029213253874331713, + 0.005399538669735193, + 0.008565117605030537, + 0.012090209871530533, + 0.0059350961819291115, + 0.0078063481487333775, + 0.006019073072820902, + 0.005608082748949528, + 0.005265981890261173, + 0.016699928790330887, + 0.08138739317655563, + 0.014432566240429878, + 0.013805942609906197, + 0.01961449719965458, + 0.013534478843212128, + 0.02840598300099373, + 0.0673801377415657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.08413136005401611 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008662031963467598 + ], + "opposite_logprob_tokens": [ + -4.758662223815918 + ], + "logprob_reward": -0.008662031963467598, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662031963467598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14067386090755463, + 0.1425723433494568, + 0.05287126824259758, + 0.07015082985162735, + 0.11737177520990372, + 0.07854732125997543, + 0.021269382908940315, + 0.03092818893492222, + 0.03336138650774956, + 0.3122536242008209 + ], + "credit_map": [ + 0.012349426746368408, + 0.027260592207312584, + 0.018070369958877563, + 0.017031827941536903, + 0.023319996893405914, + 0.014915751293301582, + 0.01095365546643734, + 0.01677226461470127, + 0.016779296100139618, + 0.02509337104856968, + 0.012749433517456055, + 0.01334820594638586, + 0.02222413569688797, + 0.024249466136097908, + 0.011861670762300491, + 0.01626676321029663, + 0.006874502170830965, + 0.0064089857041835785, + 0.007302260026335716, + 0.006607827730476856, + 0.006511148996651173, + 0.00965543370693922, + 0.004636501427739859, + 0.004874617327004671, + 0.009355937130749226, + 0.009228482842445374, + 0.011459455825388432, + 0.011845048516988754, + 0.007747895549982786, + 0.008048515766859055, + 0.004732353612780571, + 0.007733148988336325, + 0.012938616797327995, + 0.016098717227578163, + 0.01222915481775999, + 0.014260513707995415, + 0.018451793119311333, + 0.02120192162692547, + 0.009912259876728058, + 0.012278799898922443, + 0.013770179823040962, + 0.006123061757534742, + 0.01666492596268654, + 0.008368626236915588, + 0.00871483888477087, + 0.00768679054453969, + 0.007404898293316364, + 0.009814007207751274, + 0.0020682758186012506, + 0.0034804728347808123, + 0.0022060766350477934, + 0.0027977111749351025, + 0.0023292209953069687, + 0.0026348538231104612, + 0.002984232036396861, + 0.002768539823591709, + 0.0027090823277831078, + 0.005156124942004681, + 0.004136594478040934, + 0.0030062810983508825, + 0.0053695919923484325, + 0.00288070528768003, + 0.004011507611721754, + 0.00365830329246819, + 0.003610143205150962, + 0.003772915806621313, + 0.0036822292022407055, + 0.004375285003334284, + 0.003007485531270504, + 0.006394799333065748, + 0.004056448582559824, + 0.004462078679352999, + 0.02046835608780384, + 0.10266222059726715, + 0.013751806691288948, + 0.015593829564750195, + 0.023713314905762672, + 0.01651878096163273, + 0.035667285323143005, + 0.0838780403137207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.008662031963467598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703965604305267 + ], + "opposite_logprob_tokens": [ + -2.127039670944214 + ], + "logprob_reward": -0.12703965604305267, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703965604305267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13682004809379578, + 0.11886540800333023, + 0.039579834789037704, + 0.06970690190792084, + 0.09574628621339798, + 0.09910612553358078, + 0.04506712034344673, + 0.0221345704048872, + 0.0407024547457695, + 0.3322713077068329 + ], + "credit_map": [ + 0.01119362935423851, + 0.028474509716033936, + 0.016595877707004547, + 0.0178449135273695, + 0.020823659375309944, + 0.018101179972290993, + 0.01064315065741539, + 0.013143123127520084, + 0.009496655315160751, + 0.027797361835837364, + 0.01296764425933361, + 0.011001220904290676, + 0.021632397547364235, + 0.018349401652812958, + 0.0069503178820014, + 0.010670417919754982, + 0.0069697848521173, + 0.004348446149379015, + 0.005735700484365225, + 0.004853439517319202, + 0.005245031788945198, + 0.005900122225284576, + 0.003422964597120881, + 0.0031043451745063066, + 0.006886489689350128, + 0.007959726266562939, + 0.01083288062363863, + 0.014948391355574131, + 0.00709885498508811, + 0.0075378092005848885, + 0.00577495526522398, + 0.008667794987559319, + 0.012090706266462803, + 0.013337884098291397, + 0.00981864519417286, + 0.011735889129340649, + 0.014357236213982105, + 0.016144294291734695, + 0.009438170120120049, + 0.008823463693261147, + 0.011864807456731796, + 0.006453596521168947, + 0.014399378560483456, + 0.009187609888613224, + 0.012724033556878567, + 0.011741550639271736, + 0.018385088071227074, + 0.014350065030157566, + 0.010141570121049881, + 0.007932846434414387, + 0.004021072760224342, + 0.004124751780182123, + 0.005079879891127348, + 0.0038412741851061583, + 0.0062057687900960445, + 0.003719958243891597, + 0.001527405926026404, + 0.0025636395439505577, + 0.002917736768722534, + 0.0024465443566441536, + 0.0029288313817232847, + 0.0028528738766908646, + 0.003624080214649439, + 0.0032734577544033527, + 0.004022248089313507, + 0.007262319326400757, + 0.009806751273572445, + 0.0036770382430404425, + 0.0032096293289214373, + 0.004285051487386227, + 0.003970972262322903, + 0.004468446597456932, + 0.023251300677657127, + 0.10754405707120895, + 0.016308216378092766, + 0.017817223444581032, + 0.024672869592905045, + 0.016534041613340378, + 0.03690696135163307, + 0.08923660963773727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.12703965604305267 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896840572357178 + ], + "opposite_logprob_tokens": [ + -2.5789685249328613 + ], + "logprob_reward": -0.07896840572357178, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896840572357178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19277651607990265, + 0.05691701918840408, + 0.12658117711544037, + 0.04425719007849693, + 0.021066579967737198, + 0.08228269219398499, + 0.045384883880615234, + 0.022430872544646263, + 0.026085859164595604, + 0.38221731781959534 + ], + "credit_map": [ + 0.015747474506497383, + 0.048072878271341324, + 0.021930282935500145, + 0.02379128523170948, + 0.02719537541270256, + 0.023882444947957993, + 0.012506281957030296, + 0.019650468602776527, + 0.004481630399823189, + 0.006778560113161802, + 0.008573506027460098, + 0.005162352696061134, + 0.007965822704136372, + 0.00981367751955986, + 0.005863911472260952, + 0.008277553133666515, + 0.020131224766373634, + 0.021390050649642944, + 0.019217174500226974, + 0.015032078139483929, + 0.014777563512325287, + 0.01431005448102951, + 0.009809784591197968, + 0.011913232505321503, + 0.0036994311958551407, + 0.004979586228728294, + 0.008458892814815044, + 0.005686286371201277, + 0.006502074655145407, + 0.005685140378773212, + 0.004451125394552946, + 0.004794652108103037, + 0.001547070685774088, + 0.0015944417100399733, + 0.002426593564450741, + 0.001995147205889225, + 0.00337928906083107, + 0.0037805954925715923, + 0.0037705935537815094, + 0.0025728452019393444, + 0.007720982190221548, + 0.00655457703396678, + 0.008945480920374393, + 0.012361218221485615, + 0.010319479741156101, + 0.010137280449271202, + 0.014021988958120346, + 0.012221683748066425, + 0.00899031013250351, + 0.008181851357221603, + 0.004993528127670288, + 0.004671643953770399, + 0.004919429309666157, + 0.00429542176425457, + 0.004336550831794739, + 0.004996146075427532, + 0.002255067927762866, + 0.0029757062438875437, + 0.002783112693578005, + 0.00211353856138885, + 0.003147561801597476, + 0.003438870655372739, + 0.002652401803061366, + 0.003064611693844199, + 0.0020722271874547005, + 0.006434723734855652, + 0.0025199262890964746, + 0.002911655930802226, + 0.002198012312874198, + 0.0041104573756456375, + 0.0025108368135988712, + 0.003328016260638833, + 0.025323010981082916, + 0.12852978706359863, + 0.011740167625248432, + 0.019277207553386688, + 0.02860678918659687, + 0.018974360078573227, + 0.04443548992276192, + 0.10533048212528229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07896840572357178 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812806963920593 + ], + "opposite_logprob_tokens": [ + -1.406280755996704 + ], + "logprob_reward": -0.2812806963920593, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812806963920593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2989712357521057, + 0.11209134012460709, + 0.04019210487604141, + 0.06310086697340012, + 0.10365240275859833, + 0.116454117000103, + 0.04451366886496544, + 0.029750991612672806, + 0.02707740105688572, + 0.16419586539268494 + ], + "credit_map": [ + 0.022922636941075325, + 0.07341396808624268, + 0.03161953017115593, + 0.04119912162423134, + 0.04532567039132118, + 0.03126082569360733, + 0.022071925923228264, + 0.031157545745372772, + 0.008220523595809937, + 0.018929017707705498, + 0.011107977479696274, + 0.010724175721406937, + 0.017340103164315224, + 0.020970886573195457, + 0.008112012408673763, + 0.01668664626777172, + 0.004268467891961336, + 0.005069423466920853, + 0.00525396503508091, + 0.005060046445578337, + 0.004948442801833153, + 0.006865558214485645, + 0.004531083628535271, + 0.004195116460323334, + 0.003773608710616827, + 0.006276512518525124, + 0.008786100894212723, + 0.007754113059490919, + 0.009222033433616161, + 0.011109109036624432, + 0.007354152854532003, + 0.008825237862765789, + 0.011750963516533375, + 0.011657372117042542, + 0.009325454011559486, + 0.01029282622039318, + 0.017410147935152054, + 0.022257938981056213, + 0.010059633292257786, + 0.010898063890635967, + 0.009621876291930676, + 0.009475798346102238, + 0.014750751666724682, + 0.012905700132250786, + 0.014401013031601906, + 0.012772483751177788, + 0.021046573296189308, + 0.021479928866028786, + 0.004330760799348354, + 0.006205094046890736, + 0.0049504381604492664, + 0.006375746335834265, + 0.004465189296752214, + 0.005384725518524647, + 0.004794855136424303, + 0.008006858639419079, + 0.00204979395493865, + 0.0037765544839203358, + 0.003475820878520608, + 0.0030249895062297583, + 0.00622984254732728, + 0.0035116388462483883, + 0.0034580181818455458, + 0.004224332515150309, + 0.002422034740447998, + 0.002610712544992566, + 0.004429305903613567, + 0.002270143013447523, + 0.002532100770622492, + 0.005303880199790001, + 0.002764091594144702, + 0.00474513228982687, + 0.009549970738589764, + 0.05199860781431198, + 0.009833966381847858, + 0.011084780097007751, + 0.012565077282488346, + 0.008208043873310089, + 0.018765991553664207, + 0.04218941181898117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.2812806963920593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene imbued with authentic CGI cinematic realism, a sweeping wide shot unfolds in slow, graceful motion around a massive lion's mane jellyfish suspended in the deep, abyssal azure of the ocean. The creature's massive, domed bell, emitting a vibrant, fiery orange-red hue, pulsates gently and rhythmically, its gelatinous form contracting steadily and deliberately, propelling itself forward with silent, effortless force. A vast, intricate network of countless tiny, thread-like tentacles, some as thick as ribbons, others as delicate as silk, floats behind it in a tangled, fluid mass emitting a soft, ethereal blue-green bioluminescence. In the distance, wisps of brilliant sunlight filter down into shimmering divine light, piercing the darkness and illuminating the jellyfish's translucent body, revealing its intricate internal structure and highlighting tiny grains of marine snow that drift lazily through the water like suspended dust. The camera’s fluid, arcing movements capture the majestic and serene scene from every angle, emphasizing the creature’s breathtaking size and the hypnotic beauty of its slow, graceful dance through the depths.", + "seed": 2794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_3699_seed_2794_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": -0.050466219801455735, + "vlm_logprob": -0.050466219801455735, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.050466219801455735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016633637249469757 + ], + "opposite_logprob_tokens": [ + -4.3916335105896 + ], + "logprob_reward": -0.016633637249469757, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016633637249469757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33685970306396484, + 0.10591667145490646, + 0.06226126849651337, + 0.07453350722789764, + 0.08627469092607498, + 0.019087690860033035, + 0.08676015585660934, + 0.06671728938817978, + 0.04511513188481331, + 0.11647389829158783 + ], + "credit_map": [ + 0.08192510902881622, + 0.024361908435821533, + 0.046584781259298325, + 0.031164206564426422, + 0.04862412437796593, + 0.035204190760850906, + 0.021796815097332, + 0.04719853028655052, + 0.007899378426373005, + 0.009539805352687836, + 0.015064206905663013, + 0.019269440323114395, + 0.011941870674490929, + 0.02173505537211895, + 0.009731177240610123, + 0.010735728777945042, + 0.006388076581060886, + 0.00456761009991169, + 0.010715936310589314, + 0.007050055544823408, + 0.006646525114774704, + 0.011725982651114464, + 0.00663463119417429, + 0.008532444946467876, + 0.007554805371910334, + 0.006607505492866039, + 0.011640466749668121, + 0.008070715703070164, + 0.009721798822283745, + 0.009669358842074871, + 0.009473306126892567, + 0.011795544065535069, + 0.005742564797401428, + 0.010785755701363087, + 0.015074419789016247, + 0.012209420092403889, + 0.00960040744394064, + 0.01219683513045311, + 0.008363659493625164, + 0.012301626615226269, + 0.0018943310715258121, + 0.002178634749725461, + 0.002516792621463537, + 0.001807518070563674, + 0.001821802114136517, + 0.0027809145394712687, + 0.002811235375702381, + 0.0032764605712145567, + 0.006405890919268131, + 0.014456975273787975, + 0.006170154083520174, + 0.007975983433425426, + 0.007602694444358349, + 0.012462954968214035, + 0.019784651696681976, + 0.011900843121111393, + 0.0042976802214980125, + 0.005735095124691725, + 0.005182713270187378, + 0.007845492102205753, + 0.0090244822204113, + 0.006743771955370903, + 0.01850147172808647, + 0.009386577643454075, + 0.003733282908797264, + 0.005192209966480732, + 0.0038298647850751877, + 0.009348844178020954, + 0.008833114057779312, + 0.0037838437128812075, + 0.005311190616339445, + 0.005082778632640839, + 0.014135386794805527, + 0.009350119158625603, + 0.014517931267619133, + 0.016321713104844093, + 0.012113847769796848, + 0.026501024141907692, + 0.008570520207285881, + 0.014963342808187008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.016633637249469757 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00596813065931201 + ], + "opposite_logprob_tokens": [ + -5.13096809387207 + ], + "logprob_reward": -0.00596813065931201, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00596813065931201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12384401261806488, + 0.11969169229269028, + 0.062373146414756775, + 0.3844867944717407, + 0.03756564483046532, + 0.035543739795684814, + 0.08342680335044861, + 0.058445192873477936, + 0.035748012363910675, + 0.0588749535381794 + ], + "credit_map": [ + 0.015350583009421825, + 0.010883692651987076, + 0.018336329609155655, + 0.014435735531151295, + 0.019132649526000023, + 0.016582520678639412, + 0.012652358040213585, + 0.01647014729678631, + 0.011600293219089508, + 0.008405587635934353, + 0.013007017783820629, + 0.016633201390504837, + 0.01822272688150406, + 0.028273532167077065, + 0.00866137258708477, + 0.014887955039739609, + 0.004329710267484188, + 0.005336505826562643, + 0.01715496927499771, + 0.004127698019146919, + 0.0033837484661489725, + 0.007908768951892853, + 0.004421942867338657, + 0.015709802508354187, + 0.027858734130859375, + 0.03207036480307579, + 0.1197475865483284, + 0.05519683286547661, + 0.01042442861944437, + 0.034042298793792725, + 0.016093380749225616, + 0.08905316889286041, + 0.002772210631519556, + 0.002981459489092231, + 0.0037171931471675634, + 0.0036165418568998575, + 0.006622989196330309, + 0.006614817772060633, + 0.005455820821225643, + 0.005784611217677593, + 0.0032326262444257736, + 0.003961767535656691, + 0.003694121493026614, + 0.0036904122680425644, + 0.0033002805430442095, + 0.006255099084228277, + 0.004961138125509024, + 0.006448293570429087, + 0.006041227839887142, + 0.01004057563841343, + 0.006191966123878956, + 0.0074403914622962475, + 0.005134919658303261, + 0.019561270251870155, + 0.01884426362812519, + 0.010172190144658089, + 0.0031203876715153456, + 0.0055316598154604435, + 0.0035589716862887144, + 0.0063715470023453236, + 0.007758420426398516, + 0.010875537060201168, + 0.00835139025002718, + 0.012877275235950947, + 0.0032354919239878654, + 0.00461120530962944, + 0.0036624749191105366, + 0.006323680747300386, + 0.004509808961302042, + 0.004089333117008209, + 0.00464946124702692, + 0.004666555672883987, + 0.007370895706117153, + 0.004490958992391825, + 0.005745966453105211, + 0.009773901663720608, + 0.007087790407240391, + 0.011561327613890171, + 0.004601515829563141, + 0.008242594078183174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00596813065931201 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1548348069190979, + 0.11843901872634888, + 0.06041061133146286, + 0.12848533689975739, + 0.041057880967855453, + 0.021034279838204384, + 0.1250711977481842, + 0.11249350756406784, + 0.028391629457473755, + 0.209781676530838 + ], + "credit_map": [ + 0.03666459396481514, + 0.010735513642430305, + 0.023102663457393646, + 0.01845070905983448, + 0.01669011265039444, + 0.016514845192432404, + 0.01157689094543457, + 0.021099494770169258, + 0.006261460483074188, + 0.007393606938421726, + 0.009720145724713802, + 0.012396751902997494, + 0.024986812844872475, + 0.0336683914065361, + 0.009238461963832378, + 0.014773395843803883, + 0.0036764005199074745, + 0.004856786224991083, + 0.01762246899306774, + 0.005191667936742306, + 0.0052615744061768055, + 0.007853900082409382, + 0.003663355251774192, + 0.012284459546208382, + 0.009434320032596588, + 0.010940918698906898, + 0.03975810110569, + 0.01825387217104435, + 0.005896265618503094, + 0.011288897134363651, + 0.005996728781610727, + 0.02691623568534851, + 0.0024949994403868914, + 0.0029583151917904615, + 0.0024737692438066006, + 0.0026419395580887794, + 0.00721947243437171, + 0.010811656713485718, + 0.00535618094727397, + 0.0071015493012964725, + 0.001503337756730616, + 0.0020308794919401407, + 0.00230752513743937, + 0.0025236255023628473, + 0.002436236245557666, + 0.003999816719442606, + 0.0034044140484184027, + 0.00282844714820385, + 0.009750706143677235, + 0.015718994662165642, + 0.01030901912599802, + 0.009160885587334633, + 0.008308183401823044, + 0.029637616127729416, + 0.027195274829864502, + 0.014990520663559437, + 0.00518900528550148, + 0.009333737194538116, + 0.004838896449655294, + 0.01141155045479536, + 0.013287733308970928, + 0.015401528216898441, + 0.030489087104797363, + 0.02254197932779789, + 0.0020223644096404314, + 0.00308639626018703, + 0.00211803219281137, + 0.003549405140802264, + 0.0043400926515460014, + 0.0047651734203100204, + 0.0030020426493138075, + 0.005508123897016048, + 0.022806232795119286, + 0.015966219827532768, + 0.02402309700846672, + 0.03054378740489483, + 0.01991286128759384, + 0.050588443875312805, + 0.016685726121068, + 0.02925531007349491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.020581839606165886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1379048228263855, + 0.12175650149583817, + 0.08018089085817337, + 0.2747320830821991, + 0.09656553715467453, + 0.09874509274959564, + 0.0840376615524292, + 0.050255972892045975, + 0.022584253922104836, + 0.03323712944984436 + ], + "credit_map": [ + 0.03232964873313904, + 0.011975416913628578, + 0.01574188843369484, + 0.013413837179541588, + 0.021331900730729103, + 0.014828111976385117, + 0.008040969260036945, + 0.020243050530552864, + 0.009505133144557476, + 0.010241110809147358, + 0.013061403296887875, + 0.028842303901910782, + 0.013964504934847355, + 0.022658756002783775, + 0.01081740204244852, + 0.012665895745158195, + 0.007114588748663664, + 0.006723575759679079, + 0.014242087490856647, + 0.007959738373756409, + 0.00798240676522255, + 0.012670494616031647, + 0.009672035463154316, + 0.013815972954034805, + 0.017515242099761963, + 0.02482924796640873, + 0.08944232016801834, + 0.039486683905124664, + 0.00983581691980362, + 0.01956871710717678, + 0.013680104166269302, + 0.06037397310137749, + 0.0061156549490988255, + 0.00973302498459816, + 0.009676894173026085, + 0.010172128677368164, + 0.01287958025932312, + 0.017515109851956367, + 0.013573737815022469, + 0.016899414360523224, + 0.0098725575953722, + 0.01346316933631897, + 0.010912448167800903, + 0.009679718874394894, + 0.01356988213956356, + 0.014432593248784542, + 0.013971825130283833, + 0.012842908501625061, + 0.006627554539591074, + 0.010164154693484306, + 0.006852597463876009, + 0.0054190331138670444, + 0.005985323805361986, + 0.021486954763531685, + 0.017718492075800896, + 0.009783555753529072, + 0.003141784807667136, + 0.00423626322299242, + 0.0035745094064623117, + 0.004890069831162691, + 0.009517300873994827, + 0.010699919424951077, + 0.006331741809844971, + 0.007864385843276978, + 0.0015037467237561941, + 0.0021071797236800194, + 0.0017763050273060799, + 0.0028736430685967207, + 0.003605097532272339, + 0.0045508877374231815, + 0.0021718039643019438, + 0.003995593171566725, + 0.002879387466236949, + 0.004038647748529911, + 0.0038546116556972265, + 0.007354660890996456, + 0.0032482012175023556, + 0.0034481575712561607, + 0.004301016218960285, + 0.004112447611987591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.007628121413290501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20151937007904053 + ], + "opposite_logprob_tokens": [ + -1.7015193700790405 + ], + "logprob_reward": -0.20151937007904053, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20151937007904053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13987821340560913, + 0.1647566854953766, + 0.09367739409208298, + 0.16236285865306854, + 0.04979359731078148, + 0.03994004428386688, + 0.04221928119659424, + 0.19451820850372314, + 0.0237102210521698, + 0.08914344757795334 + ], + "credit_map": [ + 0.018626399338245392, + 0.011202615685760975, + 0.02104618214070797, + 0.014374220743775368, + 0.02367474138736725, + 0.019644491374492645, + 0.012777164578437805, + 0.018532399088144302, + 0.011743401177227497, + 0.01336782518774271, + 0.023817064240574837, + 0.022348545491695404, + 0.036961156874895096, + 0.018740056082606316, + 0.016972852870821953, + 0.02080577425658703, + 0.00819967407733202, + 0.007579068653285503, + 0.025704817846417427, + 0.008586720563471317, + 0.006073116324841976, + 0.01117025502026081, + 0.007593607529997826, + 0.018770135939121246, + 0.01179010421037674, + 0.013198593631386757, + 0.04964440315961838, + 0.023659776896238327, + 0.004713601898401976, + 0.014242022298276424, + 0.008405536413192749, + 0.03670882061123848, + 0.0031911777332425117, + 0.003284268081188202, + 0.004767349921166897, + 0.004374203737825155, + 0.009961476549506187, + 0.008682211861014366, + 0.007752693258225918, + 0.007780213374644518, + 0.003732106415554881, + 0.004707265645265579, + 0.004211811814457178, + 0.003073766827583313, + 0.004716203082352877, + 0.007579258177429438, + 0.004685701336711645, + 0.007233929354697466, + 0.004509837832301855, + 0.008901453576982021, + 0.003198286285623908, + 0.0040397667326033115, + 0.004360049031674862, + 0.006139968987554312, + 0.0056783766485750675, + 0.005391543731093407, + 0.011243266984820366, + 0.01742587983608246, + 0.009836247190833092, + 0.020837275311350822, + 0.03306771442294121, + 0.03929028660058975, + 0.03690752387046814, + 0.025910012423992157, + 0.001359706511721015, + 0.0021700209472328424, + 0.001971087185665965, + 0.003876327769830823, + 0.0030579841695725918, + 0.0035066825803369284, + 0.002952407579869032, + 0.004816002212464809, + 0.008061857894062996, + 0.006520621478557587, + 0.011343397200107574, + 0.009151286445558071, + 0.00717106694355607, + 0.024301396682858467, + 0.008187294006347656, + 0.014406529255211353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.20151937007904053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene imbued with authentic CGI cinematic realism, a sweeping wide shot unfolds in slow, graceful motion around a massive lion's mane jellyfish suspended in the deep, abyssal azure of the ocean. The creature's massive, domed bell, emitting a vibrant, fiery orange-red hue, pulsates gently and rhythmically, its gelatinous form contracting steadily and deliberately, propelling itself forward with silent, effortless force. A vast, intricate network of countless tiny, thread-like tentacles, some as thick as ribbons, others as delicate as silk, floats behind it in a tangled, fluid mass emitting a soft, ethereal blue-green bioluminescence. In the distance, wisps of brilliant sunlight filter down into shimmering divine light, piercing the darkness and illuminating the jellyfish's translucent body, revealing its intricate internal structure and highlighting tiny grains of marine snow that drift lazily through the water like suspended dust. The camera’s fluid, arcing movements capture the majestic and serene scene from every angle, emphasizing the creature’s breathtaking size and the hypnotic beauty of its slow, graceful dance through the depths.", + "seed": 2795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_3/index_3699_seed_2795_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": -0.03390244338661432, + "vlm_logprob": -0.03390244338661432, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.03390244338661432, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011124256998300552 + ], + "opposite_logprob_tokens": [ + -4.761124134063721 + ], + "logprob_reward": -0.011124256998300552, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011124256998300552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2077740877866745, + 0.14341306686401367, + 0.06634686142206192, + 0.1324082612991333, + 0.07222240418195724, + 0.04558641090989113, + 0.04924985021352768, + 0.042354438453912735, + 0.15824556350708008, + 0.08239908516407013 + ], + "credit_map": [ + 0.015321456827223301, + 0.026412740349769592, + 0.021067772060632706, + 0.03553179278969765, + 0.02547856606543064, + 0.03405304253101349, + 0.022068338468670845, + 0.027840374037623405, + 0.010281207971274853, + 0.02703896164894104, + 0.016356082633137703, + 0.025397174060344696, + 0.013939998112618923, + 0.015593205578625202, + 0.01685950718820095, + 0.01794692873954773, + 0.005056012887507677, + 0.00599559023976326, + 0.007355836220085621, + 0.009556100703775883, + 0.00755376135930419, + 0.009776074439287186, + 0.01059312280267477, + 0.010460367426276207, + 0.012516391463577747, + 0.011500817723572254, + 0.008812565356492996, + 0.010231313295662403, + 0.016617383807897568, + 0.03412232547998428, + 0.017075393348932266, + 0.02153206802904606, + 0.005385270342230797, + 0.00804095808416605, + 0.013447863049805164, + 0.007599546108394861, + 0.009538541547954082, + 0.00933453906327486, + 0.009684314951300621, + 0.009191365912556648, + 0.0037425446789711714, + 0.0032067704014480114, + 0.0055063748732209206, + 0.0067346771247684956, + 0.005785772111266851, + 0.005039619747549295, + 0.008967923931777477, + 0.0066027264110744, + 0.0028608820866793394, + 0.006902283523231745, + 0.005578105803579092, + 0.009923046454787254, + 0.006253896281123161, + 0.005111199803650379, + 0.008007359690964222, + 0.004613079130649567, + 0.0033580532763153315, + 0.0046726614236831665, + 0.006956126075237989, + 0.004763339646160603, + 0.004714533220976591, + 0.006083618383854628, + 0.006007464602589607, + 0.00579863740131259, + 0.01211001817137003, + 0.01563614420592785, + 0.01884349249303341, + 0.031973108649253845, + 0.014760859310626984, + 0.015713931992650032, + 0.02650095336139202, + 0.022707050666213036, + 0.009052868001163006, + 0.009916639886796474, + 0.0063292705453932285, + 0.006884268019348383, + 0.0067067816853523254, + 0.013257274404168129, + 0.009201767854392529, + 0.021050220355391502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.011124256998300552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002833639271557331 + ], + "opposite_logprob_tokens": [ + -5.877833843231201 + ], + "logprob_reward": -0.002833639271557331, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002833639271557331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15985292196273804, + 0.21043461561203003, + 0.08691859990358353, + 0.06309502571821213, + 0.040705565363168716, + 0.09568163007497787, + 0.03507811203598976, + 0.11205270886421204, + 0.05730418488383293, + 0.1388765275478363 + ], + "credit_map": [ + 0.015283205546438694, + 0.01994910277426243, + 0.021936871111392975, + 0.020156268030405045, + 0.02137448452413082, + 0.024054350331425667, + 0.01881072297692299, + 0.01828792504966259, + 0.016825737431645393, + 0.03953985869884491, + 0.022180916741490364, + 0.033816128969192505, + 0.030237670987844467, + 0.018827637657523155, + 0.02526550181210041, + 0.023741193115711212, + 0.007068671751767397, + 0.012578682042658329, + 0.01070629432797432, + 0.014212125912308693, + 0.01036848220974207, + 0.010617968626320362, + 0.009518476203083992, + 0.011847913265228271, + 0.005310416221618652, + 0.004267929587513208, + 0.004623389802873135, + 0.007094505243003368, + 0.007163696922361851, + 0.016502883285284042, + 0.007673075422644615, + 0.010459131561219692, + 0.0034233820624649525, + 0.004306921269744635, + 0.007255021017044783, + 0.005019518546760082, + 0.006386253517121077, + 0.003932752180844545, + 0.005771331489086151, + 0.00461038900539279, + 0.006394031923264265, + 0.004789028782397509, + 0.009037771262228489, + 0.010070784948766232, + 0.00706993043422699, + 0.005738300271332264, + 0.029863949865102768, + 0.022717837244272232, + 0.0035690218210220337, + 0.004106795880943537, + 0.00421133590862155, + 0.005858005490154028, + 0.003291226690635085, + 0.004381003323942423, + 0.005416834261268377, + 0.004243893548846245, + 0.01083938218653202, + 0.015052683651447296, + 0.013272054493427277, + 0.015843026340007782, + 0.01298731379210949, + 0.013426223769783974, + 0.012487731873989105, + 0.018144311383366585, + 0.0055265421979129314, + 0.005996108520776033, + 0.007900728844106197, + 0.007244272157549858, + 0.006004157941788435, + 0.006212993990629911, + 0.0092341722920537, + 0.009185216389596462, + 0.015170779079198837, + 0.023748720064759254, + 0.016520174220204353, + 0.015454435721039772, + 0.013349448330700397, + 0.020077450200915337, + 0.015451704151928425, + 0.01910383813083172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.002833639271557331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286210373044014 + ], + "opposite_logprob_tokens": [ + -3.7732861042022705 + ], + "logprob_reward": -0.023286210373044014, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286210373044014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09957488626241684, + 0.049937620759010315, + 0.10126572102308273, + 0.09011179208755493, + 0.05200926214456558, + 0.23128677904605865, + 0.07391555607318878, + 0.06468984484672546, + 0.06871887296438217, + 0.16848959028720856 + ], + "credit_map": [ + 0.01020725816488266, + 0.007600019685924053, + 0.016493558883666992, + 0.013933907262980938, + 0.011387401260435581, + 0.01561076007783413, + 0.008362342603504658, + 0.015979645773768425, + 0.006059963256120682, + 0.005591683089733124, + 0.005420780275017023, + 0.009875370189547539, + 0.004978394601494074, + 0.0060177757404744625, + 0.005800994578748941, + 0.006192662753164768, + 0.01194852590560913, + 0.010866908356547356, + 0.011079687625169754, + 0.0122674023732543, + 0.011507517658174038, + 0.016008028760552406, + 0.013137771748006344, + 0.014449880458414555, + 0.009889351204037666, + 0.006196276284754276, + 0.007579461205750704, + 0.008813345804810524, + 0.01261072140187025, + 0.01980232074856758, + 0.008254477754235268, + 0.01696583442389965, + 0.004043142776936293, + 0.007084370590746403, + 0.00760670518502593, + 0.005522450432181358, + 0.008753330446779728, + 0.0048370747826993465, + 0.008520186878740788, + 0.005641999654471874, + 0.015471534803509712, + 0.013554354198276997, + 0.019119415432214737, + 0.02687562257051468, + 0.017796756699681282, + 0.011857532896101475, + 0.07039687037467957, + 0.05621467903256416, + 0.006489339284598827, + 0.006013107020407915, + 0.010443790815770626, + 0.01305340975522995, + 0.0062513104639947414, + 0.007210673298686743, + 0.011392924003303051, + 0.01306100096553564, + 0.00730134965851903, + 0.010272542014718056, + 0.006905386224389076, + 0.007770467549562454, + 0.0077980924397706985, + 0.007292042952030897, + 0.008204028941690922, + 0.009145933203399181, + 0.0075263213366270065, + 0.009681959636509418, + 0.007180194370448589, + 0.011126305907964706, + 0.007028784602880478, + 0.00869132112711668, + 0.009756729006767273, + 0.007727257441729307, + 0.012546014040708542, + 0.0108836954459548, + 0.01181169506162405, + 0.014820259064435959, + 0.00899620819836855, + 0.043499454855918884, + 0.020017629489302635, + 0.04591464623808861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.023286210373044014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249762907624245 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249762907624245, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249762907624245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22324348986148834, + 0.0706271380186081, + 0.11122509837150574, + 0.10148885846138, + 0.10894041508436203, + 0.06488564610481262, + 0.05475931242108345, + 0.06753622740507126, + 0.03867538273334503, + 0.15861840546131134 + ], + "credit_map": [ + 0.027364781126379967, + 0.016047408804297447, + 0.04184644669294357, + 0.03199790418148041, + 0.01512488629668951, + 0.03351631015539169, + 0.021857207641005516, + 0.0354885570704937, + 0.007431800942867994, + 0.011151710525155067, + 0.007582155987620354, + 0.008992129936814308, + 0.01043874491006136, + 0.006446755491197109, + 0.00898822769522667, + 0.00959562137722969, + 0.007848449051380157, + 0.008596873842179775, + 0.0170304998755455, + 0.016231819987297058, + 0.012966153211891651, + 0.013778716325759888, + 0.018564390018582344, + 0.016208196058869362, + 0.011580453254282475, + 0.010915894992649555, + 0.012079406529664993, + 0.01143546961247921, + 0.010592183098196983, + 0.017805619165301323, + 0.013655411079525948, + 0.013424419797956944, + 0.009852767921984196, + 0.008729878813028336, + 0.025713056325912476, + 0.009570233523845673, + 0.01332539040595293, + 0.01181382592767477, + 0.013782060705125332, + 0.01615319959819317, + 0.004737683571875095, + 0.006919715087860823, + 0.009798716753721237, + 0.007357562426477671, + 0.007007793057709932, + 0.005355828441679478, + 0.012297634035348892, + 0.01141071692109108, + 0.0037233897019177675, + 0.00612874049693346, + 0.005415006075054407, + 0.009590426459908485, + 0.0053321742452681065, + 0.005292151123285294, + 0.011392692103981972, + 0.0078847361728549, + 0.006645402871072292, + 0.008554995991289616, + 0.009891406632959843, + 0.007918084040284157, + 0.00828539952635765, + 0.009263377636671066, + 0.00767508102580905, + 0.009302481077611446, + 0.0041321562603116035, + 0.004963709507137537, + 0.0049242787063121796, + 0.005984509829431772, + 0.00304823974147439, + 0.0045127286575734615, + 0.0050625307485461235, + 0.006047227885574102, + 0.011008372530341148, + 0.013658633455634117, + 0.010622656904160976, + 0.011795303784310818, + 0.010255350731313229, + 0.036101460456848145, + 0.01886999048292637, + 0.046306632459163666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.005249762907624245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 lion's mane jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the jellyfish's bell a vibrant fiery orange-red color?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny thread-like tentacles float behind the jellyfish emitting a soft blue-green bioluminescence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform fluid arcing movements around the jellyfish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are wisps of brilliant sunlight visible filtering down into the deep ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12701834738254547 + ], + "opposite_logprob_tokens": [ + -2.127018451690674 + ], + "logprob_reward": -0.12701834738254547, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12701834738254547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19095370173454285, + 0.04574863612651825, + 0.02621074952185154, + 0.044287193566560745, + 0.02352161705493927, + 0.17912167310714722, + 0.0446489118039608, + 0.08852298557758331, + 0.05143830180168152, + 0.30554622411727905 + ], + "credit_map": [ + 0.03092135302722454, + 0.013512502424418926, + 0.0404212512075901, + 0.02150714583694935, + 0.015251806005835533, + 0.024948248639702797, + 0.016828671097755432, + 0.027562730014324188, + 0.008095783181488514, + 0.004113032948225737, + 0.008349454030394554, + 0.005731714423745871, + 0.0032489204313606024, + 0.005615010857582092, + 0.004291525110602379, + 0.0063031939789652824, + 0.0028813041280955076, + 0.003753795986995101, + 0.0023563336580991745, + 0.0030657195020467043, + 0.002591649768874049, + 0.0048055509105324745, + 0.002497161505743861, + 0.004259234759956598, + 0.005184330977499485, + 0.005364733282476664, + 0.0038297101855278015, + 0.004232081584632397, + 0.008237448520958424, + 0.005719667766243219, + 0.0063070510514080524, + 0.00541217066347599, + 0.0019030917901545763, + 0.002657820237800479, + 0.003639461938291788, + 0.00243545975536108, + 0.0031121131032705307, + 0.002118409611284733, + 0.00544791528955102, + 0.0022073439322412014, + 0.011123188771307468, + 0.006163198500871658, + 0.014640365727245808, + 0.018497226759791374, + 0.011587971821427345, + 0.009350857697427273, + 0.058499448001384735, + 0.04925941303372383, + 0.004087450448423624, + 0.0033004549331963062, + 0.007018861826509237, + 0.007189522963017225, + 0.0036749516148120165, + 0.003704632166773081, + 0.007223478984087706, + 0.008449560031294823, + 0.009561446495354176, + 0.014781907200813293, + 0.009114349260926247, + 0.011015311814844608, + 0.010052474215626717, + 0.012412742711603642, + 0.009310825727880001, + 0.012273937463760376, + 0.004176982678472996, + 0.005074568558484316, + 0.008421717211604118, + 0.00665607675909996, + 0.004142214078456163, + 0.0061383056454360485, + 0.007625816389918327, + 0.009202615357935429, + 0.02086763083934784, + 0.022620806470513344, + 0.020232409238815308, + 0.02216673456132412, + 0.018234029412269592, + 0.06971269100904465, + 0.037427693605422974, + 0.09428421407938004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.12701834738254547 + } + ], + "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/0000155/group_4/index_tensor([2986, 2986])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_4/index_tensor([2986, 2986])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ef8875740940cc38d8795d4d535c4265fc56fa0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_4/index_tensor([2986, 2986])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, almost macroscopic, extreme close-up focuses on a human fingertip, the intricate spirals of the fingerprint rendered in sharp detail. A jagged, dark brown splinter is clearly embedded beneath the skin's surface, causing a slight red inflammation around the point of entry. A pair of sharp-tipped, medical-grade steel tweezers moves into the frame, capturing the cold, metallic sheen from a single, harsh overhead light source. The tips of the tweezers meticulously and decisively grasp the exposed end of the splinter. With steady, deliberate pull, the tweezers slowly and firmly extract the splinter from the flesh, causing the skin to slightly bulge around the object before it is cleanly lifted. The entire scene is rendered in a stark, desaturated palette, dominated by the cool gray of the steel, the pale white tones of the skin, and the near-black of the splinter, creating an atmosphere of tension, focus, and clinical intensity that emphasizes the delicate and slightly uncomfortable nature of the procedure.", + "seed": 20025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_4/index_2986_seed_20025_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": -0.20792442616075277, + "vlm_logprob": -0.20792442616075277, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.20792442616075277, + "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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01471111923456192 + ], + "opposite_logprob_tokens": [ + -4.514710903167725 + ], + "logprob_reward": -0.01471111923456192, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01471111923456192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0956740751862526, + 0.16077208518981934, + 0.10712867230176926, + 0.018147723749279976, + 0.02148434706032276, + 0.12079248577356339, + 0.15710653364658356, + 0.11511648446321487, + 0.07823949307203293, + 0.12553808093070984 + ], + "credit_map": [ + 0.010147759690880775, + 0.01255608256906271, + 0.0286132600158453, + 0.012684610672295094, + 0.008898314088582993, + 0.005833766423165798, + 0.01052646990865469, + 0.006413827184587717, + 0.0300271175801754, + 0.04707511141896248, + 0.03805246949195862, + 0.011080054566264153, + 0.00535462936386466, + 0.005111705046147108, + 0.010467159561812878, + 0.013603860512375832, + 0.028283990919589996, + 0.016796378418803215, + 0.013500773347914219, + 0.012589539401233196, + 0.005191118456423283, + 0.006065670400857925, + 0.006449237931519747, + 0.018251972272992134, + 0.002033801283687353, + 0.00402215076610446, + 0.003821735503152013, + 0.0032055978663265705, + 0.0006551158730871975, + 0.001311801141127944, + 0.0015446352772414684, + 0.0015528870280832052, + 0.0028930206317454576, + 0.005035511218011379, + 0.004014622885733843, + 0.0021359389647841454, + 0.0009846251923590899, + 0.0016900973860174417, + 0.0018162167398259044, + 0.0029143167193979025, + 0.020306359976530075, + 0.027814116328954697, + 0.025240395218133926, + 0.014646533876657486, + 0.006605543661862612, + 0.00766023201867938, + 0.008614983409643173, + 0.009904323145747185, + 0.019642211496829987, + 0.023132825270295143, + 0.05116554722189903, + 0.023909220471978188, + 0.007954337634146214, + 0.007668226957321167, + 0.011558824218809605, + 0.012075357139110565, + 0.009964277036488056, + 0.026434704661369324, + 0.04199126362800598, + 0.008452975191175938, + 0.0033411721233278513, + 0.007213972508907318, + 0.004919297527521849, + 0.012798827141523361, + 0.01300519797950983, + 0.010879830457270145, + 0.01350462157279253, + 0.011854883283376694, + 0.005122730508446693, + 0.007446903735399246, + 0.009675169363617897, + 0.006750165484845638, + 0.017326606437563896, + 0.011713135987520218, + 0.04359925538301468, + 0.01382493320852518, + 0.009775093756616116, + 0.009787014685571194, + 0.010661209933459759, + 0.008850844576954842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 9 + ], + "selected_reward": -0.01471111923456192 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018220432102680206 + ], + "opposite_logprob_tokens": [ + -4.0182204246521 + ], + "logprob_reward": -0.018220432102680206, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018220432102680206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15269318222999573, + 0.1981067806482315, + 0.1041247695684433, + 0.032978519797325134, + 0.022045815363526344, + 0.07125494629144669, + 0.16011807322502136, + 0.0357217900454998, + 0.09006811678409576, + 0.13288795948028564 + ], + "credit_map": [ + 0.018501505255699158, + 0.020171338692307472, + 0.04426959902048111, + 0.014812790788710117, + 0.019561590626835823, + 0.010065008886158466, + 0.014682532288134098, + 0.010628818534314632, + 0.038998574018478394, + 0.07140769809484482, + 0.043069470673799515, + 0.010528012178838253, + 0.005259006749838591, + 0.005164775066077709, + 0.005821676924824715, + 0.01785758137702942, + 0.023381279781460762, + 0.009852752089500427, + 0.01624392904341221, + 0.011350119486451149, + 0.004823812749236822, + 0.006506285630166531, + 0.008496230468153954, + 0.02347036451101303, + 0.005064795725047588, + 0.010137232020497322, + 0.006578613072633743, + 0.0032260355073958635, + 0.001143451314419508, + 0.0014929377939552069, + 0.002030337695032358, + 0.0033051175996661186, + 0.0030037728138267994, + 0.0035418353509157896, + 0.005660180933773518, + 0.0029095010831952095, + 0.0009499009465798736, + 0.002198693109676242, + 0.0014601742150261998, + 0.002321759006008506, + 0.011723704636096954, + 0.009467984549701214, + 0.014709362760186195, + 0.006425174884498119, + 0.0028340185526758432, + 0.007273274473845959, + 0.006017183884978294, + 0.012804246507585049, + 0.02095307782292366, + 0.02724745310842991, + 0.028797732666134834, + 0.033391378819942474, + 0.009689783677458763, + 0.01078022364526987, + 0.016199426725506783, + 0.013058992102742195, + 0.00482213357463479, + 0.0066504329442977905, + 0.007862628437578678, + 0.004963016603142023, + 0.0011828317074105144, + 0.002814669394865632, + 0.0034149165730923414, + 0.004011161625385284, + 0.011952393688261509, + 0.01126808114349842, + 0.014922291040420532, + 0.013092413544654846, + 0.005644293036311865, + 0.009693526662886143, + 0.0159724373370409, + 0.00752269197255373, + 0.020432282239198685, + 0.015948154032230377, + 0.04123036935925484, + 0.01711534522473812, + 0.010743360035121441, + 0.009812400676310062, + 0.008093921467661858, + 0.009512127377092838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.018220432102680206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962916649878025 + ], + "opposite_logprob_tokens": [ + -5.13096284866333 + ], + "logprob_reward": -0.005962916649878025, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962916649878025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07168721407651901, + 0.11266960948705673, + 0.05823928490281105, + 0.0188086349517107, + 0.020322276279330254, + 0.10990168154239655, + 0.01595417596399784, + 0.039175111800432205, + 0.06104818359017372, + 0.4921938180923462 + ], + "credit_map": [ + 0.008932674303650856, + 0.00957543682307005, + 0.023062394931912422, + 0.008065672591328621, + 0.0072449506260454655, + 0.004226604010909796, + 0.005823804996907711, + 0.004755676724016666, + 0.015511685982346535, + 0.03428042680025101, + 0.028366127982735634, + 0.007257549557834864, + 0.0037860118318349123, + 0.005091424565762281, + 0.008002839982509613, + 0.01037354301661253, + 0.005547475069761276, + 0.009095877408981323, + 0.013202717527747154, + 0.009317455813288689, + 0.0034952929709106684, + 0.004481217823922634, + 0.0058537935838103294, + 0.0072454530745744705, + 0.0026569408364593983, + 0.004123337101191282, + 0.004330051131546497, + 0.0024150616955012083, + 0.0007361469324678183, + 0.0014525403967127204, + 0.0013267523609101772, + 0.0017678035655990243, + 0.002392450114712119, + 0.0028095864690840244, + 0.0033142264001071453, + 0.002500381087884307, + 0.001433291588909924, + 0.002280178712680936, + 0.0032183134462684393, + 0.0023738464806228876, + 0.01715591549873352, + 0.01403862051665783, + 0.018056312575936317, + 0.014502355828881264, + 0.007079060655087233, + 0.012881827540695667, + 0.00978544820100069, + 0.016402138397097588, + 0.001837954856455326, + 0.002672460861504078, + 0.0047902618534862995, + 0.0022096026223152876, + 0.0008612187812104821, + 0.0010792118264362216, + 0.0012479794677346945, + 0.0012554851127788424, + 0.004589910618960857, + 0.007804760709404945, + 0.009791670367121696, + 0.0043280101381242275, + 0.0020186984911561012, + 0.0030561434105038643, + 0.0037583899684250355, + 0.003827525768429041, + 0.008059185929596424, + 0.007161570247262716, + 0.009110154584050179, + 0.008973577059805393, + 0.003966083750128746, + 0.007270371541380882, + 0.011092850007116795, + 0.005414395127445459, + 0.07955175638198853, + 0.05421437695622444, + 0.20276610553264618, + 0.03688998892903328, + 0.030103683471679688, + 0.027749761939048767, + 0.02309059351682663, + 0.03782755136489868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.005962916649878025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636326290667057 + ], + "opposite_logprob_tokens": [ + -3.6513633728027344 + ], + "logprob_reward": -0.02636326290667057, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636326290667057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06256330758333206, + 0.21971072256565094, + 0.06078691780567169, + 0.014743770472705364, + 0.023607950657606125, + 0.06888600438833237, + 0.0511067733168602, + 0.172988623380661, + 0.04315561428666115, + 0.28245028853416443 + ], + "credit_map": [ + 0.006708267144858837, + 0.008987422101199627, + 0.020626801997423172, + 0.006380944047123194, + 0.006682884879410267, + 0.0046814121305942535, + 0.004727847408503294, + 0.0037677225191146135, + 0.03288555517792702, + 0.0713309571146965, + 0.03958909213542938, + 0.012751666828989983, + 0.00829294789582491, + 0.009499222040176392, + 0.019864199683070183, + 0.025497091934084892, + 0.014400692656636238, + 0.006324118468910456, + 0.010042430832982063, + 0.0063134897500276566, + 0.0026523771230131388, + 0.00357501907274127, + 0.004349729046225548, + 0.013129063881933689, + 0.0027112269308418036, + 0.0028551495634019375, + 0.0029710878152400255, + 0.002337498590350151, + 0.0006316909566521645, + 0.0010789850493893027, + 0.0008239811868406832, + 0.0013341503217816353, + 0.0033434582874178886, + 0.007421047426760197, + 0.0046962834894657135, + 0.001967739313840866, + 0.0007970983861014247, + 0.001279689953662455, + 0.0016226480947807431, + 0.0024799855891615152, + 0.013787682168185711, + 0.01156364195048809, + 0.010182849131524563, + 0.008755574002861977, + 0.0033343411050736904, + 0.005890452768653631, + 0.004613871220499277, + 0.010757588781416416, + 0.005735821556299925, + 0.007039146032184362, + 0.01568474993109703, + 0.00707976333796978, + 0.0031026769429445267, + 0.0038713987451046705, + 0.004756416659802198, + 0.0038367973174899817, + 0.015122231096029282, + 0.038836367428302765, + 0.06094393879175186, + 0.012204153463244438, + 0.00577698415145278, + 0.011056272312998772, + 0.009217340499162674, + 0.019831322133541107, + 0.0059822918847203255, + 0.008542082272469997, + 0.004072696901857853, + 0.004587819799780846, + 0.003013614099472761, + 0.0044412678107619286, + 0.008406526409089565, + 0.004109315574169159, + 0.04417334496974945, + 0.031196298077702522, + 0.10795845091342926, + 0.023232903331518173, + 0.023584425449371338, + 0.016605379059910774, + 0.01474094670265913, + 0.020958533510565758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.02636326290667057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9743643999099731 + ], + "opposite_logprob_tokens": [ + -0.47436439990997314 + ], + "logprob_reward": -0.9743643999099731, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743643999099731, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17249011993408203, + 0.10331176966428757, + 0.039480675011873245, + 0.02856086753308773, + 0.036234114319086075, + 0.10257130861282349, + 0.03324753791093826, + 0.09291937202215195, + 0.23776404559612274, + 0.15342025458812714 + ], + "credit_map": [ + 0.023373790085315704, + 0.025355108082294464, + 0.031703874468803406, + 0.017339102923870087, + 0.023512404412031174, + 0.01767316646873951, + 0.020292585715651512, + 0.013240098021924496, + 0.019710544496774673, + 0.03538782522082329, + 0.018713215366005898, + 0.0055091241374611855, + 0.003922800533473492, + 0.0046593546867370605, + 0.006257819477468729, + 0.009151090867817402, + 0.0072255246341228485, + 0.005900689400732517, + 0.007707993034273386, + 0.005328605882823467, + 0.0023311118129640818, + 0.0023118997924029827, + 0.0032809018157422543, + 0.0053939479403197765, + 0.003825077787041664, + 0.004925479646772146, + 0.005247293505817652, + 0.0034116797614842653, + 0.0016747426707297564, + 0.002738114446401596, + 0.003909184131771326, + 0.002829295815899968, + 0.005283116828650236, + 0.008746746927499771, + 0.006652007345110178, + 0.004183600656688213, + 0.0013414140557870269, + 0.002501028822734952, + 0.003401809837669134, + 0.004124388564378023, + 0.015186808072030544, + 0.023992612957954407, + 0.023547664284706116, + 0.01256762444972992, + 0.004826967604458332, + 0.005990775302052498, + 0.008124987594783306, + 0.008333864621818066, + 0.003380910027772188, + 0.004041774198412895, + 0.007488546893000603, + 0.0060319313779473305, + 0.002178456634283066, + 0.0018238569609820843, + 0.005254844203591347, + 0.0030472162179648876, + 0.014523312449455261, + 0.019172221422195435, + 0.014842107892036438, + 0.014301888644695282, + 0.0040082773193717, + 0.007668780628591776, + 0.006892325822263956, + 0.011510457843542099, + 0.04247673600912094, + 0.029056930914521217, + 0.03868534043431282, + 0.034288644790649414, + 0.015284497290849686, + 0.025013886392116547, + 0.03125755861401558, + 0.021700454875826836, + 0.022002538666129112, + 0.01616567373275757, + 0.05914369598031044, + 0.015507728792726994, + 0.00876797828823328, + 0.010519388131797314, + 0.010007527656853199, + 0.011305735446512699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.9743643999099731 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, almost macroscopic, extreme close-up focuses on a human fingertip, the intricate spirals of the fingerprint rendered in sharp detail. A jagged, dark brown splinter is clearly embedded beneath the skin's surface, causing a slight red inflammation around the point of entry. A pair of sharp-tipped, medical-grade steel tweezers moves into the frame, capturing the cold, metallic sheen from a single, harsh overhead light source. The tips of the tweezers meticulously and decisively grasp the exposed end of the splinter. With steady, deliberate pull, the tweezers slowly and firmly extract the splinter from the flesh, causing the skin to slightly bulge around the object before it is cleanly lifted. The entire scene is rendered in a stark, desaturated palette, dominated by the cool gray of the steel, the pale white tones of the skin, and the near-black of the splinter, creating an atmosphere of tension, focus, and clinical intensity that emphasizes the delicate and slightly uncomfortable nature of the procedure.", + "seed": 20026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_4/index_2986_seed_20026_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": -0.12790580512955785, + "vlm_logprob": -0.12790580512955785, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.12790580512955785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01778184063732624 + ], + "opposite_logprob_tokens": [ + -5.267781734466553 + ], + "logprob_reward": -0.01778184063732624, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01778184063732624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13437315821647644, + 0.061367154121398926, + 0.08123070001602173, + 0.05715702101588249, + 0.0765102282166481, + 0.2836950123310089, + 0.1751946210861206, + 0.039795465767383575, + 0.028653141111135483, + 0.062023553997278214 + ], + "credit_map": [ + 0.012161096557974815, + 0.020871222019195557, + 0.03271695598959923, + 0.015361192636191845, + 0.007867934182286263, + 0.015899907797574997, + 0.00779853155836463, + 0.0216963030397892, + 0.005837394390255213, + 0.007652021013200283, + 0.009900682605803013, + 0.01260504312813282, + 0.005401670001447201, + 0.006423804443329573, + 0.0068468148820102215, + 0.006699725054204464, + 0.007852347567677498, + 0.00997211504727602, + 0.012925868853926659, + 0.01473511103540659, + 0.004277047701179981, + 0.004037779755890369, + 0.012624708004295826, + 0.014805709943175316, + 0.008604024536907673, + 0.007664447650313377, + 0.007815985940396786, + 0.013919992372393608, + 0.003988962154835463, + 0.0021058237180113792, + 0.004620525985956192, + 0.008437251672148705, + 0.009337672963738441, + 0.008912199176847935, + 0.015301181934773922, + 0.013933301903307438, + 0.006288264412432909, + 0.005311695393174887, + 0.005766591522842646, + 0.011659309267997742, + 0.035534005612134933, + 0.044583510607481, + 0.08482072502374649, + 0.05208912864327431, + 0.022128771990537643, + 0.012261809781193733, + 0.01496833935379982, + 0.017308687791228294, + 0.02139579877257347, + 0.02626573108136654, + 0.041204292327165604, + 0.02159939706325531, + 0.023366030305624008, + 0.008080758154392242, + 0.020310133695602417, + 0.012972461991012096, + 0.0033469025511294603, + 0.004280790686607361, + 0.005002221092581749, + 0.00953351054340601, + 0.004214257933199406, + 0.0020667798817157745, + 0.004546976648271084, + 0.0068040224723517895, + 0.0030280048958957195, + 0.004634842276573181, + 0.0039038530085235834, + 0.004513454623520374, + 0.002626645378768444, + 0.003933174069970846, + 0.003039730479940772, + 0.0029734347481280565, + 0.008522083051502705, + 0.014679551124572754, + 0.008378736674785614, + 0.007049762178212404, + 0.004826407413929701, + 0.0058253114111721516, + 0.00825483724474907, + 0.004486857447773218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.01778184063732624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004112834110856056 + ], + "opposite_logprob_tokens": [ + -5.504112720489502 + ], + "logprob_reward": -0.004112834110856056, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004112834110856056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2945694923400879, + 0.15488937497138977, + 0.13338914513587952, + 0.0441586896777153, + 0.054328739643096924, + 0.0905960351228714, + 0.07086418569087982, + 0.04602532833814621, + 0.041398730129003525, + 0.06978025287389755 + ], + "credit_map": [ + 0.025908607989549637, + 0.03167346492409706, + 0.030455097556114197, + 0.06146474555134773, + 0.04205404967069626, + 0.03971496969461441, + 0.029026851058006287, + 0.03427169471979141, + 0.01565377414226532, + 0.013742264360189438, + 0.02416316233575344, + 0.02719770558178425, + 0.01635073684155941, + 0.019223392009735107, + 0.01479312777519226, + 0.02376522123813629, + 0.010632033459842205, + 0.015829116106033325, + 0.019169608131051064, + 0.021327676251530647, + 0.0068524968810379505, + 0.008365522138774395, + 0.01981450617313385, + 0.03139819949865341, + 0.005413163919001818, + 0.007090399041771889, + 0.00710288854315877, + 0.00849928893148899, + 0.0031694520730525255, + 0.002661667298525572, + 0.005884661339223385, + 0.0043371678330004215, + 0.003520010272040963, + 0.013735202141106129, + 0.016583885997533798, + 0.007813964039087296, + 0.0026343066710978746, + 0.002119976095855236, + 0.003998142201453447, + 0.003923257812857628, + 0.010400963947176933, + 0.017418522387742996, + 0.02910294383764267, + 0.015147138386964798, + 0.0066426401026546955, + 0.0032748656813055277, + 0.004328036215156317, + 0.004280927125364542, + 0.007544385734945536, + 0.010711563751101494, + 0.016906460747122765, + 0.010021691210567951, + 0.009769316762685776, + 0.002277079503983259, + 0.006906271446496248, + 0.006727417465299368, + 0.005173149984329939, + 0.006084111053496599, + 0.006024159025400877, + 0.00826009176671505, + 0.0041082859970629215, + 0.0023997067473828793, + 0.0055848085321486, + 0.008391015231609344, + 0.0032598022371530533, + 0.004839387256652117, + 0.008745773695409298, + 0.007482948247343302, + 0.0032478589564561844, + 0.004245304968208075, + 0.004573090933263302, + 0.005004563834518194, + 0.007276101503521204, + 0.02050546370446682, + 0.00958178285509348, + 0.006396805867552757, + 0.005746934562921524, + 0.005845252424478531, + 0.007679496891796589, + 0.006748417858034372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004112834110856056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639596704393625 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639596704393625, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639596704393625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2172337770462036, + 0.07416204363107681, + 0.27427589893341064, + 0.08349958807229996, + 0.043194349855184555, + 0.07278667390346527, + 0.04403108358383179, + 0.058691587299108505, + 0.08531563729047775, + 0.04680931568145752 + ], + "credit_map": [ + 0.016061071306467056, + 0.023905977606773376, + 0.034939367324113846, + 0.03281133249402046, + 0.022112935781478882, + 0.029079394415020943, + 0.01661372184753418, + 0.041710015386343, + 0.005935379303991795, + 0.007322228047996759, + 0.01087396964430809, + 0.010947396978735924, + 0.007265864871442318, + 0.008305905386805534, + 0.008626455441117287, + 0.014884845353662968, + 0.022986898198723793, + 0.030976571142673492, + 0.033880043774843216, + 0.05856978893280029, + 0.018876466900110245, + 0.01562916301190853, + 0.03905945643782616, + 0.05429752916097641, + 0.011649306863546371, + 0.011061922647058964, + 0.012114428915083408, + 0.01603168435394764, + 0.006114856339991093, + 0.004690778441727161, + 0.009729267098009586, + 0.012107361108064651, + 0.00379724008962512, + 0.006989162415266037, + 0.009493223391473293, + 0.005595471244305372, + 0.0047397855669260025, + 0.002877254504710436, + 0.004773694556206465, + 0.004928521811962128, + 0.010432546027004719, + 0.007543148472905159, + 0.013034305535256863, + 0.014627967029809952, + 0.007830242626369, + 0.004627730697393417, + 0.006143324542790651, + 0.008547414094209671, + 0.004514271859079599, + 0.005715792533010244, + 0.007088319398462772, + 0.009160132147371769, + 0.005974121857434511, + 0.0025446717627346516, + 0.004105520434677601, + 0.00492825685068965, + 0.0075793154537677765, + 0.0077447229996323586, + 0.0073413485661149025, + 0.010375038720667362, + 0.006766664795577526, + 0.003595269750803709, + 0.007520704064518213, + 0.007768535986542702, + 0.00940756220370531, + 0.008510557934641838, + 0.027368934825062752, + 0.010883226059377193, + 0.004586252849549055, + 0.005129410419613123, + 0.007386318873614073, + 0.012043379247188568, + 0.004935271106660366, + 0.005284144077450037, + 0.006077049300074577, + 0.007870951667428017, + 0.004930899012833834, + 0.004979284945875406, + 0.008709806017577648, + 0.004021916538476944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.007639596704393625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709511160850525 + ], + "opposite_logprob_tokens": [ + -3.4087095260620117 + ], + "logprob_reward": -0.033709511160850525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709511160850525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23779849708080292, + 0.04705379158258438, + 0.07497900724411011, + 0.05645342171192169, + 0.048707641661167145, + 0.18351420760154724, + 0.10640063881874084, + 0.010868764482438564, + 0.1290133148431778, + 0.10521073639392853 + ], + "credit_map": [ + 0.01777808740735054, + 0.0279628224670887, + 0.04581247270107269, + 0.04920759052038193, + 0.019036786630749702, + 0.03239918127655983, + 0.017238959670066833, + 0.02836260199546814, + 0.003493716474622488, + 0.006046278402209282, + 0.006148910615593195, + 0.0071375928819179535, + 0.005897166673094034, + 0.005646928213536739, + 0.006803709547966719, + 0.005879491101950407, + 0.005830991081893444, + 0.008667011745274067, + 0.010394607670605183, + 0.01167643629014492, + 0.004985584877431393, + 0.0052565112709999084, + 0.012638718821108341, + 0.015529141761362553, + 0.009318063966929913, + 0.008754497393965721, + 0.008421353995800018, + 0.011668818071484566, + 0.0038567509036511183, + 0.0027316221967339516, + 0.005912894383072853, + 0.0057894205674529076, + 0.005082832649350166, + 0.00887774396687746, + 0.008840572088956833, + 0.009360743686556816, + 0.003944887779653072, + 0.002930524991825223, + 0.0038193464279174805, + 0.005850993562489748, + 0.0230892114341259, + 0.02876395732164383, + 0.05418179929256439, + 0.03471291810274124, + 0.014737502671778202, + 0.007648160215467215, + 0.009592877700924873, + 0.010787777602672577, + 0.012422447092831135, + 0.016093900427222252, + 0.02565385401248932, + 0.014507378451526165, + 0.014127334579825401, + 0.0036709709092974663, + 0.011374036781489849, + 0.008550713770091534, + 0.0008240562165156007, + 0.0016482220962643623, + 0.001980075379833579, + 0.001622888259589672, + 0.0011014947667717934, + 0.0007331279339268804, + 0.0015384674770757556, + 0.0014204323524609208, + 0.014222314581274986, + 0.016757899895310402, + 0.023501738905906677, + 0.021081523969769478, + 0.010232624597847462, + 0.015835095196962357, + 0.014136475510895252, + 0.013245631009340286, + 0.012598656117916107, + 0.03656785562634468, + 0.015870144590735435, + 0.008698515594005585, + 0.006907557137310505, + 0.00652959942817688, + 0.008698009885847569, + 0.009340389631688595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.033709511160850525 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762852430343628 + ], + "opposite_logprob_tokens": [ + -0.8262852430343628 + ], + "logprob_reward": -0.5762852430343628, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762852430343628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4705973267555237, + 0.04576614126563072, + 0.07040305435657501, + 0.09879481047391891, + 0.04062233120203018, + 0.10715891420841217, + 0.025637097656726837, + 0.024944940581917763, + 0.04945651441812515, + 0.06661887466907501 + ], + "credit_map": [ + 0.03651535511016846, + 0.056222103536129, + 0.05460499972105026, + 0.11618220806121826, + 0.055287811905145645, + 0.055232420563697815, + 0.04960307478904724, + 0.046949367970228195, + 0.004159312695264816, + 0.006710775196552277, + 0.006327613722532988, + 0.007957492023706436, + 0.004584075417369604, + 0.0038157443050295115, + 0.004813661333173513, + 0.0073974705301225185, + 0.005766266956925392, + 0.009247122332453728, + 0.010008194483816624, + 0.016292141750454903, + 0.003893373068422079, + 0.003550862893462181, + 0.00790226086974144, + 0.013742835260927677, + 0.021467097103595734, + 0.014170347712934017, + 0.0152203980833292, + 0.02206565998494625, + 0.005345745477825403, + 0.003794085932895541, + 0.007011552341282368, + 0.00971993152052164, + 0.003712017321959138, + 0.007829608395695686, + 0.010165572166442871, + 0.007338418625295162, + 0.0025557216722518206, + 0.0017766528762876987, + 0.002751920372247696, + 0.0044924234971404076, + 0.009705785661935806, + 0.02676224336028099, + 0.034943681210279465, + 0.01636878214776516, + 0.006848790217190981, + 0.0029950321186333895, + 0.0033899263944476843, + 0.00614467728883028, + 0.0030472881626337767, + 0.0037147551774978638, + 0.005356892477720976, + 0.003915841691195965, + 0.0034007790964096785, + 0.0009890587534755468, + 0.002534176455810666, + 0.002678307704627514, + 0.003004714846611023, + 0.003552443580701947, + 0.00362189463339746, + 0.005452768411487341, + 0.0021647552493959665, + 0.0014481652760878205, + 0.0028773746453225613, + 0.0028228273149579763, + 0.004744755569845438, + 0.006545303389430046, + 0.009197335690259933, + 0.007175724487751722, + 0.0034032315015792847, + 0.004925272427499294, + 0.00687022041529417, + 0.006594672799110413, + 0.009085095487535, + 0.007984747178852558, + 0.009387449361383915, + 0.010547495447099209, + 0.007828304544091225, + 0.00765468692407012, + 0.006669582333415747, + 0.0074615185149014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.5762852430343628 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, almost macroscopic, extreme close-up focuses on a human fingertip, the intricate spirals of the fingerprint rendered in sharp detail. A jagged, dark brown splinter is clearly embedded beneath the skin's surface, causing a slight red inflammation around the point of entry. A pair of sharp-tipped, medical-grade steel tweezers moves into the frame, capturing the cold, metallic sheen from a single, harsh overhead light source. The tips of the tweezers meticulously and decisively grasp the exposed end of the splinter. With steady, deliberate pull, the tweezers slowly and firmly extract the splinter from the flesh, causing the skin to slightly bulge around the object before it is cleanly lifted. The entire scene is rendered in a stark, desaturated palette, dominated by the cool gray of the steel, the pale white tones of the skin, and the near-black of the splinter, creating an atmosphere of tension, focus, and clinical intensity that emphasizes the delicate and slightly uncomfortable nature of the procedure.", + "seed": 20027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_4/index_2986_seed_20027_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": -0.15520864315330982, + "vlm_logprob": -0.15520864315330982, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.15520864315330982, + "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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009155316278338432 + ], + "opposite_logprob_tokens": [ + -5.7591552734375 + ], + "logprob_reward": -0.009155316278338432, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009155316278338432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11960898339748383, + 0.16254469752311707, + 0.32564496994018555, + 0.03664521500468254, + 0.04774196818470955, + 0.04383661597967148, + 0.07677655667066574, + 0.034967366605997086, + 0.11700429022312164, + 0.03522932901978493 + ], + "credit_map": [ + 0.01076375413686037, + 0.01262250542640686, + 0.011513984762132168, + 0.031704433262348175, + 0.01030701957643032, + 0.0156683549284935, + 0.010596571490168571, + 0.01643236353993416, + 0.016650892794132233, + 0.018752358853816986, + 0.023174723610281944, + 0.02427493780851364, + 0.021859217435121536, + 0.01855640485882759, + 0.018614191561937332, + 0.02066197246313095, + 0.03566792979836464, + 0.046697087585926056, + 0.0682920590043068, + 0.04201463609933853, + 0.022734157741069794, + 0.027636051177978516, + 0.049066513776779175, + 0.03353654965758324, + 0.0037768909242004156, + 0.004029832314699888, + 0.007470914628356695, + 0.005941938143223524, + 0.003230101428925991, + 0.0034244495909661055, + 0.005280612967908382, + 0.0034904740750789642, + 0.007422795053571463, + 0.006636190693825483, + 0.006027380935847759, + 0.008528790436685085, + 0.0052190134301781654, + 0.0050454456359148026, + 0.0036526957992464304, + 0.005209657363593578, + 0.005086083896458149, + 0.005546740256249905, + 0.005213649943470955, + 0.010716364718973637, + 0.004550352692604065, + 0.004130946472287178, + 0.0036061119753867388, + 0.00498636718839407, + 0.009023399092257023, + 0.010409756563603878, + 0.010725541040301323, + 0.01835430972278118, + 0.005129735451191664, + 0.005935173947364092, + 0.009279574267566204, + 0.007919072173535824, + 0.004481087438762188, + 0.004807569086551666, + 0.0046758330427110195, + 0.006416575983166695, + 0.002218641573563218, + 0.0016597764333710074, + 0.004755046684294939, + 0.005952836479991674, + 0.012149697169661522, + 0.014599049463868141, + 0.016206147149205208, + 0.02131071127951145, + 0.013191038742661476, + 0.0062165954150259495, + 0.01751597970724106, + 0.015815075486898422, + 0.0035642727743834257, + 0.005041781347244978, + 0.004869386553764343, + 0.005580069497227669, + 0.003627260448411107, + 0.002040966646745801, + 0.003911094274371862, + 0.006594495382159948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.009155316278338432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061815746128559 + ], + "opposite_logprob_tokens": [ + -3.895618200302124 + ], + "logprob_reward": -0.02061815746128559, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061815746128559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32523709535598755, + 0.1029275581240654, + 0.07107126712799072, + 0.06345322728157043, + 0.06108517199754715, + 0.040492795407772064, + 0.04583818465471268, + 0.03838100656867027, + 0.22521741688251495, + 0.026296371594071388 + ], + "credit_map": [ + 0.028987029567360878, + 0.03981991112232208, + 0.03790434077382088, + 0.06131786108016968, + 0.03462115675210953, + 0.050805430859327316, + 0.03450223430991173, + 0.03727912902832031, + 0.011166260577738285, + 0.01425595860928297, + 0.014656602405011654, + 0.022268496453762054, + 0.010037621483206749, + 0.007597403135150671, + 0.011404607445001602, + 0.011540612205862999, + 0.008441874757409096, + 0.009861908853054047, + 0.011863773688673973, + 0.011915438808500767, + 0.006003826856613159, + 0.004808085039258003, + 0.010499208234250546, + 0.007677149027585983, + 0.006994176656007767, + 0.007894370704889297, + 0.010286218486726284, + 0.01522558368742466, + 0.0036336099728941917, + 0.005045064724981785, + 0.005890896078199148, + 0.008483304642140865, + 0.008356474339962006, + 0.007138247601687908, + 0.008709701709449291, + 0.011266050860285759, + 0.007541080936789513, + 0.005711325444281101, + 0.005272604990750551, + 0.007089683320373297, + 0.004710888955742121, + 0.004955715965479612, + 0.005562727339565754, + 0.008625289425253868, + 0.00421105045825243, + 0.004036324564367533, + 0.0039903391152620316, + 0.004400463309139013, + 0.004956269171088934, + 0.006523709744215012, + 0.006401638500392437, + 0.011257165111601353, + 0.0037854183465242386, + 0.0035867728292942047, + 0.004843005910515785, + 0.004484204575419426, + 0.004008923657238483, + 0.0059972465969622135, + 0.005707329139113426, + 0.007983770221471786, + 0.0024150789249688387, + 0.0016975251492112875, + 0.004492033738642931, + 0.006079099606722593, + 0.024002287536859512, + 0.02676481194794178, + 0.027766413986682892, + 0.04638016223907471, + 0.027998419478535652, + 0.012507584877312183, + 0.03385958448052406, + 0.02593814767897129, + 0.002716071205213666, + 0.0044661154970526695, + 0.0028905782382935286, + 0.006030958611518145, + 0.002386680571362376, + 0.001140907290391624, + 0.0032821563072502613, + 0.0033829023595899343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.02061815746128559 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004106542095541954 + ], + "opposite_logprob_tokens": [ + -5.504106521606445 + ], + "logprob_reward": -0.004106542095541954, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004106542095541954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11922311782836914, + 0.1477007269859314, + 0.09686307609081268, + 0.0813450962305069, + 0.09166380763053894, + 0.07729561626911163, + 0.05766703188419342, + 0.04446561634540558, + 0.21836677193641663, + 0.06540915369987488 + ], + "credit_map": [ + 0.010318326763808727, + 0.016087008640170097, + 0.016348039731383324, + 0.018426131457090378, + 0.012747041881084442, + 0.016739284619688988, + 0.014631269499659538, + 0.013926017098128796, + 0.01454020757228136, + 0.021230032667517662, + 0.0224516149610281, + 0.03338172659277916, + 0.011791687458753586, + 0.012380245141685009, + 0.016660192981362343, + 0.015265027992427349, + 0.013971704058349133, + 0.013845025561749935, + 0.016285063698887825, + 0.015317123383283615, + 0.008491122163832188, + 0.005459730979055166, + 0.012976294383406639, + 0.010517008602619171, + 0.008415396325290203, + 0.009734578430652618, + 0.010667214170098305, + 0.009438327513635159, + 0.015524482354521751, + 0.010296748951077461, + 0.006374301388859749, + 0.010894042439758778, + 0.01005571149289608, + 0.01143310870975256, + 0.010840605944395065, + 0.019114192575216293, + 0.013912193477153778, + 0.00964418612420559, + 0.007465677335858345, + 0.009198126383125782, + 0.007112369406968355, + 0.01067496556788683, + 0.010648212395608425, + 0.015233736485242844, + 0.00892800185829401, + 0.0064805299043655396, + 0.00981940608471632, + 0.008398395031690598, + 0.005394581239670515, + 0.0066239540465176105, + 0.008313493803143501, + 0.012085138820111752, + 0.005164331756532192, + 0.006784453522413969, + 0.006722632329910994, + 0.006578443106263876, + 0.005043459590524435, + 0.006306599359959364, + 0.007740158587694168, + 0.006387353874742985, + 0.003598825540393591, + 0.003163259709253907, + 0.006136900279670954, + 0.006089061964303255, + 0.02299550734460354, + 0.02605312503874302, + 0.0266126599162817, + 0.04075222834944725, + 0.02767997421324253, + 0.012559199705719948, + 0.03503072261810303, + 0.026683343574404716, + 0.008210297673940659, + 0.00816225353628397, + 0.006980753969401121, + 0.013755165971815586, + 0.005638330243527889, + 0.003798207500949502, + 0.009852712042629719, + 0.00901143066585064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.004106542095541954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048672039061784744 + ], + "opposite_logprob_tokens": [ + -3.0486719608306885 + ], + "logprob_reward": -0.048672039061784744, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048672039061784744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08705556392669678, + 0.08381731808185577, + 0.047639839351177216, + 0.06271369010210037, + 0.09489037096500397, + 0.06941837072372437, + 0.06766067445278168, + 0.08823388069868088, + 0.3863488435745239, + 0.01222140621393919 + ], + "credit_map": [ + 0.008510565385222435, + 0.012909467332065105, + 0.010429086163640022, + 0.012044170871376991, + 0.009443072602152824, + 0.013971776701509953, + 0.009998834691941738, + 0.009748587384819984, + 0.009046439081430435, + 0.011129098944365978, + 0.011058316566050053, + 0.014585290104150772, + 0.008811673149466515, + 0.007653271779417992, + 0.011920069344341755, + 0.0096131581813097, + 0.004455511923879385, + 0.005615120753645897, + 0.005381012801080942, + 0.010913469828665257, + 0.005662261974066496, + 0.005000869277864695, + 0.005441722925752401, + 0.005169870797544718, + 0.006451490800827742, + 0.007282808423042297, + 0.010661562904715538, + 0.013072414323687553, + 0.00566843943670392, + 0.005267253611236811, + 0.005756062921136618, + 0.008553660474717617, + 0.009914541617035866, + 0.011742200702428818, + 0.012437691912055016, + 0.020629679784178734, + 0.012616562657058239, + 0.009615438990294933, + 0.007522124797105789, + 0.01041212398558855, + 0.007040219381451607, + 0.007825622335076332, + 0.008901620283722878, + 0.01281675137579441, + 0.010506070218980312, + 0.008184152655303478, + 0.006765526253730059, + 0.0073784091509878635, + 0.0072358823381364346, + 0.009459419175982475, + 0.00936332531273365, + 0.019862063229084015, + 0.004206339828670025, + 0.004760449286550283, + 0.0059214308857917786, + 0.006851761136204004, + 0.00989032257348299, + 0.011386759579181671, + 0.012712972238659859, + 0.013346959836781025, + 0.007638786919414997, + 0.006548763252794743, + 0.013412700034677982, + 0.013296617195010185, + 0.0418534129858017, + 0.04391590505838394, + 0.050101373344659805, + 0.07909071445465088, + 0.047024525701999664, + 0.022497836500406265, + 0.05951102823019028, + 0.04235405474901199, + 0.0014043233823031187, + 0.0014099935069680214, + 0.0014487366424873471, + 0.002545080380514264, + 0.0013714114902541041, + 0.0006905784830451012, + 0.001789243659004569, + 0.0015620384365320206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.048672039061784744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6934911608695984 + ], + "opposite_logprob_tokens": [ + -0.6934911608695984 + ], + "logprob_reward": -0.6934911608695984, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934911608695984, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20085273683071136, + 0.22608068585395813, + 0.05350895971059799, + 0.05664803460240364, + 0.17556504905223846, + 0.07386744022369385, + 0.08538252860307693, + 0.03514749929308891, + 0.06434684991836548, + 0.02860022522509098 + ], + "credit_map": [ + 0.01753333769738674, + 0.020358920097351074, + 0.02699829824268818, + 0.04178572818636894, + 0.021161485463380814, + 0.026091543957591057, + 0.02199217677116394, + 0.02493124268949032, + 0.026412062346935272, + 0.02711448259651661, + 0.04032354801893234, + 0.044662926346063614, + 0.01822860725224018, + 0.01428270898759365, + 0.02622249349951744, + 0.028833851218223572, + 0.007366053760051727, + 0.006318666040897369, + 0.010433538816869259, + 0.008672842755913734, + 0.004100404679775238, + 0.004240972455590963, + 0.006116718053817749, + 0.006259763613343239, + 0.005595117341727018, + 0.006903364323079586, + 0.008798248134553432, + 0.011093867011368275, + 0.0060785324312746525, + 0.005535050295293331, + 0.005515830125659704, + 0.007128025405108929, + 0.016634788364171982, + 0.025534501299262047, + 0.023470986634492874, + 0.038182053714990616, + 0.02048039808869362, + 0.016914943233132362, + 0.0165419802069664, + 0.017805393785238266, + 0.005752748344093561, + 0.01111605204641819, + 0.01324131153523922, + 0.01569935493171215, + 0.006773658562451601, + 0.004972412716597319, + 0.008446494117379189, + 0.00786540750414133, + 0.007212282624095678, + 0.014115679077804089, + 0.013333561830222607, + 0.019093193113803864, + 0.007770161144435406, + 0.005752520635724068, + 0.009284350089728832, + 0.008820783346891403, + 0.004760385025292635, + 0.004227638244628906, + 0.00541820889338851, + 0.0067871855571866035, + 0.0021799143869429827, + 0.0018540213350206614, + 0.004376763012260199, + 0.005543386097997427, + 0.005053365137428045, + 0.00783388502895832, + 0.008510079234838486, + 0.01430879533290863, + 0.007529758848249912, + 0.0032449625432491302, + 0.010108227841556072, + 0.007757778745144606, + 0.0031519460026174784, + 0.00439111003652215, + 0.003930536098778248, + 0.006429650355130434, + 0.0023090119939297438, + 0.0011497983941808343, + 0.004164602607488632, + 0.003073569620028138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.6934911608695984 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, almost macroscopic, extreme close-up focuses on a human fingertip, the intricate spirals of the fingerprint rendered in sharp detail. A jagged, dark brown splinter is clearly embedded beneath the skin's surface, causing a slight red inflammation around the point of entry. A pair of sharp-tipped, medical-grade steel tweezers moves into the frame, capturing the cold, metallic sheen from a single, harsh overhead light source. The tips of the tweezers meticulously and decisively grasp the exposed end of the splinter. With steady, deliberate pull, the tweezers slowly and firmly extract the splinter from the flesh, causing the skin to slightly bulge around the object before it is cleanly lifted. The entire scene is rendered in a stark, desaturated palette, dominated by the cool gray of the steel, the pale white tones of the skin, and the near-black of the splinter, creating an atmosphere of tension, focus, and clinical intensity that emphasizes the delicate and slightly uncomfortable nature of the procedure.", + "seed": 20028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_4/index_2986_seed_20028_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": -0.2946377947926521, + "vlm_logprob": -0.2946377947926521, + "vlm_yes_no_margin": 2.424999988079071, + "vlm_token_credit": -0.2946377947926521, + "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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012280801311135292 + ], + "opposite_logprob_tokens": [ + -5.3872809410095215 + ], + "logprob_reward": -0.012280801311135292, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012280801311135292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2524247169494629, + 0.11752501875162125, + 0.03153296560049057, + 0.09828102588653564, + 0.0304361954331398, + 0.0731031745672226, + 0.055506519973278046, + 0.19247472286224365, + 0.0765366405248642, + 0.07217908650636673 + ], + "credit_map": [ + 0.025066949427127838, + 0.030153147876262665, + 0.045924894511699677, + 0.03935086354613304, + 0.037306539714336395, + 0.026998279616236687, + 0.028567777946591377, + 0.01905624009668827, + 0.011459625326097012, + 0.014240209944546223, + 0.021782590076327324, + 0.017951665446162224, + 0.010873781517148018, + 0.016982151195406914, + 0.012194469571113586, + 0.012040513567626476, + 0.0030096517875790596, + 0.003991709090769291, + 0.0050823017954826355, + 0.005572445224970579, + 0.002149202860891819, + 0.004189521539956331, + 0.004191380925476551, + 0.0033467498142272234, + 0.01275668479502201, + 0.01453770138323307, + 0.015883011743426323, + 0.010249847546219826, + 0.011269924230873585, + 0.012296813540160656, + 0.009619333781301975, + 0.011667708866298199, + 0.0033521633595228195, + 0.003164729569107294, + 0.005439698696136475, + 0.003430258249863982, + 0.0038396306335926056, + 0.004774345085024834, + 0.003180230036377907, + 0.0032551372423768044, + 0.010882900096476078, + 0.012144273146986961, + 0.017178494483232498, + 0.007473493926227093, + 0.005621361080557108, + 0.007814247161149979, + 0.006215415894985199, + 0.005772982258349657, + 0.006595801096409559, + 0.007276416290551424, + 0.015926439315080643, + 0.007019944489002228, + 0.0038424022495746613, + 0.005001791287213564, + 0.005314638372510672, + 0.004529081750661135, + 0.014858441427350044, + 0.015524462796747684, + 0.02937915362417698, + 0.03850385919213295, + 0.03365156054496765, + 0.021984584629535675, + 0.01977771334350109, + 0.018794940784573555, + 0.006721216253936291, + 0.00745367631316185, + 0.014198428019881248, + 0.010404678992927074, + 0.009563208557665348, + 0.011731171980500221, + 0.00836003478616476, + 0.008104213513433933, + 0.008659603074193, + 0.010182537138462067, + 0.011916148476302624, + 0.008541950024664402, + 0.011833085678517818, + 0.007998411543667316, + 0.007198484614491463, + 0.005848863162100315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.012280801311135292 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023325225338339806 + ], + "opposite_logprob_tokens": [ + -3.773325204849243 + ], + "logprob_reward": -0.023325225338339806, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023325225338339806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2543584108352661, + 0.09723715484142303, + 0.04321441054344177, + 0.22586779296398163, + 0.08174639195203781, + 0.03557603061199188, + 0.07018792629241943, + 0.07339046895503998, + 0.05475622043013573, + 0.0636652335524559 + ], + "credit_map": [ + 0.02832542173564434, + 0.03353986516594887, + 0.04667417332530022, + 0.023134103044867516, + 0.047510530799627304, + 0.03441828861832619, + 0.019472677260637283, + 0.02128334529697895, + 0.011039556935429573, + 0.015168149024248123, + 0.011842272244393826, + 0.013910730369389057, + 0.01668490096926689, + 0.011145106516778469, + 0.00938537809997797, + 0.008061052300035954, + 0.00398966483771801, + 0.004165162798017263, + 0.007494068238884211, + 0.007712301332503557, + 0.006052749697118998, + 0.005059279501438141, + 0.004391134716570377, + 0.0043500447645783424, + 0.023973509669303894, + 0.030220258980989456, + 0.033949337899684906, + 0.02418648824095726, + 0.03020981140434742, + 0.039260994642972946, + 0.01923936977982521, + 0.024828005582094193, + 0.008728213608264923, + 0.008212029933929443, + 0.013435622677206993, + 0.007695126812905073, + 0.013287543319165707, + 0.010791908949613571, + 0.010128960013389587, + 0.00946697499603033, + 0.004221345763653517, + 0.004460759926587343, + 0.008380520157516003, + 0.004307151772081852, + 0.0024868829641491175, + 0.005204449873417616, + 0.0032548257149755955, + 0.0032600904814898968, + 0.009445645846426487, + 0.009599445387721062, + 0.01257895864546299, + 0.008905288763344288, + 0.005037249531596899, + 0.010278827510774136, + 0.005724412389099598, + 0.008618091233074665, + 0.005502632353454828, + 0.006376987788826227, + 0.010584579780697823, + 0.007591927889734507, + 0.012727535329759121, + 0.014567198231816292, + 0.007015777286142111, + 0.009023822844028473, + 0.008707229048013687, + 0.010223163291811943, + 0.0077645257115364075, + 0.0071077230386435986, + 0.00528737111017108, + 0.005429478827863932, + 0.005081404931843281, + 0.005155324470251799, + 0.011767680756747723, + 0.012505619786679745, + 0.007149347569793463, + 0.00588028272613883, + 0.010206446051597595, + 0.006541762035340071, + 0.004578283987939358, + 0.005035806447267532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.023325225338339806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135151267051697 + ], + "opposite_logprob_tokens": [ + -1.3135151863098145 + ], + "logprob_reward": -0.3135151267051697, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135151267051697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446099430322647, + 0.05775423347949982, + 0.07893509417772293, + 0.24665391445159912, + 0.08673715591430664, + 0.0547281950712204, + 0.06823549419641495, + 0.05619175732135773, + 0.13907456398010254, + 0.06707970798015594 + ], + "credit_map": [ + 0.01757570169866085, + 0.020130189135670662, + 0.023468254134058952, + 0.01977519877254963, + 0.02392474189400673, + 0.015185758471488953, + 0.012265686877071857, + 0.012284411117434502, + 0.005805489141494036, + 0.0062732030637562275, + 0.011617470532655716, + 0.010074383579194546, + 0.0074483370408415794, + 0.005174041260033846, + 0.0062319329008460045, + 0.005129375960677862, + 0.008377636782824993, + 0.010986789129674435, + 0.015168702229857445, + 0.010747796855866909, + 0.008939016610383987, + 0.008860930800437927, + 0.0077491444535553455, + 0.008105074055492878, + 0.026973921805620193, + 0.03184056282043457, + 0.04479037970304489, + 0.031170791015028954, + 0.03756368160247803, + 0.030414728447794914, + 0.02299216203391552, + 0.02090766839683056, + 0.011014408431947231, + 0.009315816685557365, + 0.022538913413882256, + 0.008644967339932919, + 0.008239654824137688, + 0.011775687336921692, + 0.007891133427619934, + 0.007316575385630131, + 0.005736831109970808, + 0.006587203126400709, + 0.011796335689723492, + 0.006391433533281088, + 0.004205218981951475, + 0.0083413515239954, + 0.0055839745327830315, + 0.006085842847824097, + 0.007135082967579365, + 0.00816374272108078, + 0.011215581558644772, + 0.006915938574820757, + 0.006486881989985704, + 0.012887316755950451, + 0.005300326272845268, + 0.010130625218153, + 0.011955734342336655, + 0.012164251878857613, + 0.00771669577807188, + 0.0039750332944095135, + 0.004205427132546902, + 0.007685462478548288, + 0.003682873910292983, + 0.004806278273463249, + 0.029819263145327568, + 0.03497830405831337, + 0.015271652489900589, + 0.015007466077804565, + 0.014954361133277416, + 0.009199460968375206, + 0.010636666789650917, + 0.009207394905388355, + 0.010447493754327297, + 0.012478805147111416, + 0.009918928146362305, + 0.005643906537443399, + 0.012520259246230125, + 0.007084951736032963, + 0.004691548179835081, + 0.00429381662979722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.3135151267051697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005671411752701 + ], + "opposite_logprob_tokens": [ + -2.695056676864624 + ], + "logprob_reward": -0.07005671411752701, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005671411752701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11542599648237228, + 0.04309339448809624, + 0.07358773052692413, + 0.4092426598072052, + 0.049691442400217056, + 0.022903913632035255, + 0.04520740360021591, + 0.09433259069919586, + 0.08523024618625641, + 0.061284616589546204 + ], + "credit_map": [ + 0.014183778315782547, + 0.014866679906845093, + 0.024874277412891388, + 0.01187150739133358, + 0.017905371263623238, + 0.013101845048367977, + 0.00828472152352333, + 0.010337810963392258, + 0.004694027826189995, + 0.005790357477962971, + 0.008400725200772285, + 0.005514315329492092, + 0.006369971204549074, + 0.004902961663901806, + 0.0034311716444790363, + 0.003989866003394127, + 0.006805629935115576, + 0.008399217389523983, + 0.013659161515533924, + 0.010175769217312336, + 0.01046228501945734, + 0.008908886462450027, + 0.007004844956099987, + 0.008171935565769672, + 0.04844292998313904, + 0.05437411740422249, + 0.07195477187633514, + 0.03442053869366646, + 0.05526850000023842, + 0.06006871536374092, + 0.03737374395132065, + 0.04733934998512268, + 0.005802907049655914, + 0.005706199910491705, + 0.012570016086101532, + 0.005093683954328299, + 0.005487225484102964, + 0.005415032152086496, + 0.00516725517809391, + 0.004449123982340097, + 0.002898501930758357, + 0.003229000838473439, + 0.005581799894571304, + 0.0026906330604106188, + 0.0016426150687038898, + 0.0028156761545687914, + 0.0018450176576152444, + 0.0022006689105182886, + 0.006593532394617796, + 0.0068262177519500256, + 0.009691787883639336, + 0.005801151506602764, + 0.0031882626935839653, + 0.005592632573097944, + 0.0032971599139273167, + 0.004216656554490328, + 0.011987793259322643, + 0.012974242679774761, + 0.010588526725769043, + 0.007663436699658632, + 0.015156778506934643, + 0.016806676983833313, + 0.008585086092352867, + 0.010570045560598373, + 0.007898686453700066, + 0.009371812455356121, + 0.018433770164847374, + 0.014295127242803574, + 0.010564721189439297, + 0.008052070625126362, + 0.009905108250677586, + 0.0067089456133544445, + 0.01211045403033495, + 0.010888488031923771, + 0.00827834289520979, + 0.005082213319838047, + 0.00923687219619751, + 0.006376906298100948, + 0.004224864766001701, + 0.005086473189294338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.07005671411752701 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a human fingertip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a jagged dark brown splinter visible embedded beneath the skin's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are sharp-tipped medical-grade steel tweezers visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the tweezers grasp the splinter and extract it from the flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single harsh overhead light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0540111064910889 + ], + "opposite_logprob_tokens": [ + -0.4290110468864441 + ], + "logprob_reward": -1.0540111064910889, + "yes_no_margin_reward": -0.6250000596046448, + "token_credit_reward": -1.0540111064910889, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22174625098705292, + 0.0690242350101471, + 0.07800301909446716, + 0.1448778361082077, + 0.0899861603975296, + 0.047562990337610245, + 0.04245565086603165, + 0.16003718972206116, + 0.08151289075613022, + 0.06479380279779434 + ], + "credit_map": [ + 0.022438952699303627, + 0.029576029628515244, + 0.04433741047978401, + 0.028555719181895256, + 0.030462468042969704, + 0.028493929654359818, + 0.02182842418551445, + 0.016053322702646255, + 0.007730519864708185, + 0.011005819775164127, + 0.008787138387560844, + 0.00936825666576624, + 0.011527959257364273, + 0.0075355288572609425, + 0.007633789908140898, + 0.005435217637568712, + 0.008085204288363457, + 0.012592517770826817, + 0.010603317059576511, + 0.013087750412523746, + 0.0070557910948991776, + 0.009294700808823109, + 0.008878201246261597, + 0.008405535481870174, + 0.020593611523509026, + 0.026123235002160072, + 0.026811709627509117, + 0.01488774549216032, + 0.014321422204375267, + 0.018407845869660378, + 0.011444281786680222, + 0.012287985533475876, + 0.010046924464404583, + 0.010176683776080608, + 0.025374386459589005, + 0.008278805762529373, + 0.005066622979938984, + 0.013270338997244835, + 0.00933529157191515, + 0.008437108248472214, + 0.005993926897644997, + 0.00773335387930274, + 0.009902335703372955, + 0.004095642361789942, + 0.0033604681957513094, + 0.0077504427172243595, + 0.003303937613964081, + 0.005422886461019516, + 0.00582408532500267, + 0.005957420915365219, + 0.010861750692129135, + 0.005624096840620041, + 0.002256059553474188, + 0.0037078887689858675, + 0.005122429691255093, + 0.0031019174493849277, + 0.01693463884294033, + 0.017301253974437714, + 0.04092325270175934, + 0.026427604258060455, + 0.011660566553473473, + 0.016277173534035683, + 0.017664341256022453, + 0.012848363257944584, + 0.012495706789195538, + 0.01263502798974514, + 0.02047569863498211, + 0.011816518381237984, + 0.005841431673616171, + 0.0070909964852035046, + 0.006084590684622526, + 0.005072925239801407, + 0.008254267275333405, + 0.008240912109613419, + 0.012215024791657925, + 0.009074758738279343, + 0.00774156441912055, + 0.007583648897707462, + 0.005271901376545429, + 0.006411725655198097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -1.0540111064910889 + } + ], + "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/0000155/group_5/index_tensor([7397, 7397])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_5/index_tensor([7397, 7397])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e06a9456418ad8ef9b3bf8a2ce059f3290a476d2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_5/index_tensor([7397, 7397])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An unsettling, tilted, high-angle shot places the viewer in a scene almost engulfed in pure darkness. A strong, sharply directed beam of cold white light streams down from the upper and side of the frame, precisely illuminating an arthropod creature emerging from the deep shadows. The creature's deep black chitinous carapace is covered in irregular, bony knobs and protuberances. Its surface, glossy and smooth as if coated with oil, reflects sharp, glaring highlights from the single light source, while the rest of its body blends completely into the bottomless shadows. With a slow, deliberate rhythm, it slowly raises its front half, its long, jointed legs carefully adjusting their footing on the damp rocky ground, its front arching upward with a powerful force, as if performing some ancient ritual. Though slow, this movement is imbued with a menacing, poised momentum. The entire composition is filled with strong chiaroscuro (light and dark contrasts). The light sculpts every detail of the creature's silhouette, creating a huge contrast with the surrounding darkness, perfectly recreating Caravaggio's classical tension and creating a suffocating, silent atmosphere mixed with fear and classical beauty.", + "seed": 7234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_5/index_7397_seed_7234_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": -0.019038160529453306, + "vlm_logprob": -0.019038160529453306, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.019038160529453306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05249451845884323 + ], + "opposite_logprob_tokens": [ + -3.677494525909424 + ], + "logprob_reward": -0.05249451845884323, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05249451845884323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06708423048257828, + 0.014196332544088364, + 0.012400340288877487, + 0.015830297023057938, + 0.012078916653990746, + 0.213063046336174, + 0.6081177592277527, + 0.02196667343378067, + 0.026542911306023598, + 0.008719470351934433 + ], + "credit_map": [ + 0.006900292821228504, + 0.005760646890848875, + 0.018404657021164894, + 0.007167479954659939, + 0.005804067477583885, + 0.006705523934215307, + 0.005845280829817057, + 0.010496286675333977, + 0.0015972532564774156, + 0.0014822710072621703, + 0.0022031692788004875, + 0.0016034654108807445, + 0.001539929653517902, + 0.001602443982847035, + 0.002300094813108444, + 0.0018677054904401302, + 0.0009686664561741054, + 0.0011113592190667987, + 0.0011589787900447845, + 0.0009745785500854254, + 0.001972209895029664, + 0.0029661909211426973, + 0.0007530190632678568, + 0.0024953377433121204, + 0.0014036233769729733, + 0.0018079798901453614, + 0.002094038762152195, + 0.0021276106126606464, + 0.0016093621961772442, + 0.002260268200188875, + 0.0022184501867741346, + 0.0023089637979865074, + 0.001397099462337792, + 0.0012696669436991215, + 0.0010156860807910562, + 0.0005850665620528162, + 0.0012903718743473291, + 0.0027426378801465034, + 0.0009574209689162672, + 0.002820966299623251, + 0.00486280582845211, + 0.001823222148232162, + 0.0016557107446715236, + 0.00512287812307477, + 0.003473937977105379, + 0.09325166046619415, + 0.010970129631459713, + 0.09190269559621811, + 0.011005993001163006, + 0.004448658786714077, + 0.009451227262616158, + 0.013773169368505478, + 0.008528847247362137, + 0.2710771858692169, + 0.02986891195178032, + 0.2599637508392334, + 0.002884246176108718, + 0.0020932273473590612, + 0.0024989221710711718, + 0.0026341078337281942, + 0.002747142221778631, + 0.00295273307710886, + 0.0027894568629562855, + 0.0033668370451778173, + 0.002881412860006094, + 0.0017456666100770235, + 0.0019454776775091887, + 0.002081009792163968, + 0.002569261472672224, + 0.006196900736540556, + 0.005395907908678055, + 0.003727275412529707, + 0.0012638942571356893, + 0.0007336924318224192, + 0.0008327553514391184, + 0.0005523278377950191, + 0.0011731209233403206, + 0.001071393140591681, + 0.002004128647968173, + 0.0010881578782573342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.05249451845884323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330903708934784 + ], + "opposite_logprob_tokens": [ + -3.7733089923858643 + ], + "logprob_reward": -0.02330903708934784, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330903708934784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12384112179279327, + 0.05092687904834747, + 0.027002597227692604, + 0.012105938978493214, + 0.021818185225129128, + 0.3063499331474304, + 0.38208094239234924, + 0.029360754415392876, + 0.02553166262805462, + 0.020981978625059128 + ], + "credit_map": [ + 0.00830871518701315, + 0.010726670734584332, + 0.009522881358861923, + 0.009105238132178783, + 0.018732639029622078, + 0.020587174221873283, + 0.026836220175027847, + 0.020021585747599602, + 0.003606425365433097, + 0.0056243278086185455, + 0.006279792636632919, + 0.0061173331923782825, + 0.005375607870519161, + 0.005937997717410326, + 0.011427098885178566, + 0.006558296736329794, + 0.0019467623205855489, + 0.0027607062365859747, + 0.002157376380637288, + 0.002313721925020218, + 0.0040405853651463985, + 0.006359907332807779, + 0.0017891557654365897, + 0.005634382367134094, + 0.002137166215106845, + 0.0012616447638720274, + 0.0015432219952344894, + 0.0015600459882989526, + 0.0010040274355560541, + 0.0016852424014359713, + 0.001075695501640439, + 0.0018388940952718258, + 0.0028940492775291204, + 0.0018009794875979424, + 0.0031102951616048813, + 0.0017897129291668534, + 0.0016160371014848351, + 0.0033000940456986427, + 0.0028866510838270187, + 0.004420365672558546, + 0.007382551208138466, + 0.002994221169501543, + 0.0035576955415308475, + 0.007103911135345697, + 0.0060685365460813046, + 0.13312682509422302, + 0.01586560718715191, + 0.13025058805942535, + 0.0066005815751850605, + 0.002676783362403512, + 0.005820872262120247, + 0.008644680492579937, + 0.005256213713437319, + 0.17092974483966827, + 0.01907160133123398, + 0.16308046877384186, + 0.0046575251035392284, + 0.003523868275806308, + 0.003220121143385768, + 0.0033836709335446358, + 0.0032881570514291525, + 0.004205037839710712, + 0.004128691740334034, + 0.002953681629151106, + 0.003628338687121868, + 0.0018633855506777763, + 0.0022229847963899374, + 0.00225364463403821, + 0.002714064670726657, + 0.004782269708812237, + 0.004876330494880676, + 0.003190644085407257, + 0.0029548669699579477, + 0.0015772837214171886, + 0.0017837673658505082, + 0.001566292718052864, + 0.0019389638910070062, + 0.002847653115168214, + 0.006102897226810455, + 0.0022102543152868748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.02330903708934784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823891816660762 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002823891816660762, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823891816660762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0300449188798666, + 0.05999501422047615, + 0.011802990920841694, + 0.022137094289064407, + 0.02573033608496189, + 0.3575064539909363, + 0.4247901737689972, + 0.01836453005671501, + 0.03710852935910225, + 0.012519887648522854 + ], + "credit_map": [ + 0.0022438387386500835, + 0.0028989894781261683, + 0.0042334794998168945, + 0.0030809962190687656, + 0.004597916733473539, + 0.003763460787013173, + 0.004427434876561165, + 0.004798809997737408, + 0.005686567630618811, + 0.009232176467776299, + 0.010253834538161755, + 0.0074886782094836235, + 0.005576111376285553, + 0.006156533025205135, + 0.007407343480736017, + 0.008193783462047577, + 0.0011209616204723716, + 0.0012345662107691169, + 0.0011751424754038453, + 0.0015325129497796297, + 0.0010829427046701312, + 0.002210968639701605, + 0.0013075193855911493, + 0.0021383787970989943, + 0.0022128236014395952, + 0.0020104367285966873, + 0.003773665986955166, + 0.002119910204783082, + 0.0022400622256100178, + 0.0036122221499681473, + 0.00263338815420866, + 0.003534589195623994, + 0.0035914569161832333, + 0.002856925828382373, + 0.003314813831821084, + 0.0015883190790191293, + 0.0019080123165622354, + 0.004078238736838102, + 0.0027816216461360455, + 0.00561095355078578, + 0.008682151325047016, + 0.003765298519283533, + 0.00427620904520154, + 0.008545495569705963, + 0.0067543769255280495, + 0.15466205775737762, + 0.018783096224069595, + 0.1520378440618515, + 0.007940861396491528, + 0.003478041850030422, + 0.00673100259155035, + 0.009897565469145775, + 0.006059408653527498, + 0.1890096366405487, + 0.021137140691280365, + 0.18053656816482544, + 0.004067771136760712, + 0.002431435976177454, + 0.002166682155802846, + 0.0016189281595870852, + 0.0016963907983154058, + 0.0023014163598418236, + 0.0022729209158569574, + 0.00180898851249367, + 0.005702368449419737, + 0.0037269708700478077, + 0.002802253467962146, + 0.00277672684751451, + 0.0038697088602930307, + 0.00655758660286665, + 0.006649855058640242, + 0.005023065954446793, + 0.0013549450086429715, + 0.001465167966671288, + 0.0010322972666472197, + 0.00124111317563802, + 0.0015646859537810087, + 0.002160362433642149, + 0.002083391882479191, + 0.0016179268714040518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.002823891816660762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1039208173751831, + 0.08119052648544312, + 0.03456813842058182, + 0.023857658728957176, + 0.05338433384895325, + 0.3860029876232147, + 0.15952178835868835, + 0.08297871053218842, + 0.047341905534267426, + 0.027233052998781204 + ], + "credit_map": [ + 0.009505774825811386, + 0.009601481258869171, + 0.020943226292729378, + 0.010508782230317593, + 0.012297660112380981, + 0.011264974251389503, + 0.01428182702511549, + 0.015517096035182476, + 0.006824430078268051, + 0.008641415275633335, + 0.01348335761576891, + 0.00977354496717453, + 0.008563288487493992, + 0.009734791703522205, + 0.012941727414727211, + 0.011227982118725777, + 0.0036318767815828323, + 0.0038108117878437042, + 0.004451824817806482, + 0.00318573578260839, + 0.006522313691675663, + 0.005212648771703243, + 0.003240481950342655, + 0.00451244693249464, + 0.0019790837541222572, + 0.002079274272546172, + 0.0040782359428703785, + 0.0023457559291273355, + 0.0027280780486762524, + 0.003825155086815357, + 0.003448516596108675, + 0.0033735623583197594, + 0.00804923102259636, + 0.005837134085595608, + 0.006671819370239973, + 0.0037813824601471424, + 0.005849361419677734, + 0.006966057233512402, + 0.006226277910172939, + 0.010003074072301388, + 0.014118724502623081, + 0.008227796293795109, + 0.008991803973913193, + 0.010615437291562557, + 0.011232887394726276, + 0.15716850757598877, + 0.021269306540489197, + 0.15437857806682587, + 0.007885366678237915, + 0.0031137478072196245, + 0.005919207353144884, + 0.00722544826567173, + 0.004925289656966925, + 0.06097758561372757, + 0.011590080335736275, + 0.057885076850652695, + 0.010098849423229694, + 0.00732663506641984, + 0.008059827610850334, + 0.009667536243796349, + 0.011219370178878307, + 0.012631245888769627, + 0.013629314489662647, + 0.01034594140946865, + 0.006470558233559132, + 0.0045519196428358555, + 0.006259603891521692, + 0.005527956411242485, + 0.004251265432685614, + 0.007258104160428047, + 0.006456018425524235, + 0.006566486321389675, + 0.0029852776788175106, + 0.0017168589401990175, + 0.0021522114984691143, + 0.002359224483370781, + 0.0035455096513032913, + 0.004258362110704184, + 0.007328098174184561, + 0.0028875120915472507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.016066256910562515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004970983718521893 + ], + "opposite_logprob_tokens": [ + -7.625496864318848 + ], + "logprob_reward": -0.0004970983718521893, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004970983718521893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03967510908842087, + 0.03899029269814491, + 0.04854656755924225, + 0.018684478476643562, + 0.030360765755176544, + 0.2721722424030304, + 0.44494351744651794, + 0.027623433619737625, + 0.036962419748306274, + 0.04204115644097328 + ], + "credit_map": [ + 0.003658576402813196, + 0.0029648402705788612, + 0.0075647104531526566, + 0.0037969062104821205, + 0.004005251452326775, + 0.004767047241330147, + 0.007161314599215984, + 0.005756461061537266, + 0.0025331147480756044, + 0.0032698335126042366, + 0.004188328981399536, + 0.004726201295852661, + 0.004041297361254692, + 0.006094006355851889, + 0.008801978081464767, + 0.005335534457117319, + 0.003373874817043543, + 0.004759419709444046, + 0.003973069600760937, + 0.00429706834256649, + 0.00917226355522871, + 0.01089819148182869, + 0.004151586908847094, + 0.007921095006167889, + 0.0018952031387016177, + 0.0019364063628017902, + 0.0031301998533308506, + 0.0022715204395353794, + 0.0012962687760591507, + 0.003010638989508152, + 0.002423158148303628, + 0.002721083117648959, + 0.004395302850753069, + 0.003098957473412156, + 0.003879899624735117, + 0.0020449936855584383, + 0.0019678587559610605, + 0.004913810174912214, + 0.00313854799605906, + 0.00692139333114028, + 0.006695074029266834, + 0.0028292122296988964, + 0.0035583185963332653, + 0.00680434750393033, + 0.004323068540543318, + 0.11748072504997253, + 0.01443219929933548, + 0.11604930460453033, + 0.00830703042447567, + 0.0039376565255224705, + 0.006898657884448767, + 0.010519489645957947, + 0.006663463544100523, + 0.1976228803396225, + 0.022189659997820854, + 0.18880467116832733, + 0.002930667018517852, + 0.0026163035072386265, + 0.002714789705350995, + 0.003825309220701456, + 0.0033118713181465864, + 0.004891275428235531, + 0.004077186807990074, + 0.0032560303807258606, + 0.006025838200002909, + 0.003173637669533491, + 0.002935594879090786, + 0.003040590789169073, + 0.003962289541959763, + 0.006966389715671539, + 0.0067502586171031, + 0.004107820801436901, + 0.005057364236563444, + 0.0040204403921961784, + 0.002637845929712057, + 0.003926088567823172, + 0.004135720431804657, + 0.005511687137186527, + 0.010526432655751705, + 0.006225578486919403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.0004970983718521893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsettling, tilted, high-angle shot places the viewer in a scene almost engulfed in pure darkness. A strong, sharply directed beam of cold white light streams down from the upper and side of the frame, precisely illuminating an arthropod creature emerging from the deep shadows. The creature's deep black chitinous carapace is covered in irregular, bony knobs and protuberances. Its surface, glossy and smooth as if coated with oil, reflects sharp, glaring highlights from the single light source, while the rest of its body blends completely into the bottomless shadows. With a slow, deliberate rhythm, it slowly raises its front half, its long, jointed legs carefully adjusting their footing on the damp rocky ground, its front arching upward with a powerful force, as if performing some ancient ritual. Though slow, this movement is imbued with a menacing, poised momentum. The entire composition is filled with strong chiaroscuro (light and dark contrasts). The light sculpts every detail of the creature's silhouette, creating a huge contrast with the surrounding darkness, perfectly recreating Caravaggio's classical tension and creating a suffocating, silent atmosphere mixed with fear and classical beauty.", + "seed": 7235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_5/index_7397_seed_7235_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": -0.029241074935998767, + "vlm_logprob": -0.029241074935998767, + "vlm_yes_no_margin": 4.8249999523162845, + "vlm_token_credit": -0.029241074935998767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10285503417253494 + ], + "opposite_logprob_tokens": [ + -3.1028549671173096 + ], + "logprob_reward": -0.10285503417253494, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.10285503417253494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13529451191425323, + 0.16805963218212128, + 0.05413157492876053, + 0.39741256833076477, + 0.046180639415979385, + 0.07287627458572388, + 0.04888860136270523, + 0.015021833591163158, + 0.024599621072411537, + 0.03753477707505226 + ], + "credit_map": [ + 0.017456430941820145, + 0.017949707806110382, + 0.017722636461257935, + 0.012401380576193333, + 0.01475012581795454, + 0.01781887374818325, + 0.01750037632882595, + 0.019694983959197998, + 0.0200248621404171, + 0.018861211836338043, + 0.020897021517157555, + 0.01857135444879532, + 0.019005276262760162, + 0.027179401367902756, + 0.020804161205887794, + 0.022716332226991653, + 0.007671710103750229, + 0.006991671398282051, + 0.0076954360119998455, + 0.005440911743789911, + 0.007542535662651062, + 0.004590090364217758, + 0.007175972685217857, + 0.007023245096206665, + 0.0560847632586956, + 0.05265520513057709, + 0.07212971895933151, + 0.05355430394411087, + 0.0491446889936924, + 0.029668031260371208, + 0.05132105201482773, + 0.03285481408238411, + 0.007551525719463825, + 0.005309329833835363, + 0.004626784007996321, + 0.00472051789984107, + 0.006396095268428326, + 0.0038543990813195705, + 0.005565845873206854, + 0.008156141266226768, + 0.010445359162986279, + 0.010221946984529495, + 0.007774901110678911, + 0.009138003922998905, + 0.008986253291368484, + 0.004174455534666777, + 0.010755741968750954, + 0.011379610747098923, + 0.009297171607613564, + 0.0042458390817046165, + 0.004470080602914095, + 0.005245401989668608, + 0.0048704552464187145, + 0.005995749495923519, + 0.0072684516198933125, + 0.007495452184230089, + 0.0019347332417964935, + 0.0025517677422612906, + 0.0017430409789085388, + 0.001931186649017036, + 0.0015314470510929823, + 0.0014614129904657602, + 0.0014735495205968618, + 0.0023946957662701607, + 0.0035241611767560244, + 0.0038904494140297174, + 0.0031917798332870007, + 0.002853859681636095, + 0.0028779623098671436, + 0.0027959675062447786, + 0.002824680181220174, + 0.0026407595723867416, + 0.0036238415632396936, + 0.002921595238149166, + 0.006240352988243103, + 0.004147697705775499, + 0.003573345486074686, + 0.0056355358101427555, + 0.004963116720318794, + 0.006429293658584356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.10285503417253494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020617924630641937 + ], + "opposite_logprob_tokens": [ + -3.895617961883545 + ], + "logprob_reward": -0.020617924630641937, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020617924630641937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11130959540605545, + 0.04362476244568825, + 0.05164029449224472, + 0.16905546188354492, + 0.09182922542095184, + 0.21548612415790558, + 0.08606427162885666, + 0.03154975548386574, + 0.03418566659092903, + 0.1652548611164093 + ], + "credit_map": [ + 0.012296566739678383, + 0.019778836518526077, + 0.015843557193875313, + 0.01568339392542839, + 0.00980345718562603, + 0.0104833310469985, + 0.01028943806886673, + 0.017131011933088303, + 0.006452817935496569, + 0.005080853123217821, + 0.008906244300305843, + 0.0060108317993581295, + 0.0032998884562402964, + 0.0048565431497991085, + 0.0037779740523546934, + 0.0052396091632544994, + 0.008120086044073105, + 0.005264968611299992, + 0.009037121199071407, + 0.005164450500160456, + 0.004352315329015255, + 0.005872941110283136, + 0.006393809802830219, + 0.0074346051551401615, + 0.025899138301610947, + 0.011966940015554428, + 0.015001101419329643, + 0.017485441640019417, + 0.024754777550697327, + 0.015176155604422092, + 0.04038504883646965, + 0.018386859446763992, + 0.013591608963906765, + 0.008449233137071133, + 0.01285046711564064, + 0.010949719697237015, + 0.012399500235915184, + 0.009085489436984062, + 0.012765354476869106, + 0.011737850494682789, + 0.027220990508794785, + 0.031817417591810226, + 0.027298858389258385, + 0.03239356726408005, + 0.02584267035126686, + 0.021223271265625954, + 0.02451932802796364, + 0.02517000585794449, + 0.013358048163354397, + 0.006108708214014769, + 0.005599732976406813, + 0.007531738840043545, + 0.008509484119713306, + 0.021191511303186417, + 0.01067892275750637, + 0.013086125254631042, + 0.0043725380674004555, + 0.004033603705465794, + 0.005971167236566544, + 0.004854824859648943, + 0.0027666869573295116, + 0.00273123849183321, + 0.002750094048678875, + 0.004069602116942406, + 0.00405971659347415, + 0.004826702177524567, + 0.004064307548105717, + 0.0036751413717865944, + 0.004964036401361227, + 0.004014125559478998, + 0.004510903265327215, + 0.004070732742547989, + 0.019805289804935455, + 0.013743075542151928, + 0.01988983154296875, + 0.01745612733066082, + 0.016253076493740082, + 0.02424481138586998, + 0.02363770082592964, + 0.030224939808249474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.020617924630641937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102030768990517 + ], + "opposite_logprob_tokens": [ + -5.5041022300720215 + ], + "logprob_reward": -0.004102030768990517, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102030768990517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09793944656848907, + 0.09071417152881622, + 0.13339099287986755, + 0.0925557017326355, + 0.11120086908340454, + 0.050430066883563995, + 0.11589260399341583, + 0.04168426990509033, + 0.14784394204616547, + 0.11834795773029327 + ], + "credit_map": [ + 0.011886795982718468, + 0.016978327184915543, + 0.012977815233170986, + 0.01390603743493557, + 0.008661583997309208, + 0.009619471617043018, + 0.009598182514309883, + 0.01431123074144125, + 0.010748322121798992, + 0.009632643312215805, + 0.013272237032651901, + 0.011141175404191017, + 0.009399443864822388, + 0.016841361299157143, + 0.009704381227493286, + 0.009974601678550243, + 0.019392892718315125, + 0.01645110361278057, + 0.017003746703267097, + 0.015153524465858936, + 0.01617986336350441, + 0.018100235611200333, + 0.012690985575318336, + 0.01841863803565502, + 0.014142594300210476, + 0.010553300380706787, + 0.009394153952598572, + 0.010026681236922741, + 0.01292969286441803, + 0.007179506588727236, + 0.020023541525006294, + 0.008306227624416351, + 0.015606040135025978, + 0.01020774058997631, + 0.01568726822733879, + 0.015626009553670883, + 0.012065134011209011, + 0.010384919121861458, + 0.01714634895324707, + 0.014477411285042763, + 0.007865329273045063, + 0.005637509282678366, + 0.004916617181152105, + 0.0070083956234157085, + 0.005223910324275494, + 0.0050874133594334126, + 0.007024835329502821, + 0.007666058838367462, + 0.017429472878575325, + 0.0075119370594620705, + 0.008675836957991123, + 0.013362729921936989, + 0.01313334796577692, + 0.026269642636179924, + 0.01465584896504879, + 0.014853783883154392, + 0.005061339121311903, + 0.009254823438823223, + 0.005121500231325626, + 0.005110805854201317, + 0.0036317443009465933, + 0.004019089508801699, + 0.00332282786257565, + 0.006162137724459171, + 0.023608272895216942, + 0.032570671290159225, + 0.010231032967567444, + 0.01609964296221733, + 0.018092572689056396, + 0.019573330879211426, + 0.014037327840924263, + 0.013631084933876991, + 0.01372313592582941, + 0.010102292522788048, + 0.016100089997053146, + 0.010869299992918968, + 0.012078527361154556, + 0.017581084743142128, + 0.01853085309267044, + 0.019362667575478554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.004102030768990517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189646303653717 + ], + "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": [ + 0.10856378078460693, + 0.1285972148180008, + 0.0734739899635315, + 0.14210017025470734, + 0.11555561423301697, + 0.1229056566953659, + 0.09185169637203217, + 0.05416686832904816, + 0.05571885406970978, + 0.10706612467765808 + ], + "credit_map": [ + 0.014428390190005302, + 0.0111920814961195, + 0.014942674897611141, + 0.013692446984350681, + 0.007263391278684139, + 0.014966847375035286, + 0.011858637444674969, + 0.02021932043135166, + 0.015355913899838924, + 0.01332512404769659, + 0.017177406698465347, + 0.017989352345466614, + 0.015401524491608143, + 0.01754341833293438, + 0.01490221917629242, + 0.016902266070246696, + 0.00981932133436203, + 0.010231832042336464, + 0.010607516393065453, + 0.008727253414690495, + 0.008995437994599342, + 0.007545975502580404, + 0.008723310194909573, + 0.00882334541529417, + 0.01827998273074627, + 0.020244410261511803, + 0.02385612763464451, + 0.01563774049282074, + 0.016813041642308235, + 0.01204353291541338, + 0.019871143624186516, + 0.015354189090430737, + 0.0178076159209013, + 0.011581925675272942, + 0.01283907238394022, + 0.015856577083468437, + 0.01382623240351677, + 0.01241787988692522, + 0.017153501510620117, + 0.014072815887629986, + 0.015394228510558605, + 0.01605081744492054, + 0.01816665753722191, + 0.02033395506441593, + 0.014912987127900124, + 0.009450094774365425, + 0.01639523170888424, + 0.012201692909002304, + 0.015878604725003242, + 0.005579724442213774, + 0.007861598394811153, + 0.008634957484900951, + 0.008750976994633675, + 0.014697466045618057, + 0.013671863824129105, + 0.016776515170931816, + 0.007260048296302557, + 0.007038809359073639, + 0.008382687345147133, + 0.007555493153631687, + 0.004735259339213371, + 0.005939339753240347, + 0.00570089602842927, + 0.007554339710623026, + 0.0068092430010437965, + 0.009051823057234287, + 0.007203579414635897, + 0.006623114459216595, + 0.005343697499483824, + 0.006233097519725561, + 0.00718386797234416, + 0.007270435802638531, + 0.012593582272529602, + 0.009648061357438564, + 0.012560696341097355, + 0.016131669282913208, + 0.008411603048443794, + 0.015736788511276245, + 0.01456817053258419, + 0.017415564507246017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.018189646303653717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004407388041727245 + ], + "opposite_logprob_tokens": [ + -7.75044059753418 + ], + "logprob_reward": -0.0004407388041727245, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004407388041727245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11674360185861588, + 0.09836920350790024, + 0.05812273547053337, + 0.18019017577171326, + 0.0744018405675888, + 0.053130000829696655, + 0.072562575340271, + 0.07859475165605545, + 0.06086621806025505, + 0.20701885223388672 + ], + "credit_map": [ + 0.01642308011651039, + 0.013228924944996834, + 0.017176145687699318, + 0.01109888032078743, + 0.008051611483097076, + 0.012885231524705887, + 0.014374900609254837, + 0.02350483275949955, + 0.010169937275350094, + 0.009128456935286522, + 0.014831854961812496, + 0.013265039771795273, + 0.012268784455955029, + 0.01523149199783802, + 0.011614852584898472, + 0.01185878086835146, + 0.007846600376069546, + 0.00815581064671278, + 0.008102168329060078, + 0.004656215663999319, + 0.007942565716803074, + 0.007407442666590214, + 0.006531005259603262, + 0.007480928208678961, + 0.028198366984725, + 0.02339191362261772, + 0.027896979823708534, + 0.03041364625096321, + 0.024006903171539307, + 0.013981555588543415, + 0.013435984961688519, + 0.01886482536792755, + 0.011276443488895893, + 0.007417658809572458, + 0.008824874646961689, + 0.008653665892779827, + 0.00902744010090828, + 0.005787534173578024, + 0.01264051254838705, + 0.010773707181215286, + 0.008126594126224518, + 0.004894366953521967, + 0.0059576681815087795, + 0.007533328142017126, + 0.005617829505354166, + 0.0048514134250581264, + 0.007157968357205391, + 0.008990832604467869, + 0.010850176215171814, + 0.004556892439723015, + 0.007150446064770222, + 0.00989238265901804, + 0.008183949626982212, + 0.010722294449806213, + 0.011146130971610546, + 0.010060301050543785, + 0.012877749279141426, + 0.010810411535203457, + 0.01173689030110836, + 0.009219232015311718, + 0.006943050771951675, + 0.007136720698326826, + 0.006885354872792959, + 0.012985342182219028, + 0.007217261008918285, + 0.006900185253471136, + 0.005517145618796349, + 0.0062317391857504845, + 0.010260535404086113, + 0.007261369377374649, + 0.007990769110620022, + 0.009487210772931576, + 0.02488579787313938, + 0.016038594767451286, + 0.02721332013607025, + 0.019928067922592163, + 0.02049432508647442, + 0.0312824584543705, + 0.0349428616464138, + 0.032233428210020065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0004407388041727245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsettling, tilted, high-angle shot places the viewer in a scene almost engulfed in pure darkness. A strong, sharply directed beam of cold white light streams down from the upper and side of the frame, precisely illuminating an arthropod creature emerging from the deep shadows. The creature's deep black chitinous carapace is covered in irregular, bony knobs and protuberances. Its surface, glossy and smooth as if coated with oil, reflects sharp, glaring highlights from the single light source, while the rest of its body blends completely into the bottomless shadows. With a slow, deliberate rhythm, it slowly raises its front half, its long, jointed legs carefully adjusting their footing on the damp rocky ground, its front arching upward with a powerful force, as if performing some ancient ritual. Though slow, this movement is imbued with a menacing, poised momentum. The entire composition is filled with strong chiaroscuro (light and dark contrasts). The light sculpts every detail of the creature's silhouette, creating a huge contrast with the surrounding darkness, perfectly recreating Caravaggio's classical tension and creating a suffocating, silent atmosphere mixed with fear and classical beauty.", + "seed": 7236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_5/index_7397_seed_7236_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": -0.026828865247080103, + "vlm_logprob": -0.026828865247080103, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.026828865247080103, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06899150460958481 + ], + "opposite_logprob_tokens": [ + -3.1939914226531982 + ], + "logprob_reward": -0.06899150460958481, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06899150460958481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25296735763549805, + 0.1437711864709854, + 0.03023000806570053, + 0.28377705812454224, + 0.06374388933181763, + 0.027088802307844162, + 0.02214546874165535, + 0.026102788746356964, + 0.09211031347513199, + 0.05806306004524231 + ], + "credit_map": [ + 0.043348222970962524, + 0.026059864088892937, + 0.033284589648246765, + 0.0222130436450243, + 0.028974216431379318, + 0.034529659897089005, + 0.036759354174137115, + 0.027798444032669067, + 0.012684346176683903, + 0.006802611518651247, + 0.010248529724776745, + 0.011483598500490189, + 0.011300304904580116, + 0.03179827332496643, + 0.03050837107002735, + 0.02894514799118042, + 0.0028327959589660168, + 0.0029432715382426977, + 0.0036933887749910355, + 0.0038966776337474585, + 0.0020895437337458134, + 0.00444172415882349, + 0.007041103672236204, + 0.003291504690423608, + 0.006667790003120899, + 0.006656083278357983, + 0.004407140891999006, + 0.007716297172009945, + 0.0053053502924740314, + 0.12398427724838257, + 0.013768305070698261, + 0.11527185142040253, + 0.0055334074422717094, + 0.003739200532436371, + 0.005059409886598587, + 0.005564954597502947, + 0.008837517350912094, + 0.011326580308377743, + 0.01300527062267065, + 0.010677551850676537, + 0.0035886988043785095, + 0.0014421531232073903, + 0.0038857038598507643, + 0.002808427670970559, + 0.003782605519518256, + 0.002378651173785329, + 0.0047528743743896484, + 0.004449689760804176, + 0.001964556984603405, + 0.0015315692871809006, + 0.0029877666383981705, + 0.003504630411043763, + 0.002793440828099847, + 0.0031901563052088022, + 0.0030574966222047806, + 0.0031158532947301865, + 0.002444375306367874, + 0.0013793054968118668, + 0.003052370622754097, + 0.0025389338843524456, + 0.003426241921260953, + 0.00521830515936017, + 0.0044035762548446655, + 0.0036396824289113283, + 0.011280057020485401, + 0.008119473233819008, + 0.008134972304105759, + 0.013318726792931557, + 0.007432765793055296, + 0.01681436412036419, + 0.013733409345149994, + 0.013276548124849796, + 0.005094505380839109, + 0.004958653822541237, + 0.004587881732732058, + 0.004190271720290184, + 0.004420401994138956, + 0.008135472424328327, + 0.01375644002109766, + 0.012919431552290916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.06899150460958481 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637290023267269 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.02637290023267269, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637290023267269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13827727735042572, + 0.4111384451389313, + 0.06805432587862015, + 0.06064982712268829, + 0.02919352427124977, + 0.04146992042660713, + 0.03438039869070053, + 0.026586143299937248, + 0.1074904128909111, + 0.08275964856147766 + ], + "credit_map": [ + 0.024068566039204597, + 0.014875736087560654, + 0.017605358734726906, + 0.013119890354573727, + 0.013589431531727314, + 0.017173893749713898, + 0.022301817312836647, + 0.015542594715952873, + 0.05126526206731796, + 0.025245094671845436, + 0.029899848625063896, + 0.03492424264550209, + 0.030760662630200386, + 0.07224328815937042, + 0.10752954334020615, + 0.05927053838968277, + 0.007004124578088522, + 0.006306146737188101, + 0.00926375761628151, + 0.007101063150912523, + 0.004949852824211121, + 0.010583451949059963, + 0.01578180119395256, + 0.007064139004796743, + 0.004636796656996012, + 0.00447022495791316, + 0.005505013279616833, + 0.006265698932111263, + 0.004322735592722893, + 0.014636273495852947, + 0.007093382067978382, + 0.013719705864787102, + 0.0024383298587054014, + 0.001433741650544107, + 0.0036793630570173264, + 0.0018352445913478732, + 0.002980026649311185, + 0.004733700305223465, + 0.00644656689837575, + 0.005646553821861744, + 0.006263347342610359, + 0.0030609359964728355, + 0.0060972124338150024, + 0.005430152174085379, + 0.004573316778987646, + 0.005101362708956003, + 0.00643443688750267, + 0.004509160295128822, + 0.0037921385373920202, + 0.0020354820881038904, + 0.004093566443771124, + 0.005484926514327526, + 0.004052131436765194, + 0.0056196777150034904, + 0.0046102325432002544, + 0.0046922434121370316, + 0.0033021667040884495, + 0.0017736689187586308, + 0.003057264955714345, + 0.002779678674414754, + 0.0035897030029445887, + 0.003889282699674368, + 0.00470339460298419, + 0.003490985371172428, + 0.009792409837245941, + 0.009751029312610626, + 0.011912389658391476, + 0.01412865798920393, + 0.009598786011338234, + 0.01838034763932228, + 0.017322033643722534, + 0.01660476066172123, + 0.0069393436424434185, + 0.006037745159119368, + 0.007158041466027498, + 0.012906650081276894, + 0.004233199637383223, + 0.006559180561453104, + 0.020276421681046486, + 0.01864907145500183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02637290023267269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464076129719615 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464076129719615, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464076129719615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19332259893417358, + 0.07979541271924973, + 0.04383296146988869, + 0.10157521069049835, + 0.03734925761818886, + 0.029546966776251793, + 0.059019215404987335, + 0.043039921671152115, + 0.22348614037036896, + 0.18903231620788574 + ], + "credit_map": [ + 0.02284836955368519, + 0.012435353361070156, + 0.03476562723517418, + 0.011842245236039162, + 0.019865000620484352, + 0.027154959738254547, + 0.038416504859924316, + 0.025994524359703064, + 0.005423009395599365, + 0.0036779679358005524, + 0.00642343470826745, + 0.005333296488970518, + 0.006721437443047762, + 0.019872356206178665, + 0.014135444536805153, + 0.018208462744951248, + 0.003204927546903491, + 0.0035773152485489845, + 0.003937294706702232, + 0.0044966829009354115, + 0.003236170159652829, + 0.007374780718237162, + 0.012870093807578087, + 0.00513569638133049, + 0.005123509559780359, + 0.004479700233787298, + 0.012243141420185566, + 0.007698822766542435, + 0.005229816306382418, + 0.02975689247250557, + 0.009445048868656158, + 0.02759828232228756, + 0.003541176440194249, + 0.0021899836137890816, + 0.004033740144222975, + 0.0025732899084687233, + 0.007300402969121933, + 0.005469306837767363, + 0.00758875859901309, + 0.004652597941458225, + 0.0025095471646636724, + 0.001902346615679562, + 0.004065863322466612, + 0.0036805225536227226, + 0.004103473387658596, + 0.0038898943457752466, + 0.005235384218394756, + 0.004159937147051096, + 0.00529830064624548, + 0.004443950951099396, + 0.006405177991837263, + 0.008903213776648045, + 0.008299560286104679, + 0.010031469166278839, + 0.007989027537405491, + 0.0076485187746584415, + 0.005047723185271025, + 0.0027349365409463644, + 0.004341588821262121, + 0.00516888964921236, + 0.0058562238700687885, + 0.006241956260055304, + 0.008105100132524967, + 0.005543507635593414, + 0.020195960998535156, + 0.017071589827537537, + 0.022489847615361214, + 0.03772025927901268, + 0.015179269015789032, + 0.04288419336080551, + 0.034088656306266785, + 0.03385636955499649, + 0.015305078588426113, + 0.017650222405791283, + 0.014901147224009037, + 0.015576991252601147, + 0.018074003979563713, + 0.02173115499317646, + 0.04335792362689972, + 0.042435768991708755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.00464076129719615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369913622736931 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.03369913622736931, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369913622736931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20305941998958588, + 0.12465862929821014, + 0.09813201427459717, + 0.12484818696975708, + 0.10948607325553894, + 0.08657039701938629, + 0.06294449418783188, + 0.07646263390779495, + 0.029975680634379387, + 0.08386247605085373 + ], + "credit_map": [ + 0.03444704785943031, + 0.02208649553358555, + 0.027705198153853416, + 0.01820501871407032, + 0.02182709611952305, + 0.026398666203022003, + 0.028606010600924492, + 0.023783870041370392, + 0.012168195098638535, + 0.006331889424473047, + 0.011740569956600666, + 0.0072431801818311214, + 0.009972657077014446, + 0.028909994289278984, + 0.027586501091718674, + 0.020705625414848328, + 0.008660662919282913, + 0.006566750351339579, + 0.009208490140736103, + 0.0103795500472188, + 0.005231842864304781, + 0.020849138498306274, + 0.0204604621976614, + 0.016775112599134445, + 0.007900296710431576, + 0.005444792564958334, + 0.008063272573053837, + 0.009504118002951145, + 0.005652422551065683, + 0.0401560440659523, + 0.009310384280979633, + 0.038816846907138824, + 0.006547996308654547, + 0.0058726659044623375, + 0.009739231318235397, + 0.008079170249402523, + 0.008751040324568748, + 0.02649369090795517, + 0.014940733090043068, + 0.029061535373330116, + 0.006945195607841015, + 0.005680456757545471, + 0.016341468319296837, + 0.008445014245808125, + 0.009352573193609715, + 0.009812766686081886, + 0.014788434840738773, + 0.015204482711851597, + 0.00805090181529522, + 0.004452117253094912, + 0.007674127817153931, + 0.007251602131873369, + 0.004575977101922035, + 0.011587309651076794, + 0.010341841727495193, + 0.009010612964630127, + 0.006941060069948435, + 0.004938998259603977, + 0.009602121077477932, + 0.01068560965359211, + 0.010919211432337761, + 0.010027508251369, + 0.013045188039541245, + 0.010302928276360035, + 0.0031779343262314796, + 0.0020673321560025215, + 0.0037620458751916885, + 0.0038192542269825935, + 0.002614616183564067, + 0.00426391139626503, + 0.005627183709293604, + 0.004643400199711323, + 0.008837008848786354, + 0.009439661167562008, + 0.009917034767568111, + 0.011311103589832783, + 0.007217324338853359, + 0.01361994631588459, + 0.012079966254532337, + 0.011440422385931015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.03369913622736931 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004400238685775548 + ], + "opposite_logprob_tokens": [ + -7.7504401206970215 + ], + "logprob_reward": -0.0004400238685775548, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004400238685775548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14175504446029663, + 0.1556074321269989, + 0.032190147787332535, + 0.3007218837738037, + 0.03822790086269379, + 0.017887365072965622, + 0.03742404654622078, + 0.026708101853728294, + 0.22322601079940796, + 0.026252148672938347 + ], + "credit_map": [ + 0.024899495765566826, + 0.012398489750921726, + 0.018711725249886513, + 0.00946736615151167, + 0.016931593418121338, + 0.01802060566842556, + 0.023476617410779, + 0.01784912496805191, + 0.01898665353655815, + 0.008982028812170029, + 0.012334661558270454, + 0.014987013302743435, + 0.011105340905487537, + 0.02525220438838005, + 0.040452562272548676, + 0.02350695990025997, + 0.0040223486721515656, + 0.003353771986439824, + 0.0036890492774546146, + 0.004288442898541689, + 0.0026052386965602636, + 0.00489962100982666, + 0.006435710936784744, + 0.0028959631454199553, + 0.005618090275675058, + 0.006725863087922335, + 0.005365916993469, + 0.00764646427705884, + 0.005111884791404009, + 0.132808119058609, + 0.013614769093692303, + 0.12383075058460236, + 0.002101510763168335, + 0.0014984660083428025, + 0.002742316573858261, + 0.0021723592653870583, + 0.0035422800574451685, + 0.009768194518983364, + 0.005654138047248125, + 0.010748633183538914, + 0.0020097410306334496, + 0.0011941695120185614, + 0.002477556699886918, + 0.002237144159153104, + 0.002457479014992714, + 0.0016911699203774333, + 0.0030907050240784883, + 0.002729396801441908, + 0.0033247373066842556, + 0.0023470488376915455, + 0.00410444475710392, + 0.005084712989628315, + 0.004212391097098589, + 0.007208263501524925, + 0.00564124621450901, + 0.005501197651028633, + 0.003456626320257783, + 0.0014931600308045745, + 0.003035013098269701, + 0.0029399648774415255, + 0.0031047204975038767, + 0.003918569069355726, + 0.005337764509022236, + 0.003422281239181757, + 0.020946450531482697, + 0.019308608025312424, + 0.02190006896853447, + 0.03562753647565842, + 0.016565635800361633, + 0.04160826653242111, + 0.03398507460951805, + 0.03328436613082886, + 0.002900507301092148, + 0.002829695586115122, + 0.003241198370233178, + 0.00308672315441072, + 0.0030559347942471504, + 0.003725058864802122, + 0.0040582395158708096, + 0.0033547866623848677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.0004400238685775548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An unsettling, tilted, high-angle shot places the viewer in a scene almost engulfed in pure darkness. A strong, sharply directed beam of cold white light streams down from the upper and side of the frame, precisely illuminating an arthropod creature emerging from the deep shadows. The creature's deep black chitinous carapace is covered in irregular, bony knobs and protuberances. Its surface, glossy and smooth as if coated with oil, reflects sharp, glaring highlights from the single light source, while the rest of its body blends completely into the bottomless shadows. With a slow, deliberate rhythm, it slowly raises its front half, its long, jointed legs carefully adjusting their footing on the damp rocky ground, its front arching upward with a powerful force, as if performing some ancient ritual. Though slow, this movement is imbued with a menacing, poised momentum. The entire composition is filled with strong chiaroscuro (light and dark contrasts). The light sculpts every detail of the creature's silhouette, creating a huge contrast with the surrounding darkness, perfectly recreating Caravaggio's classical tension and creating a suffocating, silent atmosphere mixed with fear and classical beauty.", + "seed": 7237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_5/index_7397_seed_7237_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": -0.023802154674194755, + "vlm_logprob": -0.023802154674194755, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.023802154674194755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0971493199467659 + ], + "opposite_logprob_tokens": [ + -2.97214937210083 + ], + "logprob_reward": -0.0971493199467659, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0971493199467659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16664926707744598, + 0.04979134723544121, + 0.05855339765548706, + 0.051845669746398926, + 0.09447196871042252, + 0.08407711237668991, + 0.05345689505338669, + 0.04481584206223488, + 0.2940952777862549, + 0.10224317759275436 + ], + "credit_map": [ + 0.018519945442676544, + 0.029297325760126114, + 0.019258998334407806, + 0.01798383891582489, + 0.025584513321518898, + 0.021952828392386436, + 0.01721959002315998, + 0.016832225024700165, + 0.0042438805103302, + 0.009084762074053288, + 0.006120115052908659, + 0.004885248839855194, + 0.006744151469320059, + 0.0070887962356209755, + 0.004383940249681473, + 0.00724045280367136, + 0.0075958543457090855, + 0.007975275628268719, + 0.004588690120726824, + 0.004780002404004335, + 0.006847858894616365, + 0.008058863691985607, + 0.005723932757973671, + 0.012982918880879879, + 0.008977225981652737, + 0.006261380389332771, + 0.006480540614575148, + 0.005167930852621794, + 0.0053065768443048, + 0.008046031929552555, + 0.00653119059279561, + 0.0050747934728860855, + 0.009356614202260971, + 0.00623320834711194, + 0.02147948555648327, + 0.007588157430291176, + 0.01120598055422306, + 0.01472750585526228, + 0.008951201103627682, + 0.01492981892079115, + 0.008405433036386967, + 0.007001710589975119, + 0.0073681906796991825, + 0.008463352918624878, + 0.021405266597867012, + 0.011088448576629162, + 0.012865331955254078, + 0.007479377556592226, + 0.007632162421941757, + 0.00545245036482811, + 0.0049242256209254265, + 0.004612737335264683, + 0.006113752257078886, + 0.006750218570232391, + 0.007991249673068523, + 0.009980103932321072, + 0.0040147737599909306, + 0.0030409677419811487, + 0.004092241171747446, + 0.004607363138347864, + 0.0035723529290407896, + 0.006398423574864864, + 0.0088680824264884, + 0.010221635922789574, + 0.022644151002168655, + 0.026881594210863113, + 0.02873888611793518, + 0.042097024619579315, + 0.048250462859869, + 0.027819570153951645, + 0.06622625142335892, + 0.03143736347556114, + 0.016083113849163055, + 0.009980469942092896, + 0.01230145338922739, + 0.011591039597988129, + 0.012246919795870781, + 0.011026077903807163, + 0.01565285213291645, + 0.013361247256398201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0971493199467659 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007661127485334873 + ], + "opposite_logprob_tokens": [ + -4.88266134262085 + ], + "logprob_reward": -0.007661127485334873, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007661127485334873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10981632024049759, + 0.27663105726242065, + 0.19299212098121643, + 0.041564635932445526, + 0.0321989431977272, + 0.017185134813189507, + 0.07964849472045898, + 0.0414663664996624, + 0.14872463047504425, + 0.05977224186062813 + ], + "credit_map": [ + 0.010568014346063137, + 0.014010732062160969, + 0.012774650938808918, + 0.013434325344860554, + 0.016892313957214355, + 0.016183646395802498, + 0.013267476111650467, + 0.012685166671872139, + 0.04511592909693718, + 0.0728473886847496, + 0.02681293524801731, + 0.01539669744670391, + 0.033658020198345184, + 0.03820132464170456, + 0.015226797200739384, + 0.029371974989771843, + 0.029525386169552803, + 0.04451441392302513, + 0.01985248550772667, + 0.012974904850125313, + 0.017864659428596497, + 0.026993662118911743, + 0.011161545291543007, + 0.030105073004961014, + 0.008554854430258274, + 0.005996589548885822, + 0.0045270551927387714, + 0.004012411925941706, + 0.0036890991032123566, + 0.006291727069765329, + 0.004860143177211285, + 0.003632759442552924, + 0.0038218360859900713, + 0.0025488664396107197, + 0.005969788879156113, + 0.0036599456798285246, + 0.003297803457826376, + 0.0047952416352927685, + 0.003836682066321373, + 0.004268779419362545, + 0.0021560087334364653, + 0.0019529799465090036, + 0.001983341760933399, + 0.001781173050403595, + 0.0034302324056625366, + 0.001833347138017416, + 0.002571946009993553, + 0.0014761066995561123, + 0.01124302763491869, + 0.007085827179253101, + 0.008707514964044094, + 0.009014179930090904, + 0.00905501376837492, + 0.009670832194387913, + 0.012818398885428905, + 0.012053697369992733, + 0.0044630891643464565, + 0.003380544949322939, + 0.006761413533240557, + 0.005999736022204161, + 0.0034772823564708233, + 0.00566649716347456, + 0.006102803163230419, + 0.005614999681711197, + 0.010416490025818348, + 0.013931956142187119, + 0.015158125199377537, + 0.02206011489033699, + 0.02476651780307293, + 0.014477704651653767, + 0.03227357938885689, + 0.015640156343579292, + 0.007503170054405928, + 0.0032110672909766436, + 0.00611661234870553, + 0.006254148203879595, + 0.007247240748256445, + 0.007120735477656126, + 0.013963783159852028, + 0.008355491794645786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.007661127485334873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822227543219924 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822227543219924, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822227543219924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24797970056533813, + 0.13352105021476746, + 0.08291825652122498, + 0.1309434324502945, + 0.07226735353469849, + 0.039341848343610764, + 0.07806839793920517, + 0.08208781480789185, + 0.05209021642804146, + 0.08078192919492722 + ], + "credit_map": [ + 0.02668694220483303, + 0.036692939698696136, + 0.029945291578769684, + 0.028876904398202896, + 0.037176281213760376, + 0.03817083686590195, + 0.02319902926683426, + 0.027231475338339806, + 0.0201646089553833, + 0.033281225711107254, + 0.012963034212589264, + 0.008935341611504555, + 0.01833101361989975, + 0.017613666132092476, + 0.009022203274071217, + 0.013209942728281021, + 0.010522719472646713, + 0.013405375182628632, + 0.0077169411815702915, + 0.009302391670644283, + 0.010335491970181465, + 0.01100179273635149, + 0.008516114205121994, + 0.01211742777377367, + 0.02105528861284256, + 0.014073505997657776, + 0.013797902502119541, + 0.014878001064062119, + 0.020181357860565186, + 0.02005649171769619, + 0.01091426145285368, + 0.015986619517207146, + 0.009076421149075031, + 0.009869148023426533, + 0.008165881969034672, + 0.008985823951661587, + 0.010080469772219658, + 0.00958932377398014, + 0.00811549462378025, + 0.00838479120284319, + 0.004209971521049738, + 0.004416897427290678, + 0.0033839717507362366, + 0.004944460466504097, + 0.006262665148824453, + 0.004815778695046902, + 0.007187055889517069, + 0.004121049307286739, + 0.00770689407363534, + 0.008246822282671928, + 0.006423358805477619, + 0.0067156897857785225, + 0.013180511072278023, + 0.00931620318442583, + 0.013555854558944702, + 0.012923060916364193, + 0.007479134947061539, + 0.006416899152100086, + 0.012256558984518051, + 0.011886837892234325, + 0.008885519579052925, + 0.01131268497556448, + 0.011625745333731174, + 0.012224429287016392, + 0.005742477253079414, + 0.005262396298348904, + 0.004359326791018248, + 0.008104962296783924, + 0.007747023366391659, + 0.007053733337670565, + 0.007418020162731409, + 0.006402274593710899, + 0.01305165234953165, + 0.00571801234036684, + 0.008051790297031403, + 0.00896925013512373, + 0.012513693422079086, + 0.011488760821521282, + 0.012397028505802155, + 0.00859173946082592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.002822227543219924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107332669198513 + ], + "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": [ + 0.09605675935745239, + 0.056050918996334076, + 0.10558181256055832, + 0.10122169554233551, + 0.11233507841825485, + 0.26704490184783936, + 0.08955427259206772, + 0.057397231459617615, + 0.05198761075735092, + 0.06276965141296387 + ], + "credit_map": [ + 0.007579762954264879, + 0.013498576357960701, + 0.012170152738690376, + 0.011407552286982536, + 0.010207002982497215, + 0.01491324882954359, + 0.0135597949847579, + 0.012720680795609951, + 0.004563919734209776, + 0.006518388632684946, + 0.0055711399763822556, + 0.007253626361489296, + 0.006905163638293743, + 0.008426358923316002, + 0.007955202832818031, + 0.0088571235537529, + 0.010711601935327053, + 0.01113586314022541, + 0.010528349317610264, + 0.007665126584470272, + 0.013803557492792606, + 0.015824727714061737, + 0.00859925989061594, + 0.027313336730003357, + 0.01767653040587902, + 0.013401634059846401, + 0.010004336014389992, + 0.011913393624126911, + 0.012813101522624493, + 0.014487013220787048, + 0.012503051199018955, + 0.008422640152275562, + 0.020163439214229584, + 0.008524991571903229, + 0.020018842071294785, + 0.009568635374307632, + 0.012605312280356884, + 0.018132396042346954, + 0.008968405425548553, + 0.014353073202073574, + 0.027261875569820404, + 0.019539520144462585, + 0.023266689851880074, + 0.024083072319626808, + 0.07387390732765198, + 0.03607909381389618, + 0.040569499135017395, + 0.02237127348780632, + 0.011138060130178928, + 0.0077396659180521965, + 0.010840527713298798, + 0.010966329835355282, + 0.013770122081041336, + 0.010279507376253605, + 0.011868353933095932, + 0.012951713986694813, + 0.0067080422304570675, + 0.005069256294518709, + 0.004965754691511393, + 0.007109975907951593, + 0.010985155589878559, + 0.007316101808100939, + 0.0067775859497487545, + 0.008465362712740898, + 0.006345134694129229, + 0.0039579663425683975, + 0.004873034078627825, + 0.007466087583452463, + 0.007004999089986086, + 0.0047842469066381454, + 0.011473511345684528, + 0.006082640029489994, + 0.008290495723485947, + 0.0048135267570614815, + 0.006691965274512768, + 0.00773023534566164, + 0.0084443474188447, + 0.008032378740608692, + 0.011202672496438026, + 0.007564038969576359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.01107332669198513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle with a tilted composition?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a strong, sharply directed beam of cold white light streaming from the upper and side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central subject an arthropod creature with a deep black chitinous carapace covered in irregular, bony knobs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature slowly raising its front half with long, jointed legs adjusting on damp rocky ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit strong chiaroscuro with the creature's glossy surface reflecting sharp highlights against deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003047717036679387 + ], + "opposite_logprob_tokens": [ + -8.12530517578125 + ], + "logprob_reward": -0.0003047717036679387, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0003047717036679387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14430633187294006, + 0.12904728949069977, + 0.08884494006633759, + 0.05363275855779648, + 0.085411936044693, + 0.05927170813083649, + 0.09313013404607773, + 0.08884838968515396, + 0.12713882327079773, + 0.1303676962852478 + ], + "credit_map": [ + 0.016367828473448753, + 0.020470663905143738, + 0.014620711095631123, + 0.017006762325763702, + 0.02096487209200859, + 0.021074047312140465, + 0.017042318359017372, + 0.016759127378463745, + 0.02045736089348793, + 0.03018558770418167, + 0.012557041831314564, + 0.008808854967355728, + 0.014821574091911316, + 0.016673831269145012, + 0.01258491724729538, + 0.01295812614262104, + 0.013186058960855007, + 0.016909325495362282, + 0.006663402076810598, + 0.0074232290498912334, + 0.011476965621113777, + 0.012182015925645828, + 0.008534601889550686, + 0.012469333596527576, + 0.010061494074761868, + 0.0055782487615942955, + 0.006801773328334093, + 0.005096536595374346, + 0.005978500936180353, + 0.00844384916126728, + 0.0065946755930781364, + 0.005077680107206106, + 0.016624273732304573, + 0.008234125562012196, + 0.013283533044159412, + 0.007639187853783369, + 0.00951386895030737, + 0.013646158389747143, + 0.00772458640858531, + 0.008746202103793621, + 0.007621945813298225, + 0.006168984342366457, + 0.004948023706674576, + 0.005345007870346308, + 0.014382735826075077, + 0.008006072603166103, + 0.007291649933904409, + 0.005507289431989193, + 0.014065139926970005, + 0.008182475343346596, + 0.009216838516294956, + 0.008417400531470776, + 0.013991707935929298, + 0.011760655790567398, + 0.014486744999885559, + 0.013009173795580864, + 0.00979455653578043, + 0.008456025272607803, + 0.012072763405740261, + 0.011112547479569912, + 0.01058141142129898, + 0.01170356385409832, + 0.014070335775613785, + 0.011057190597057343, + 0.010324359871447086, + 0.010277505032718182, + 0.010853776708245277, + 0.016541576012969017, + 0.022719483822584152, + 0.01279861107468605, + 0.029482023790478706, + 0.014141479507088661, + 0.018430262804031372, + 0.011880901642143726, + 0.01627231575548649, + 0.014994018711149693, + 0.01467140018939972, + 0.01512944046407938, + 0.020565524697303772, + 0.018423832952976227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0003047717036679387 + } + ], + "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/0000155/group_6/index_tensor([3583, 3583])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_tensor([3583, 3583])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a3d759d3db0fc1a60bcadeca031d00f8d62afa1a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_tensor([3583, 3583])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, tight close-up immerses the viewer in a fierce battle between two leopard lizards on a dark, wet metal floor. Their bumpy, alien-like, banded skin, a detailed mosaic of light yellow and dark spots, shimmers and reflects the scene in distorted ripples. The only illumination comes from the relentless, out-of-focus glow of futuristic neon signs in the background, which cast ever-shifting, high-contrast electric blues and vibrant fuchsias across their writhing bodies. One lizard viciously bites at the other's flank, jaws clenched, while the defender struggles desperately, its spindly limbs and tail slapping against the wet surface, sending tiny splashes of water flying. Their tiny claws scrape frantically against the cold steel as they struggle for purchase, their desperate movements and the overpowering synthetic light combining to create a raw, tense, and distinctly dystopian sci-fi atmosphere.", + "seed": 14047, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_3583_seed_14047_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": -0.06047742832452059, + "vlm_logprob": -0.06047742832452059, + "vlm_yes_no_margin": 3.375000023841858, + "vlm_token_credit": -0.06047742832452059, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10279673337936401 + ], + "opposite_logprob_tokens": [ + -2.352796792984009 + ], + "logprob_reward": -0.10279673337936401, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10279673337936401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15961100161075592, + 0.08809500187635422, + 0.09238476306200027, + 0.09544822573661804, + 0.09652499854564667, + 0.09560389816761017, + 0.12749508023262024, + 0.11897816509008408, + 0.047933150082826614, + 0.07792568951845169 + ], + "credit_map": [ + 0.0159915741533041, + 0.023774512112140656, + 0.01628410629928112, + 0.02308657579123974, + 0.024690108373761177, + 0.015037722885608673, + 0.024034660309553146, + 0.016711745411157608, + 0.00924710649996996, + 0.007590813562273979, + 0.012375901453197002, + 0.010809505358338356, + 0.010631708428263664, + 0.01229447964578867, + 0.01384107768535614, + 0.011304414831101894, + 0.010341422632336617, + 0.01205403357744217, + 0.010746476240456104, + 0.011482971720397472, + 0.010420742444694042, + 0.010642334818840027, + 0.0110786659643054, + 0.015618130564689636, + 0.0161198191344738, + 0.009054677560925484, + 0.011875210329890251, + 0.011494842357933521, + 0.009348643943667412, + 0.019859816879034042, + 0.007538017351180315, + 0.010157204233109951, + 0.006715003401041031, + 0.0079315435141325, + 0.006663855165243149, + 0.009283486753702164, + 0.009691497310996056, + 0.027250273153185844, + 0.008067894726991653, + 0.02092145010828972, + 0.007688082288950682, + 0.008715610019862652, + 0.009080542251467705, + 0.005981628317385912, + 0.007793912198394537, + 0.014700062572956085, + 0.0263474490493536, + 0.015296622179448605, + 0.011379046365618706, + 0.013606525957584381, + 0.018108809366822243, + 0.014619884081184864, + 0.009040027856826782, + 0.027894573286175728, + 0.015419894829392433, + 0.017426330596208572, + 0.009454941377043724, + 0.0052870227955281734, + 0.006036453880369663, + 0.006748364772647619, + 0.007891986519098282, + 0.032099660485982895, + 0.012567900121212006, + 0.03889184817671776, + 0.005909504368901253, + 0.004685561638325453, + 0.007584541570395231, + 0.0077798631973564625, + 0.003728881012648344, + 0.005203952081501484, + 0.006206986960023642, + 0.006833861581981182, + 0.007599478121846914, + 0.008119712583720684, + 0.011963441967964172, + 0.012179824523627758, + 0.006173557136207819, + 0.009558865800499916, + 0.011421128176152706, + 0.010909684002399445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.10279673337936401 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359664276242256 + ], + "opposite_logprob_tokens": [ + -3.6513595581054688 + ], + "logprob_reward": -0.026359664276242256, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359664276242256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18149502575397491, + 0.06277301907539368, + 0.063997283577919, + 0.05514155328273773, + 0.050555747002363205, + 0.11107538640499115, + 0.23199157416820526, + 0.10078831017017365, + 0.049074504524469376, + 0.09310761839151382 + ], + "credit_map": [ + 0.02377048321068287, + 0.014393758960068226, + 0.036105334758758545, + 0.02602665312588215, + 0.012409777380526066, + 0.014623787254095078, + 0.024114398285746574, + 0.030050817877054214, + 0.00679814862087369, + 0.006218975409865379, + 0.01339141745120287, + 0.009615234099328518, + 0.009342457167804241, + 0.006410432979464531, + 0.004678658675402403, + 0.0063176932744681835, + 0.006702111568301916, + 0.007548307068645954, + 0.007882424630224705, + 0.008341710083186626, + 0.009271090850234032, + 0.007049282547086477, + 0.0067632910795509815, + 0.010439068078994751, + 0.006592411547899246, + 0.004331398755311966, + 0.007404011208564043, + 0.0032790089026093483, + 0.007332137320190668, + 0.012171671725809574, + 0.006248220335692167, + 0.007782689295709133, + 0.0043971654959023, + 0.005811416078358889, + 0.004272253718227148, + 0.007125428412109613, + 0.006401869934052229, + 0.009937294758856297, + 0.0068707941100001335, + 0.0057395282201468945, + 0.005389440339058638, + 0.006387076806277037, + 0.00808128621429205, + 0.006511795800179243, + 0.007512085139751434, + 0.012475595809519291, + 0.04433251917362213, + 0.020385587587952614, + 0.023679455742239952, + 0.030337410047650337, + 0.028395112603902817, + 0.026455141603946686, + 0.02566368319094181, + 0.030860234051942825, + 0.034548792988061905, + 0.03205176815390587, + 0.007527574431151152, + 0.005225207656621933, + 0.005520202685147524, + 0.005234687589108944, + 0.007137073669582605, + 0.037499021738767624, + 0.009570824913680553, + 0.0230737142264843, + 0.006191683001816273, + 0.005477637052536011, + 0.008342772722244263, + 0.005823157727718353, + 0.004468254279345274, + 0.0068123419769108295, + 0.006311084609478712, + 0.005647572688758373, + 0.009152116253972054, + 0.0077935936860740185, + 0.015881389379501343, + 0.014729464426636696, + 0.0076750339940190315, + 0.007664664648473263, + 0.01785397343337536, + 0.012357380241155624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.026359664276242256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349330320954323 + ], + "opposite_logprob_tokens": [ + -3.6513493061065674 + ], + "logprob_reward": -0.026349330320954323, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349330320954323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16502295434474945, + 0.043998539447784424, + 0.04981336370110512, + 0.18399544060230255, + 0.1201491728425026, + 0.12122085690498352, + 0.06162736564874649, + 0.14195962250232697, + 0.03583051264286041, + 0.07638219743967056 + ], + "credit_map": [ + 0.019235974177718163, + 0.01949978992342949, + 0.017972256988286972, + 0.02347293309867382, + 0.017948823049664497, + 0.02069547213613987, + 0.02244812622666359, + 0.023749586194753647, + 0.004192442167550325, + 0.004735129419714212, + 0.005002617835998535, + 0.0050988104194402695, + 0.009435058571398258, + 0.005709381308406591, + 0.004542892798781395, + 0.005282212048768997, + 0.004308235831558704, + 0.00492594949901104, + 0.007702942471951246, + 0.007140562403947115, + 0.008383054286241531, + 0.005584107246249914, + 0.00505007803440094, + 0.006718433927744627, + 0.01625973917543888, + 0.015784932300448418, + 0.03562252223491669, + 0.02367381565272808, + 0.02288598008453846, + 0.025748539716005325, + 0.021222151815891266, + 0.022797759622335434, + 0.006709105800837278, + 0.01165622379630804, + 0.009035224094986916, + 0.014489586465060711, + 0.010053957812488079, + 0.026467720046639442, + 0.014519312418997288, + 0.027218030765652657, + 0.0052223303355276585, + 0.006902586203068495, + 0.010346921160817146, + 0.0077359117567539215, + 0.005179785192012787, + 0.025118280202150345, + 0.04822922497987747, + 0.012485822662711143, + 0.0061125620268285275, + 0.006751521956175566, + 0.009126017801463604, + 0.005923651158809662, + 0.006783261429518461, + 0.011313089169561863, + 0.007131482940167189, + 0.008485779166221619, + 0.00829192902892828, + 0.005448800045996904, + 0.006522781681269407, + 0.006939241662621498, + 0.00800093449652195, + 0.04949311912059784, + 0.011790426447987556, + 0.0454723946750164, + 0.005129206459969282, + 0.00833084899932146, + 0.003599090501666069, + 0.00556730292737484, + 0.0036284029483795166, + 0.002610491355881095, + 0.004378167446702719, + 0.002587000373750925, + 0.007128329947590828, + 0.0066785323433578014, + 0.014951401390135288, + 0.011708540841937065, + 0.00716537656262517, + 0.006110230926424265, + 0.008545216172933578, + 0.014094567857682705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.026349330320954323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094076342880726 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094076342880726, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094076342880726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12009884417057037, + 0.10000760853290558, + 0.13660083711147308, + 0.03957276791334152, + 0.0253599900752306, + 0.10341955721378326, + 0.06866313517093658, + 0.3351573646068573, + 0.023386716842651367, + 0.047733232378959656 + ], + "credit_map": [ + 0.01379034761339426, + 0.019060881808400154, + 0.01114574447274208, + 0.016080452129244804, + 0.01524152047932148, + 0.012087934650480747, + 0.012800557538866997, + 0.019891394302248955, + 0.010316482745110989, + 0.009700694121420383, + 0.010870647616684437, + 0.013062261044979095, + 0.015812180936336517, + 0.015090467408299446, + 0.01297613512724638, + 0.012178732082247734, + 0.018906377255916595, + 0.026768924668431282, + 0.01790621504187584, + 0.015142625197768211, + 0.016893278807401657, + 0.013323662802577019, + 0.010901582427322865, + 0.0167581494897604, + 0.003821415128186345, + 0.005192428827285767, + 0.005382640287280083, + 0.006013953126966953, + 0.00545828090980649, + 0.0049604736268520355, + 0.004023728892207146, + 0.004719844087958336, + 0.002605872694402933, + 0.0026148930191993713, + 0.0021884706802666187, + 0.0025985115207731724, + 0.0022506106179207563, + 0.0055304053239524364, + 0.0021480524446815252, + 0.005423170048743486, + 0.005258710123598576, + 0.005063849035650492, + 0.007655961439013481, + 0.006512152496725321, + 0.007623202633112669, + 0.012809203937649727, + 0.0414966382086277, + 0.016999829560518265, + 0.005514545366168022, + 0.0074411677196621895, + 0.007500975858420134, + 0.008033942431211472, + 0.009202522225677967, + 0.00939413346350193, + 0.01054102648049593, + 0.011034813709557056, + 0.01752212643623352, + 0.008819677866995335, + 0.008082038722932339, + 0.01130107045173645, + 0.016671134158968925, + 0.1290581375360489, + 0.02703462354838848, + 0.11666853725910187, + 0.003055959939956665, + 0.0033638766035437584, + 0.004083470441401005, + 0.002833729609847069, + 0.002387742046266794, + 0.0022288181353360415, + 0.0030005075968801975, + 0.002432609908282757, + 0.004591191653162241, + 0.0036801148671656847, + 0.007969444617629051, + 0.00564807653427124, + 0.006397123448550701, + 0.0058708288706839085, + 0.004990356974303722, + 0.008586092852056026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.004094076342880726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278733730316162 + ], + "opposite_logprob_tokens": [ + -2.017787456512451 + ], + "logprob_reward": -0.14278733730316162, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14278733730316162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07896890491247177, + 0.04341692104935646, + 0.09862768650054932, + 0.03635560721158981, + 0.034098733216524124, + 0.22191236913204193, + 0.13851022720336914, + 0.1151963323354721, + 0.04412909597158432, + 0.1887841373682022 + ], + "credit_map": [ + 0.008150704205036163, + 0.011913510039448738, + 0.008382237516343594, + 0.010530718602240086, + 0.009594939649105072, + 0.006502613890916109, + 0.013248593546450138, + 0.010645589791238308, + 0.005142432637512684, + 0.004078565165400505, + 0.006056312471628189, + 0.004612122196704149, + 0.006069480441510677, + 0.006421579979360104, + 0.006233926862478256, + 0.004802502226084471, + 0.008696740493178368, + 0.010101237334311008, + 0.014039370231330395, + 0.012963764369487762, + 0.01546631008386612, + 0.010330205783247948, + 0.011671836487948895, + 0.015358217060565948, + 0.005038361996412277, + 0.003575695678591728, + 0.0036540378350764513, + 0.006040893495082855, + 0.003778208279982209, + 0.0065202550031244755, + 0.004238645546138287, + 0.00350951193831861, + 0.0028902231715619564, + 0.003589214291423559, + 0.0033273850567638874, + 0.003648975398391485, + 0.004428401123732328, + 0.007934588007628918, + 0.003378375433385372, + 0.0049015735276043415, + 0.01019257865846157, + 0.013553624041378498, + 0.015566179528832436, + 0.011095288209617138, + 0.013325304724276066, + 0.037180352956056595, + 0.08523949235677719, + 0.03575954586267471, + 0.012711511924862862, + 0.014557724818587303, + 0.015071147121489048, + 0.01706843450665474, + 0.022742610424757004, + 0.018493058159947395, + 0.022443197667598724, + 0.015422535128891468, + 0.00650629261508584, + 0.0047669303603470325, + 0.006760729476809502, + 0.006906772032380104, + 0.006187780760228634, + 0.04121517762541771, + 0.010303501971065998, + 0.03254915028810501, + 0.005395607091486454, + 0.007542212028056383, + 0.011786319315433502, + 0.0063152676448225975, + 0.003472157521173358, + 0.002407051157206297, + 0.003878111019730568, + 0.0033323706593364477, + 0.018217317759990692, + 0.014979907311499119, + 0.0276322178542614, + 0.02638292871415615, + 0.024567093700170517, + 0.022760165855288506, + 0.02968880720436573, + 0.024555709213018417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.14278733730316162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up immerses the viewer in a fierce battle between two leopard lizards on a dark, wet metal floor. Their bumpy, alien-like, banded skin, a detailed mosaic of light yellow and dark spots, shimmers and reflects the scene in distorted ripples. The only illumination comes from the relentless, out-of-focus glow of futuristic neon signs in the background, which cast ever-shifting, high-contrast electric blues and vibrant fuchsias across their writhing bodies. One lizard viciously bites at the other's flank, jaws clenched, while the defender struggles desperately, its spindly limbs and tail slapping against the wet surface, sending tiny splashes of water flying. Their tiny claws scrape frantically against the cold steel as they struggle for purchase, their desperate movements and the overpowering synthetic light combining to create a raw, tense, and distinctly dystopian sci-fi atmosphere.", + "seed": 14048, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_3583_seed_14048_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": -0.043095900770276786, + "vlm_logprob": -0.043095900770276786, + "vlm_yes_no_margin": 3.474999952316284, + "vlm_token_credit": -0.043095900770276786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08144751191139221 + ], + "opposite_logprob_tokens": [ + -2.5814476013183594 + ], + "logprob_reward": -0.08144751191139221, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08144751191139221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06207277625799179, + 0.0035515916533768177, + 0.002139432355761528, + 0.0072181252762675285, + 0.023254496976733208, + 0.0047777085565030575, + 0.002735465532168746, + 0.8907846808433533, + 0.0018696724437177181, + 0.001595994457602501 + ], + "credit_map": [ + 0.006935727782547474, + 0.007603714242577553, + 0.009728144854307175, + 0.004219752736389637, + 0.008324967697262764, + 0.012171966023743153, + 0.005425130482763052, + 0.00766337662935257, + 0.0003136297455057502, + 0.0004066552792210132, + 0.000496041146107018, + 0.0004685716121457517, + 0.00034186162520200014, + 0.000320209888741374, + 0.0007378777954727411, + 0.00046674456098116934, + 0.0003297261137049645, + 0.00022317971161101013, + 0.00027759282966144383, + 0.00023109288304112852, + 0.00027421157574281096, + 0.0003438781131990254, + 0.00020908594888169318, + 0.00025066526723094285, + 0.000783044786658138, + 0.0012724397238343954, + 0.0011926793958991766, + 0.0008697757148183882, + 0.0011597468983381987, + 0.0005570609355345368, + 0.0007148900185711682, + 0.0006684877444058657, + 0.00150717340875417, + 0.0007413031999021769, + 0.00851010624319315, + 0.001353269792161882, + 0.0007267273613251746, + 0.0008628273499198258, + 0.0011595557443797588, + 0.00839353445917368, + 0.000492411432787776, + 0.0004190745821688324, + 0.0010630112374201417, + 0.0011592325754463673, + 0.00031755282543599606, + 0.0005892355111427605, + 0.0002689899120014161, + 0.0004682005092035979, + 0.0002805678523145616, + 0.00038724884507246315, + 0.00029240481671877205, + 0.0003683333925437182, + 0.00041366874938830733, + 0.00033566189813427627, + 0.00036685020313598216, + 0.0002907299203798175, + 0.050393182784318924, + 0.014227552339434624, + 0.30593574047088623, + 0.05422140657901764, + 0.03800643980503082, + 0.025482581928372383, + 0.3393099308013916, + 0.0632079467177391, + 0.00013306459004525095, + 0.00022760778665542603, + 0.00024021034187171608, + 0.0002437246439512819, + 0.0002987004991155118, + 0.0002073160867439583, + 0.00029172818176448345, + 0.000227320400881581, + 0.00017104715516325086, + 0.00026835312019102275, + 0.0002643718325998634, + 0.00019217163207940757, + 0.000211157719604671, + 0.00016593550390098244, + 0.00015393750800285488, + 0.0001690200442681089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.08144751191139221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495835840702057 + ], + "opposite_logprob_tokens": [ + -2.9299583435058594 + ], + "logprob_reward": -0.05495835840702057, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495835840702057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29411667585372925, + 0.035313546657562256, + 0.02426070347428322, + 0.016346488147974014, + 0.07474381476640701, + 0.01721697486937046, + 0.033538416028022766, + 0.43343326449394226, + 0.04014817625284195, + 0.03088190406560898 + ], + "credit_map": [ + 0.03257258981466293, + 0.033393505960702896, + 0.04361695796251297, + 0.018847132101655006, + 0.04361805319786072, + 0.05895838513970375, + 0.028600722551345825, + 0.03450934216380119, + 0.0033234357833862305, + 0.003731976728886366, + 0.004524361342191696, + 0.005450660828500986, + 0.004379519261419773, + 0.0040395366959273815, + 0.004866916686296463, + 0.004997140262275934, + 0.005209057126194239, + 0.0021381136029958725, + 0.002610238268971443, + 0.003002764657139778, + 0.002318419748917222, + 0.0029687276110053062, + 0.002611556090414524, + 0.003401825437322259, + 0.0023739805910736322, + 0.00224143429659307, + 0.002110386034473777, + 0.0022082438226789236, + 0.0023009502328932285, + 0.0019493730505928397, + 0.001608778373338282, + 0.001553341280668974, + 0.005299374461174011, + 0.0022326665930449963, + 0.028356922790408134, + 0.004655108787119389, + 0.002512908773496747, + 0.002560994355008006, + 0.005396416876465082, + 0.02372942492365837, + 0.0027792174369096756, + 0.0015473203966394067, + 0.0037249517627060413, + 0.0014232085086405277, + 0.0011436532950028777, + 0.002256872598081827, + 0.001193088828586042, + 0.0031486612278968096, + 0.0037581773940473795, + 0.0046877204440534115, + 0.0037603601813316345, + 0.005294605623930693, + 0.0038624603766947985, + 0.0034302733838558197, + 0.005206112749874592, + 0.003538707736879587, + 0.02458101511001587, + 0.007012542802840471, + 0.14872591197490692, + 0.0262713972479105, + 0.018447531387209892, + 0.012803426943719387, + 0.16472750902175903, + 0.030863920226693153, + 0.0057891299948096275, + 0.0038666867185384035, + 0.005045406520366669, + 0.006640145089477301, + 0.004153685178607702, + 0.003572775749489665, + 0.005327500402927399, + 0.005752848461270332, + 0.003506470238789916, + 0.002780476352199912, + 0.005181317683309317, + 0.00431599048897624, + 0.0033403660636395216, + 0.0037201405502855778, + 0.0045686098746955395, + 0.0034685314167290926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.05495835840702057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184378743171692 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.018184378743171692, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018184378743171692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10355588048696518, + 0.0065296548418700695, + 0.002491921652108431, + 0.018256433308124542, + 0.03721679374575615, + 0.008986715227365494, + 0.010528395883738995, + 0.7992607355117798, + 0.006442938465625048, + 0.00673052528873086 + ], + "credit_map": [ + 0.01179550401866436, + 0.012561717070639133, + 0.0156839769333601, + 0.006642615422606468, + 0.014027412980794907, + 0.021177494898438454, + 0.009054918773472309, + 0.01261222455650568, + 0.0005398617358878255, + 0.0008968421025201678, + 0.0008958398248068988, + 0.0008557669934816658, + 0.000967322732321918, + 0.0006512482068501413, + 0.0008394519099965692, + 0.0008833211613819003, + 0.0004972384194843471, + 0.00022624252596870065, + 0.00030184537172317505, + 0.0002199892478529364, + 0.0002858437364920974, + 0.0003230761212762445, + 0.0002840751549229026, + 0.0003536106669344008, + 0.0015398630639538169, + 0.002099876757711172, + 0.0036002839915454388, + 0.0038144381251186132, + 0.0025229905731976032, + 0.00139775057323277, + 0.0015840594423934817, + 0.0016971698496490717, + 0.0023419607896357775, + 0.0010410243412479758, + 0.013685153797268867, + 0.0021615808364003897, + 0.0011305531952530146, + 0.0013562784297391772, + 0.001846357248723507, + 0.013653880916535854, + 0.0012063798494637012, + 0.0007469787378795445, + 0.0020334222353994846, + 0.0020192882511764765, + 0.0005102111608721316, + 0.0011662269243970513, + 0.0005027843872085214, + 0.0008014223421923816, + 0.0012959414161741734, + 0.0014963007997721434, + 0.0013038002653047442, + 0.0015053290408104658, + 0.001290474901907146, + 0.001089893514290452, + 0.0013915871968492866, + 0.0011550674680620432, + 0.04524077847599983, + 0.012751521542668343, + 0.27464911341667175, + 0.048551253974437714, + 0.03407800570130348, + 0.02280372753739357, + 0.30448096990585327, + 0.056705351918935776, + 0.0005796891055069864, + 0.0008625767659395933, + 0.0007693950901739299, + 0.0010091528529301286, + 0.000987177831120789, + 0.0007642883574590087, + 0.0006543918862007558, + 0.0008162661106325686, + 0.0008126721368171275, + 0.0014986391179263592, + 0.0005579120479524136, + 0.0004911237047053874, + 0.0013057091273367405, + 0.0007823205669410527, + 0.0005672735860571265, + 0.0007148738368414342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.018184378743171692 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946918856352568 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005946918856352568, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946918856352568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06034550070762634, + 0.026622124016284943, + 0.0317409448325634, + 0.03648332506418228, + 0.3131844401359558, + 0.025171393528580666, + 0.02570696733891964, + 0.3939419388771057, + 0.030636826530098915, + 0.05616651475429535 + ], + "credit_map": [ + 0.005774560384452343, + 0.009948087856173515, + 0.010543004609644413, + 0.0047889528796076775, + 0.009213877841830254, + 0.005161021836102009, + 0.00657475832849741, + 0.0083412304520607, + 0.0027195182628929615, + 0.002893787110224366, + 0.003659687703475356, + 0.003313731402158737, + 0.0023245010524988174, + 0.0019512661965563893, + 0.005432886071503162, + 0.0043267481960356236, + 0.0060735903680324554, + 0.0027411975897848606, + 0.004009914584457874, + 0.002399756573140621, + 0.003562257857993245, + 0.004802288021892309, + 0.0038948252331465483, + 0.0042571136727929115, + 0.0038039525970816612, + 0.004069863352924585, + 0.00605452386662364, + 0.005086295306682587, + 0.005963784176856279, + 0.0038791545666754246, + 0.0036878560204058886, + 0.003937895409762859, + 0.020630042999982834, + 0.009154767729341984, + 0.11624713242053986, + 0.01817128248512745, + 0.00942767970263958, + 0.012005913071334362, + 0.01601167768239975, + 0.1115359216928482, + 0.002849995158612728, + 0.0016267463797703385, + 0.008130020461976528, + 0.004667096305638552, + 0.0015970737440511584, + 0.002787237986922264, + 0.0013525631511583924, + 0.002160661155357957, + 0.00266758119687438, + 0.003977170679718256, + 0.0028320224955677986, + 0.0037212257739156485, + 0.0035223669838160276, + 0.002642398001626134, + 0.004060415551066399, + 0.00228378688916564, + 0.02243032306432724, + 0.006292041391134262, + 0.13534025847911835, + 0.023851808160543442, + 0.016848446801304817, + 0.011441171169281006, + 0.14970961213111877, + 0.02802831307053566, + 0.003900738200172782, + 0.0030044219456613064, + 0.0037446317728608847, + 0.0066060093231499195, + 0.0032073380425572395, + 0.0031551842112094164, + 0.00266438745893538, + 0.004354116041213274, + 0.008309685625135899, + 0.008848190307617188, + 0.00834044348448515, + 0.004919350612908602, + 0.006743958219885826, + 0.008032032288610935, + 0.004876831546425819, + 0.006096018943935633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.005946918856352568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054942335933446884 + ], + "opposite_logprob_tokens": [ + -2.9299423694610596 + ], + "logprob_reward": -0.054942335933446884, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054942335933446884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18037785589694977, + 0.0381707027554512, + 0.05506810173392296, + 0.06446892023086548, + 0.019255895167589188, + 0.02600274793803692, + 0.04845573380589485, + 0.505650520324707, + 0.02225089631974697, + 0.04029868543148041 + ], + "credit_map": [ + 0.018909815698862076, + 0.021421100944280624, + 0.027682675048708916, + 0.01306265126913786, + 0.025174645707011223, + 0.03394762799143791, + 0.017574593424797058, + 0.02260473370552063, + 0.003984325099736452, + 0.0036644984502345324, + 0.003859360469505191, + 0.008407756686210632, + 0.005860159173607826, + 0.002788311103358865, + 0.004250690340995789, + 0.0053556025959551334, + 0.007097144611179829, + 0.006327224895358086, + 0.007184304296970367, + 0.007756366394460201, + 0.008009771816432476, + 0.006060014013200998, + 0.0057225157506763935, + 0.006910762283951044, + 0.008159034885466099, + 0.008165272884070873, + 0.010343778878450394, + 0.009219732135534286, + 0.008598658256232738, + 0.005694168154150248, + 0.0070579624734818935, + 0.007230311632156372, + 0.002179298782721162, + 0.0029853100422769785, + 0.001522761071100831, + 0.002193779218941927, + 0.002049506176263094, + 0.0019847285002470016, + 0.00244923890568316, + 0.003891273168846965, + 0.003650614293292165, + 0.0025844434276223183, + 0.004483481869101524, + 0.004098501987755299, + 0.001647501834668219, + 0.0020756057929247618, + 0.002385709434747696, + 0.005076889414340258, + 0.004475515801459551, + 0.0060496823862195015, + 0.005358082242310047, + 0.009323513135313988, + 0.005500520579516888, + 0.005656866822391748, + 0.007374878041446209, + 0.00471667293459177, + 0.028393659740686417, + 0.008490517735481262, + 0.17263606190681458, + 0.031056927517056465, + 0.02185639552772045, + 0.014332487247884274, + 0.19265887141227722, + 0.03622560203075409, + 0.0025117869954556227, + 0.002222164999693632, + 0.0029115667566657066, + 0.004098296631127596, + 0.0027960818260908127, + 0.002021330874413252, + 0.0033410012256354094, + 0.0023486665450036526, + 0.003801418701186776, + 0.007814088836312294, + 0.004497556481510401, + 0.003049066523090005, + 0.007862587459385395, + 0.004922476131469011, + 0.003793503623455763, + 0.004557989537715912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.054942335933446884 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up immerses the viewer in a fierce battle between two leopard lizards on a dark, wet metal floor. Their bumpy, alien-like, banded skin, a detailed mosaic of light yellow and dark spots, shimmers and reflects the scene in distorted ripples. The only illumination comes from the relentless, out-of-focus glow of futuristic neon signs in the background, which cast ever-shifting, high-contrast electric blues and vibrant fuchsias across their writhing bodies. One lizard viciously bites at the other's flank, jaws clenched, while the defender struggles desperately, its spindly limbs and tail slapping against the wet surface, sending tiny splashes of water flying. Their tiny claws scrape frantically against the cold steel as they struggle for purchase, their desperate movements and the overpowering synthetic light combining to create a raw, tense, and distinctly dystopian sci-fi atmosphere.", + "seed": 14049, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_3583_seed_14049_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": -0.058973617199808356, + "vlm_logprob": -0.058973617199808356, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.058973617199808356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07253359258174896 + ], + "opposite_logprob_tokens": [ + -2.69753360748291 + ], + "logprob_reward": -0.07253359258174896, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07253359258174896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04046278074383736, + 0.06472443789243698, + 0.08870561420917511, + 0.012835711240768433, + 0.20004568994045258, + 0.04436928406357765, + 0.06707847118377686, + 0.05531838908791542, + 0.37399810552597046, + 0.05246146395802498 + ], + "credit_map": [ + 0.004792234394699335, + 0.003745575901120901, + 0.004255702719092369, + 0.005067487712949514, + 0.0036853624042123556, + 0.00617184117436409, + 0.004084366373717785, + 0.0086602121591568, + 0.009003981947898865, + 0.008066176436841488, + 0.005482921842485666, + 0.00552520714700222, + 0.004814225248992443, + 0.018714556470513344, + 0.0060562388971447945, + 0.00706113874912262, + 0.010724089108407497, + 0.007235554978251457, + 0.024562032893300056, + 0.006041687447577715, + 0.004376398865133524, + 0.026791488751769066, + 0.004402935970574617, + 0.004571425262838602, + 0.0017594898818060756, + 0.0011628217762336135, + 0.0014951553894206882, + 0.001397057087160647, + 0.0021030851639807224, + 0.0013402794720605016, + 0.0022070705890655518, + 0.0013707528123632073, + 0.03127824142575264, + 0.017714057117700577, + 0.014641125686466694, + 0.06673185527324677, + 0.024986106902360916, + 0.017547715455293655, + 0.019140737131237984, + 0.008005854673683643, + 0.004554287064820528, + 0.009770356118679047, + 0.004527100827544928, + 0.004194337874650955, + 0.008588443510234356, + 0.005437734071165323, + 0.003298717550933361, + 0.003998308442533016, + 0.009856420569121838, + 0.008549228310585022, + 0.006349978502839804, + 0.006559401284903288, + 0.0056100161746144295, + 0.008031066507101059, + 0.008116067387163639, + 0.014006298966705799, + 0.010755203664302826, + 0.005853625480085611, + 0.005569227039813995, + 0.005780443549156189, + 0.004794670268893242, + 0.007085429038852453, + 0.006403489504009485, + 0.009076302871108055, + 0.0799122005701065, + 0.04931449145078659, + 0.032273776829242706, + 0.03631847724318504, + 0.020670056343078613, + 0.026383904740214348, + 0.030821209773421288, + 0.09830403327941895, + 0.007412605453282595, + 0.006129685789346695, + 0.007779022213071585, + 0.006613289006054401, + 0.005392860621213913, + 0.005186871159821749, + 0.006283565424382687, + 0.007663566153496504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.07253359258174896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898525893688202 + ], + "opposite_logprob_tokens": [ + -2.5789852142333984 + ], + "logprob_reward": -0.07898525893688202, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898525893688202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07724833488464355, + 0.04318241402506828, + 0.15679378807544708, + 0.07295931875705719, + 0.14239533245563507, + 0.05444733798503876, + 0.06216945871710777, + 0.04202571511268616, + 0.21136997640132904, + 0.13740836083889008 + ], + "credit_map": [ + 0.008722983300685883, + 0.008720519952476025, + 0.008527098223567009, + 0.008098512887954712, + 0.005926367361098528, + 0.014470942318439484, + 0.005790171679109335, + 0.016991738229990005, + 0.007585486397147179, + 0.004960859660059214, + 0.004492785315960646, + 0.005256237927824259, + 0.004084624350070953, + 0.004393064882606268, + 0.00873249489814043, + 0.003676861524581909, + 0.022339221090078354, + 0.015387495048344135, + 0.04140383005142212, + 0.00986403413116932, + 0.008424500934779644, + 0.04190746694803238, + 0.008864974603056908, + 0.008602280169725418, + 0.00928250141441822, + 0.006532033905386925, + 0.009223912842571735, + 0.007672125939279795, + 0.008645802736282349, + 0.01368733774870634, + 0.009397082030773163, + 0.008518518880009651, + 0.012802467681467533, + 0.01829204149544239, + 0.011243672110140324, + 0.011642545461654663, + 0.03837672621011734, + 0.02253032848238945, + 0.01792767085134983, + 0.009579881094396114, + 0.006054092664271593, + 0.0065913875587284565, + 0.004341525956988335, + 0.003951380494982004, + 0.009220367297530174, + 0.01347602903842926, + 0.0032827325630933046, + 0.00752981984987855, + 0.009996351785957813, + 0.006529632024466991, + 0.006716820877045393, + 0.006100501865148544, + 0.0072680371813476086, + 0.006931000854820013, + 0.007256669458001852, + 0.011370443738996983, + 0.007220410276204348, + 0.004648644011467695, + 0.003530430141836405, + 0.00438709557056427, + 0.003448219271376729, + 0.008722266182303429, + 0.005407801829278469, + 0.004660848993808031, + 0.04044999182224274, + 0.027856631204485893, + 0.022196296602487564, + 0.022820567712187767, + 0.014520694501698017, + 0.017951155081391335, + 0.01707119308412075, + 0.04850344359874725, + 0.02284684032201767, + 0.012464025057852268, + 0.014399812556803226, + 0.010546055622398853, + 0.011851847171783447, + 0.019423380494117737, + 0.013832216151058674, + 0.0320441760122776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.07898525893688202 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343176141381264 + ], + "opposite_logprob_tokens": [ + -3.6513431072235107 + ], + "logprob_reward": -0.026343176141381264, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343176141381264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05833904817700386, + 0.07332655787467957, + 0.1540290266275406, + 0.02608402632176876, + 0.12442028522491455, + 0.022674204781651497, + 0.08112456649541855, + 0.02188851870596409, + 0.35075506567955017, + 0.08735860884189606 + ], + "credit_map": [ + 0.006834528408944607, + 0.006415690295398235, + 0.0067982422187924385, + 0.009649804793298244, + 0.007356465328484774, + 0.008800730109214783, + 0.005939024034887552, + 0.006544573698192835, + 0.007478803861886263, + 0.00848599337041378, + 0.006375555880367756, + 0.006519939750432968, + 0.004826235119253397, + 0.02160181663930416, + 0.011574975214898586, + 0.006463249679654837, + 0.01679936610162258, + 0.012817206792533398, + 0.038805246353149414, + 0.011850803159177303, + 0.006803824566304684, + 0.05151371285319328, + 0.007370225619524717, + 0.008068663999438286, + 0.004269061144441366, + 0.002347639063373208, + 0.0032271945383399725, + 0.002199948998168111, + 0.004303270019590855, + 0.0023702403996139765, + 0.003341138828545809, + 0.004025537986308336, + 0.019583597779273987, + 0.011637922376394272, + 0.011174537241458893, + 0.035218361765146255, + 0.014344465918838978, + 0.01174891646951437, + 0.01253662072122097, + 0.008175874128937721, + 0.003097217995673418, + 0.003566171508282423, + 0.003381137503311038, + 0.00206427788361907, + 0.003957686480134726, + 0.002854129998013377, + 0.0014541859272867441, + 0.0022994012106209993, + 0.011895093135535717, + 0.010481713339686394, + 0.006942030508071184, + 0.006852329708635807, + 0.007202897220849991, + 0.01120018120855093, + 0.00946348998695612, + 0.017086848616600037, + 0.003681402886286378, + 0.0022884784266352654, + 0.0027064208406955004, + 0.0027412944473326206, + 0.0019912919960916042, + 0.004118049517273903, + 0.00235884846188128, + 0.0020027346909046173, + 0.0767301544547081, + 0.043532367795705795, + 0.029683582484722137, + 0.03213938698172569, + 0.019767342135310173, + 0.024932455271482468, + 0.030971763655543327, + 0.09299806505441666, + 0.012015369720757008, + 0.010380841791629791, + 0.014990203082561493, + 0.00981900654733181, + 0.00952382292598486, + 0.00757451681420207, + 0.010000936686992645, + 0.013053915463387966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 4 + ], + "selected_reward": -0.026343176141381264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092770162969828 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092770162969828, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092770162969828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08065609633922577, + 0.02885296754539013, + 0.12435898929834366, + 0.02197970263659954, + 0.27679604291915894, + 0.02048904076218605, + 0.05503857880830765, + 0.08454672247171402, + 0.18507251143455505, + 0.12220937758684158 + ], + "credit_map": [ + 0.012319449335336685, + 0.00731293810531497, + 0.011585034430027008, + 0.011413974687457085, + 0.009000011719763279, + 0.011529654264450073, + 0.00902725849300623, + 0.008467772975564003, + 0.0038514756597578526, + 0.004137884359806776, + 0.0032974311616271734, + 0.004026066046208143, + 0.002908307360485196, + 0.0033099004067480564, + 0.004650183953344822, + 0.002671716269105673, + 0.014415828511118889, + 0.011386236175894737, + 0.03250603377819061, + 0.00908195972442627, + 0.0060493843629956245, + 0.039484184235334396, + 0.0054993778467178345, + 0.005935974884778261, + 0.003107689321041107, + 0.00258160219527781, + 0.002049180446192622, + 0.002041297033429146, + 0.003964776173233986, + 0.0023729123640805483, + 0.0023060524836182594, + 0.0035561912227422, + 0.04581466689705849, + 0.02054361253976822, + 0.01856772042810917, + 0.10586010664701462, + 0.024220556020736694, + 0.02419172041118145, + 0.025143584236502647, + 0.012454051524400711, + 0.002307357033714652, + 0.0035060588270425797, + 0.0025544501841068268, + 0.00181935157161206, + 0.004734127316623926, + 0.0019318481208756566, + 0.001844877377152443, + 0.001790970447473228, + 0.007785904686897993, + 0.007807785179466009, + 0.005063251592218876, + 0.005538190249353647, + 0.0049694739282131195, + 0.007527748588472605, + 0.006176324561238289, + 0.01016989815980196, + 0.013857683166861534, + 0.008030449971556664, + 0.007240017410367727, + 0.00835368037223816, + 0.008190159685909748, + 0.01981523633003235, + 0.009871774353086948, + 0.009187716990709305, + 0.03843818977475166, + 0.020185435190796852, + 0.018651511520147324, + 0.01513742282986641, + 0.009728429839015007, + 0.015877433121204376, + 0.016552269458770752, + 0.05050181224942207, + 0.01707572676241398, + 0.01139181386679411, + 0.019772281870245934, + 0.01607983559370041, + 0.009118362329900265, + 0.013658015988767147, + 0.011929672211408615, + 0.023183658719062805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.004092770162969828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291328817605972 + ], + "opposite_logprob_tokens": [ + -2.2379133701324463 + ], + "logprob_reward": -0.11291328817605972, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291328817605972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08256005495786667, + 0.02938833460211754, + 0.050786543637514114, + 0.09698526561260223, + 0.36312517523765564, + 0.04885717108845711, + 0.05277683958411217, + 0.019932584837079048, + 0.1703682839870453, + 0.08521972596645355 + ], + "credit_map": [ + 0.011157311499118805, + 0.012905962765216827, + 0.00783606618642807, + 0.008806336671113968, + 0.0076485611498355865, + 0.011348890140652657, + 0.007594996131956577, + 0.015261939726769924, + 0.0033690333366394043, + 0.005232241470366716, + 0.004430556204169989, + 0.0031010101083666086, + 0.003102814545854926, + 0.005016827955842018, + 0.0022271513007581234, + 0.002908702241256833, + 0.007273444905877113, + 0.009596780873835087, + 0.007230325136333704, + 0.0036218627355992794, + 0.004260512068867683, + 0.007902584969997406, + 0.0061599984765052795, + 0.0047410400584340096, + 0.01099696196615696, + 0.008777459152042866, + 0.014735092408955097, + 0.009998412802815437, + 0.011750868521630764, + 0.022570651024580002, + 0.008893665857613087, + 0.00926215946674347, + 0.044553469866514206, + 0.03920440375804901, + 0.02834370918571949, + 0.07183009386062622, + 0.07432623952627182, + 0.047033973038196564, + 0.03529612347483635, + 0.022537175565958023, + 0.0060573723167181015, + 0.005358307156711817, + 0.00759487459436059, + 0.006690536625683308, + 0.005506003275513649, + 0.004516713786870241, + 0.008036079816520214, + 0.0050972867757081985, + 0.008071507327258587, + 0.006678469944745302, + 0.0051610879600048065, + 0.005256549920886755, + 0.004116728436201811, + 0.006594841834157705, + 0.005475564859807491, + 0.011422094888985157, + 0.002800586400553584, + 0.0021431990899145603, + 0.0019974021706730127, + 0.002427305793389678, + 0.0018727855058386922, + 0.003432460827752948, + 0.0019429244566708803, + 0.003315923037007451, + 0.03022083081305027, + 0.0261648278683424, + 0.01858704350888729, + 0.021830938756465912, + 0.014867974445223808, + 0.010939586907625198, + 0.014849487692117691, + 0.03290760889649391, + 0.009324027225375175, + 0.0076263644732534885, + 0.018410664051771164, + 0.010242066346108913, + 0.009508769027888775, + 0.007710036356002092, + 0.010047318413853645, + 0.012350483797490597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.11291328817605972 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up immerses the viewer in a fierce battle between two leopard lizards on a dark, wet metal floor. Their bumpy, alien-like, banded skin, a detailed mosaic of light yellow and dark spots, shimmers and reflects the scene in distorted ripples. The only illumination comes from the relentless, out-of-focus glow of futuristic neon signs in the background, which cast ever-shifting, high-contrast electric blues and vibrant fuchsias across their writhing bodies. One lizard viciously bites at the other's flank, jaws clenched, while the defender struggles desperately, its spindly limbs and tail slapping against the wet surface, sending tiny splashes of water flying. Their tiny claws scrape frantically against the cold steel as they struggle for purchase, their desperate movements and the overpowering synthetic light combining to create a raw, tense, and distinctly dystopian sci-fi atmosphere.", + "seed": 14050, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_6/index_3583_seed_14050_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": -0.031544498167932034, + "vlm_logprob": -0.031544498167932034, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.031544498167932034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05188804864883423 + ], + "opposite_logprob_tokens": [ + -3.0518879890441895 + ], + "logprob_reward": -0.05188804864883423, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05188804864883423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18178334832191467, + 0.06984657794237137, + 0.02910938858985901, + 0.07451105117797852, + 0.027402939274907112, + 0.14401225745677948, + 0.09417369216680527, + 0.17293666303157806, + 0.15003301203250885, + 0.0561910979449749 + ], + "credit_map": [ + 0.024987807497382164, + 0.014584599062800407, + 0.03315141424536705, + 0.03327367082238197, + 0.018299754709005356, + 0.018921850249171257, + 0.017471371218562126, + 0.02109287865459919, + 0.004877353552728891, + 0.008651688694953918, + 0.005042627919465303, + 0.007040140684694052, + 0.012798392213881016, + 0.015256782062351704, + 0.006908518727868795, + 0.009271074086427689, + 0.0029733716510236263, + 0.004480297211557627, + 0.0018652136204764247, + 0.0035517544019967318, + 0.004405235406011343, + 0.004154112655669451, + 0.0039008590392768383, + 0.0037785435561090708, + 0.009461835026741028, + 0.004560976754873991, + 0.004635493736714125, + 0.01327713206410408, + 0.008009129203855991, + 0.010915355756878853, + 0.005467635113745928, + 0.01818349026143551, + 0.003025116864591837, + 0.0028089387342333794, + 0.0022661378607153893, + 0.006795309484004974, + 0.0023636389523744583, + 0.002131614601239562, + 0.005536430981010199, + 0.002475749934092164, + 0.029393227770924568, + 0.015944378450512886, + 0.011246200650930405, + 0.04506874829530716, + 0.009169385768473148, + 0.00825891736894846, + 0.018659038469195366, + 0.006272348575294018, + 0.012775441631674767, + 0.010342258028686047, + 0.011549980379641056, + 0.01231742650270462, + 0.011827194131910801, + 0.010833974927663803, + 0.01151992753148079, + 0.013007484376430511, + 0.021267101168632507, + 0.01973658613860607, + 0.013907808810472488, + 0.010927664116024971, + 0.011104204691946507, + 0.013195388950407505, + 0.04918549209833145, + 0.033612411469221115, + 0.02060740627348423, + 0.020775893703103065, + 0.018195783719420433, + 0.023393049836158752, + 0.017984263598918915, + 0.016806578263640404, + 0.020438017323613167, + 0.011832017451524734, + 0.007797838654369116, + 0.007509131915867329, + 0.006725236773490906, + 0.0104044359177351, + 0.005791875999420881, + 0.005110804922878742, + 0.008239304646849632, + 0.004612465389072895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.05188804864883423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709973096847534 + ], + "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": [ + 0.07846132665872574, + 0.01966613158583641, + 0.0072427233681082726, + 0.005927978083491325, + 0.006973114795982838, + 0.0465361513197422, + 0.012775242328643799, + 0.7750205397605896, + 0.04052518680691719, + 0.006871524266898632 + ], + "credit_map": [ + 0.005726371426135302, + 0.0056290337815880775, + 0.016269396990537643, + 0.020299827679991722, + 0.007611674256622791, + 0.007233281619846821, + 0.008040846325457096, + 0.007650899700820446, + 0.00352590368129313, + 0.002616587793454528, + 0.00199287966825068, + 0.0027096751146018505, + 0.002903192536905408, + 0.0018395102815702558, + 0.002271049190312624, + 0.0018073355313390493, + 0.0008285269141197205, + 0.0008272283594124019, + 0.0008689484093338251, + 0.0010145597625523806, + 0.0008150992798618972, + 0.000934517418500036, + 0.00104857690166682, + 0.0009052667883224785, + 0.0008975079981610179, + 0.0006974365678615868, + 0.000543189758900553, + 0.0007747544441372156, + 0.0007986201671883464, + 0.0006174042937345803, + 0.0007029306725598872, + 0.0008961350540630519, + 0.0008871419704519212, + 0.0005594034446403384, + 0.0006452960660681129, + 0.001597311464138329, + 0.0006493624532595277, + 0.0006362645654007792, + 0.0012572082923725247, + 0.0007411271799355745, + 0.011728335171937943, + 0.002937780227512121, + 0.0026364964433014393, + 0.016232801601290703, + 0.0028630252927541733, + 0.002778788562864065, + 0.005584306083619595, + 0.0017746201483532786, + 0.0015413733199238777, + 0.0016520314384251833, + 0.001343834912404418, + 0.002050909446552396, + 0.002026108093559742, + 0.0011801350628957152, + 0.0017557631945237517, + 0.0012250879080966115, + 0.09865617752075195, + 0.05107970908284187, + 0.08911929279565811, + 0.0416884683072567, + 0.03860984370112419, + 0.06771508604288101, + 0.2103797048330307, + 0.17777231335639954, + 0.0031371128279715776, + 0.0033834392670542, + 0.008485454134643078, + 0.005460196174681187, + 0.0047903601080179214, + 0.003858062671497464, + 0.0048957765102386475, + 0.006514786742627621, + 0.0008290411788038909, + 0.0008835531189106405, + 0.0009887726046144962, + 0.0008071376360021532, + 0.0006124822539277375, + 0.0010894087608903646, + 0.0008120656129904091, + 0.0008490634500049055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.033709973096847534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033700522035360336 + ], + "opposite_logprob_tokens": [ + -3.408700466156006 + ], + "logprob_reward": -0.033700522035360336, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033700522035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20435459911823273, + 0.05165610834956169, + 0.017787838354706764, + 0.018383167684078217, + 0.012015393935143948, + 0.030300838872790337, + 0.05194524675607681, + 0.5714526176452637, + 0.021144939586520195, + 0.020959267392754555 + ], + "credit_map": [ + 0.035210736095905304, + 0.016278909519314766, + 0.04190772399306297, + 0.038595911115407944, + 0.015068918466567993, + 0.0166187584400177, + 0.016189495101571083, + 0.024484138935804367, + 0.005603158380836248, + 0.006070262752473354, + 0.00485950754955411, + 0.007520759012550116, + 0.006489685736596584, + 0.009362736716866493, + 0.005111705511808395, + 0.006638294085860252, + 0.0016821554163470864, + 0.0025919515173882246, + 0.0012353711063042283, + 0.0028480638284236193, + 0.002104168524965644, + 0.002556255552917719, + 0.0025962803047150373, + 0.002173592569306493, + 0.003432878525927663, + 0.0018651374848559499, + 0.0017494474304839969, + 0.0020662767346948385, + 0.0018150187097489834, + 0.0024731166195124388, + 0.0018394043436273932, + 0.003141887718811631, + 0.0016535053728148341, + 0.0012409952469170094, + 0.0019944426603615284, + 0.0019702669233083725, + 0.0012343174312263727, + 0.0011808613780885935, + 0.001662311959080398, + 0.0010786930797621608, + 0.004453752189874649, + 0.0026200010906904936, + 0.0029864772222936153, + 0.009195041842758656, + 0.00286993570625782, + 0.0020501569379121065, + 0.0042058383114635944, + 0.0019196337088942528, + 0.0070798625238239765, + 0.00751098757609725, + 0.006158127449452877, + 0.00626911548897624, + 0.00852593407034874, + 0.005131660494953394, + 0.006447125691920519, + 0.004822434391826391, + 0.07234855741262436, + 0.055535703897476196, + 0.05237988755106926, + 0.032310549169778824, + 0.029572971165180206, + 0.04626447334885597, + 0.157242551445961, + 0.12579786777496338, + 0.0022839116863906384, + 0.0022764825262129307, + 0.0031212966423481703, + 0.004038957413285971, + 0.001755105098709464, + 0.002190219471231103, + 0.0029413311276584864, + 0.002537635387852788, + 0.002239642199128866, + 0.0029968763701617718, + 0.0038371747359633446, + 0.0029325734358280897, + 0.001588367624208331, + 0.0024391268379986286, + 0.002590078627690673, + 0.002335428027436137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.033700522035360336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636386133730412 + ], + "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": [ + 0.08794769644737244, + 0.022442972287535667, + 0.022777197882533073, + 0.08137332648038864, + 0.018056176602840424, + 0.039043862372636795, + 0.02998092584311962, + 0.6508035659790039, + 0.024166179820895195, + 0.023408133536577225 + ], + "credit_map": [ + 0.00900243129581213, + 0.008594058454036713, + 0.01384675968438387, + 0.017571231350302696, + 0.008645597845315933, + 0.010044458322227001, + 0.008947399444878101, + 0.011295744217932224, + 0.004364457447081804, + 0.0023602398578077555, + 0.0023490500170737505, + 0.0026167386677116156, + 0.002951369620859623, + 0.0021232510916888714, + 0.0026434832252562046, + 0.0030343816615641117, + 0.00225671473890543, + 0.0026668638456612825, + 0.0033159651793539524, + 0.00254507502540946, + 0.002852602396160364, + 0.0035069941077381372, + 0.0028350218199193478, + 0.002797958441078663, + 0.011650539003312588, + 0.008808838203549385, + 0.0072225467301905155, + 0.010454862378537655, + 0.01159327756613493, + 0.008618142455816269, + 0.011127524077892303, + 0.011897590011358261, + 0.002514852676540613, + 0.001209041802212596, + 0.002292081480845809, + 0.0021010746713727713, + 0.0024568340741097927, + 0.0017412652960047126, + 0.0026587308384478092, + 0.0030822940170764923, + 0.008710549212992191, + 0.0027654687874019146, + 0.0030176271684467793, + 0.013399746268987656, + 0.002389262430369854, + 0.002104918472468853, + 0.004734107758849859, + 0.0019221813417971134, + 0.004054909572005272, + 0.005095255561172962, + 0.00277019664645195, + 0.0038170802872627974, + 0.0039643882773816586, + 0.0030738231725990772, + 0.0032104908023029566, + 0.003994779661297798, + 0.08233000338077545, + 0.06325085461139679, + 0.06046748533844948, + 0.03630553185939789, + 0.03510858118534088, + 0.051973991096019745, + 0.17973454296588898, + 0.1416325867176056, + 0.0023213170934468508, + 0.00442446768283844, + 0.004282675217837095, + 0.003085260046645999, + 0.0022353834938257933, + 0.0019118963973596692, + 0.00298535218462348, + 0.0029198264237493277, + 0.0033721840009093285, + 0.0032179125119000673, + 0.003059625392779708, + 0.0040615336038172245, + 0.002530510537326336, + 0.0023793212603777647, + 0.002461242489516735, + 0.002325804438441992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.008636386133730412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two leopard lizards visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dark, wet metal floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one lizard biting the other's flank while the defender struggles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting provided by out-of-focus futuristic neon signs casting electric blues and vibrant fuchsias?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, tight close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787560924887657 + ], + "opposite_logprob_tokens": [ + -3.529787540435791 + ], + "logprob_reward": -0.029787560924887657, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787560924887657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1017959713935852, + 0.03372771292924881, + 0.024664994329214096, + 0.03982784226536751, + 0.04105544090270996, + 0.03677164763212204, + 0.03221200406551361, + 0.6325563788414001, + 0.04278373718261719, + 0.014604257419705391 + ], + "credit_map": [ + 0.007700108923017979, + 0.006479867734014988, + 0.022826669737696648, + 0.03082575649023056, + 0.008737735450267792, + 0.007741640321910381, + 0.010335816070437431, + 0.007148376666009426, + 0.0058443290181458, + 0.0035885001998394728, + 0.003842954058200121, + 0.0039943186566233635, + 0.0031572021543979645, + 0.0044828359968960285, + 0.00415524747222662, + 0.004662327002733946, + 0.0029082270339131355, + 0.0030652175191789865, + 0.002356413984671235, + 0.0026290558744221926, + 0.0038701847661286592, + 0.002643095562234521, + 0.00434403819963336, + 0.002848760224878788, + 0.004827066324651241, + 0.006043712608516216, + 0.003731226082891226, + 0.005094977095723152, + 0.006037821527570486, + 0.005187269765883684, + 0.004558670334517956, + 0.004347096662968397, + 0.005799515172839165, + 0.004334991332143545, + 0.004765667021274567, + 0.006806613877415657, + 0.005184012930840254, + 0.004400058649480343, + 0.005744484718888998, + 0.004020098131150007, + 0.0052540916949510574, + 0.0047745900228619576, + 0.0038185808807611465, + 0.008266604505479336, + 0.003163696499541402, + 0.0028198501095175743, + 0.005632394924759865, + 0.0030418382957577705, + 0.004557518754154444, + 0.004440077114850283, + 0.004756856244057417, + 0.004349144175648689, + 0.0034964182414114475, + 0.0035831567365676165, + 0.00376322609372437, + 0.003265606239438057, + 0.08030359447002411, + 0.052291397005319595, + 0.06498361378908157, + 0.0336579866707325, + 0.031691499054431915, + 0.05365686118602753, + 0.17473454773426056, + 0.14123687148094177, + 0.004352576564997435, + 0.003592072520405054, + 0.006407638546079397, + 0.004420118872076273, + 0.008972465991973877, + 0.004670283291488886, + 0.005452165845781565, + 0.004916416481137276, + 0.001706235110759735, + 0.0017281639156863093, + 0.0026269559748470783, + 0.002064202446490526, + 0.0013705838937312365, + 0.001668323646299541, + 0.0018160545732825994, + 0.001623737858608365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.029787560924887657 + } + ], + "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/0000155/group_7/index_tensor([5812, 5812])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_tensor([5812, 5812])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d5bf324d5c031941e48a855f10a64a85896cc15d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_tensor([5812, 5812])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping wide shot begins, the camera tracking smoothly horizontally from left to right, following a lone figure gliding across the vast frozen surface. The figure is a woman, her dark hair pulled back in a tight, elegant bun, and her movements graceful. She wears a deep red, long-sleeved velvet skating dress, the rich fabric striking as it ripples with her movements. Her white skates, adorned with gleaming silver blades, carve silent, clean lines across the ice. She glides effortlessly, her arms extended in a balanced, fluid pose, her expression calm and focused. The camera maintains an eye-level perspective, emphasizing the seamless connection between the skater and her environment: the iconic reflecting pool in front of the Taj Mahal, now a perfect, mirror-smooth expanse of ice. In the background, the majestic white marble mausoleum towers over the monument, its intricate details and magnificent dome softened by the morning mist. The entire scene is bathed in a soft, ethereal morning light, casting long, soft shadows and giving the white marble a warm, pale golden hue, creating an atmosphere of profound tranquility, dreaminess, and majesty.", + "seed": 2512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2512_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": -0.03588221319951117, + "vlm_logprob": -0.03588221319951117, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.03588221319951117, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0201646126806736 + ], + "opposite_logprob_tokens": [ + -4.145164489746094 + ], + "logprob_reward": -0.0201646126806736, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0201646126806736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09273923188447952, + 0.04982158914208412, + 0.05668773502111435, + 0.03686809912323952, + 0.021605169400572777, + 0.1590195596218109, + 0.1737709790468216, + 0.14758406579494476, + 0.17080067098140717, + 0.09110291302204132 + ], + "credit_map": [ + 0.013628790155053139, + 0.015021509490907192, + 0.012593301013112068, + 0.009121139533817768, + 0.015239923261106014, + 0.008803179487586021, + 0.009578424505889416, + 0.008752953261137009, + 0.00723990797996521, + 0.006881102453917265, + 0.0036029182374477386, + 0.004338597878813744, + 0.013363435864448547, + 0.005314475856721401, + 0.00520187197253108, + 0.0038792742416262627, + 0.008403228595852852, + 0.006435038056224585, + 0.00462941313162446, + 0.0046585677191615105, + 0.013710001483559608, + 0.006291619502007961, + 0.006903533358126879, + 0.005656329914927483, + 0.005283564794808626, + 0.006315902806818485, + 0.003908749204128981, + 0.0049964552745223045, + 0.006681562401354313, + 0.0037973152939230204, + 0.002836366882547736, + 0.0030481780413538218, + 0.003173286560922861, + 0.002460190560668707, + 0.003047224832698703, + 0.0032826566603034735, + 0.0038083025719970465, + 0.0020637784618884325, + 0.0018869086634367704, + 0.001882818411104381, + 0.02550159953534603, + 0.017607687041163445, + 0.015821566805243492, + 0.013421852141618729, + 0.04501952975988388, + 0.01792989857494831, + 0.01331144105643034, + 0.010405970737338066, + 0.028933493420481682, + 0.019088832661509514, + 0.018694208934903145, + 0.02264956757426262, + 0.02376394160091877, + 0.03337985649704933, + 0.015389695763587952, + 0.01187136210501194, + 0.0214560404419899, + 0.025237763300538063, + 0.011423598974943161, + 0.011404183693230152, + 0.02254221774637699, + 0.019535861909389496, + 0.015683239325881004, + 0.02030116319656372, + 0.025360794737935066, + 0.0257427841424942, + 0.021260809153318405, + 0.028434351086616516, + 0.02152014523744583, + 0.018437804654240608, + 0.012378103099763393, + 0.017665864899754524, + 0.013200225308537483, + 0.013155844993889332, + 0.009293320588767529, + 0.01920250430703163, + 0.008349157869815826, + 0.009186934679746628, + 0.008503920398652554, + 0.01021099928766489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.0201646126806736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651040494441986 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651040494441986, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651040494441986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.180279940366745, + 0.03987441956996918, + 0.09562303870916367, + 0.03816726803779602, + 0.0249340683221817, + 0.13662956655025482, + 0.14952535927295685, + 0.08255211263895035, + 0.09956976026296616, + 0.15284448862075806 + ], + "credit_map": [ + 0.025684742256999016, + 0.029009701684117317, + 0.0233216043561697, + 0.015773385763168335, + 0.02741558477282524, + 0.019059298560023308, + 0.021291987970471382, + 0.018723635002970695, + 0.006190560292452574, + 0.004401118494570255, + 0.004549936391413212, + 0.0029445309191942215, + 0.009792085736989975, + 0.004687636625021696, + 0.003117144340649247, + 0.004191404674202204, + 0.01331302896142006, + 0.010362956672906876, + 0.00834656786173582, + 0.00629048328846693, + 0.02554544061422348, + 0.008918145671486855, + 0.012806633487343788, + 0.010039781220257282, + 0.004832712933421135, + 0.004545896779745817, + 0.0032934171613305807, + 0.0033488848712295294, + 0.009539982303977013, + 0.004188304767012596, + 0.004613060969859362, + 0.003805010113865137, + 0.0035807641688734293, + 0.0026344561483711004, + 0.003583302954211831, + 0.0024287302512675524, + 0.004952610470354557, + 0.003352045314386487, + 0.002498465124517679, + 0.001903693308122456, + 0.020967718213796616, + 0.015865953639149666, + 0.013504288159310818, + 0.012911060824990273, + 0.03778988867998123, + 0.014817552641034126, + 0.01122225634753704, + 0.009550852701067924, + 0.02525104023516178, + 0.020296042785048485, + 0.013336660340428352, + 0.025048473849892616, + 0.015391591005027294, + 0.027804186567664146, + 0.012211997993290424, + 0.010185355320572853, + 0.009764646179974079, + 0.008691900409758091, + 0.010645124129951, + 0.007662060204893351, + 0.011539150029420853, + 0.0063821240328252316, + 0.01458453293889761, + 0.013282577507197857, + 0.01476866751909256, + 0.015600507147610188, + 0.013230479322373867, + 0.0197835061699152, + 0.012640131637454033, + 0.009432795457541943, + 0.007035454269498587, + 0.007078217808157206, + 0.0218197014182806, + 0.025537127628922462, + 0.013008948415517807, + 0.029956737533211708, + 0.014346711337566376, + 0.013478251174092293, + 0.016440143808722496, + 0.018256869167089462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.008651040494441986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003194113029167056 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.003194113029167056, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003194113029167056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1259133517742157, + 0.043942488729953766, + 0.09092891961336136, + 0.033938195556402206, + 0.039767030626535416, + 0.14709103107452393, + 0.17967666685581207, + 0.1435876041650772, + 0.13819701969623566, + 0.05695762112736702 + ], + "credit_map": [ + 0.01625349372625351, + 0.014497359283268452, + 0.015907255932688713, + 0.012446095235645771, + 0.016650699079036713, + 0.020633254200220108, + 0.017660275101661682, + 0.011864932253956795, + 0.005475331097841263, + 0.006474307272583246, + 0.006654154043644667, + 0.005985185503959656, + 0.0072255306877195835, + 0.0034377602860331535, + 0.005024276208132505, + 0.003665947588160634, + 0.011966385878622532, + 0.010111058130860329, + 0.009845574386417866, + 0.007779519073665142, + 0.02165338769555092, + 0.008120054379105568, + 0.010727201588451862, + 0.01072574220597744, + 0.004357409663498402, + 0.0036588518414646387, + 0.004074574913829565, + 0.004278809763491154, + 0.006277722306549549, + 0.0038566691800951958, + 0.0034622815437614918, + 0.003971877973526716, + 0.005099636968225241, + 0.004251777660101652, + 0.0047101955860853195, + 0.005386669188737869, + 0.005216329824179411, + 0.004134282004088163, + 0.0070863524451851845, + 0.0038817892782390118, + 0.023518351837992668, + 0.018213409930467606, + 0.013562050648033619, + 0.012166707776486874, + 0.04104236140847206, + 0.01566346175968647, + 0.012401236221194267, + 0.010523462668061256, + 0.03235070779919624, + 0.028591778129339218, + 0.01667753979563713, + 0.030831297859549522, + 0.01821655035018921, + 0.017850015312433243, + 0.018640689551830292, + 0.016518088057637215, + 0.01695571467280388, + 0.017259418964385986, + 0.014803648926317692, + 0.011259115301072598, + 0.02210635133087635, + 0.018642151728272438, + 0.022285424172878265, + 0.02027578093111515, + 0.0180523619055748, + 0.01734611950814724, + 0.014298876747488976, + 0.025322476401925087, + 0.021253976970911026, + 0.01390003226697445, + 0.014392695389688015, + 0.013630485162138939, + 0.008097771555185318, + 0.005790873896330595, + 0.006780205760151148, + 0.007064243778586388, + 0.008129669353365898, + 0.008660138584673405, + 0.005720493849366903, + 0.006714226678013802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.003194113029167056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635777790099382 + ], + "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": [ + 0.2782420814037323, + 0.061841972172260284, + 0.11043404787778854, + 0.06185967102646828, + 0.061541929841041565, + 0.0897577628493309, + 0.19464801251888275, + 0.052075862884521484, + 0.03562574461102486, + 0.05397298187017441 + ], + "credit_map": [ + 0.031664058566093445, + 0.03178056329488754, + 0.022122148424386978, + 0.03626787289977074, + 0.04202347621321678, + 0.026496782898902893, + 0.062105435878038406, + 0.02578170970082283, + 0.009074156172573566, + 0.011706754565238953, + 0.00464628916233778, + 0.004375627264380455, + 0.011523938737809658, + 0.008285469375550747, + 0.006605817005038261, + 0.005623909179121256, + 0.01607372798025608, + 0.012754141353070736, + 0.00977629330009222, + 0.008685032837092876, + 0.028706198558211327, + 0.012104318477213383, + 0.013231709599494934, + 0.00910261645913124, + 0.009512590244412422, + 0.008221088908612728, + 0.0045180111192166805, + 0.006257843226194382, + 0.013772084377706051, + 0.006891218479722738, + 0.007446054834872484, + 0.005240771919488907, + 0.00866161473095417, + 0.0053833648562431335, + 0.006203728262335062, + 0.00914741400629282, + 0.01267279777675867, + 0.0072096181102097034, + 0.006766480393707752, + 0.005496907979249954, + 0.01320569682866335, + 0.012373065575957298, + 0.009253999218344688, + 0.009841319173574448, + 0.021132366731762886, + 0.009248757734894753, + 0.007107411045581102, + 0.00759514095261693, + 0.031249670311808586, + 0.022983236238360405, + 0.021502530202269554, + 0.024972420185804367, + 0.022025251761078835, + 0.03307522088289261, + 0.019364072009921074, + 0.01947559043765068, + 0.006926043890416622, + 0.003459787927567959, + 0.003930800594389439, + 0.003640057286247611, + 0.009202631190419197, + 0.007664271164685488, + 0.008942421525716782, + 0.008309844881296158, + 0.004092162940651178, + 0.003948359750211239, + 0.0035382111091166735, + 0.004271650221198797, + 0.005076602566987276, + 0.004998701624572277, + 0.004761266987770796, + 0.004938785452395678, + 0.009270325303077698, + 0.006550699472427368, + 0.005822265986353159, + 0.0053937481716275215, + 0.006297050975263119, + 0.006444833241403103, + 0.006280421745032072, + 0.007913632318377495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "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 main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14276552200317383 + ], + "opposite_logprob_tokens": [ + -2.017765522003174 + ], + "logprob_reward": -0.14276552200317383, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14276552200317383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0933932512998581, + 0.05358607694506645, + 0.031347211450338364, + 0.0471222884953022, + 0.07201626896858215, + 0.3296070694923401, + 0.10475954413414001, + 0.10975093394517899, + 0.085581935942173, + 0.07283540815114975 + ], + "credit_map": [ + 0.011372464708983898, + 0.012990441173315048, + 0.00640268437564373, + 0.010882831178605556, + 0.016227219253778458, + 0.01082866732031107, + 0.017411956563591957, + 0.007276988122612238, + 0.006693268194794655, + 0.00726248137652874, + 0.004813103936612606, + 0.003969572950154543, + 0.008900294080376625, + 0.007127753458917141, + 0.006813566666096449, + 0.008006037212908268, + 0.004100187681615353, + 0.005040814634412527, + 0.003232138231396675, + 0.004148459527641535, + 0.005728268995881081, + 0.003110797144472599, + 0.0034432036336511374, + 0.0025433411356061697, + 0.006594275124371052, + 0.005459751468151808, + 0.0039572399109601974, + 0.0063502732664346695, + 0.009996443055570126, + 0.004881022498011589, + 0.004935508593916893, + 0.004947772715240717, + 0.010798818431794643, + 0.007639885414391756, + 0.008948571048676968, + 0.008520658127963543, + 0.010150831192731857, + 0.012137182988226414, + 0.006745823659002781, + 0.007074503228068352, + 0.053514186292886734, + 0.04218202084302902, + 0.03049689158797264, + 0.026495421305298805, + 0.0970851257443428, + 0.033839453011751175, + 0.025929292663931847, + 0.0200646985322237, + 0.014972032979130745, + 0.015048099681735039, + 0.010148289613425732, + 0.01520617213100195, + 0.015424077399075031, + 0.012887774035334587, + 0.011792732402682304, + 0.00928037241101265, + 0.015648240223526955, + 0.013654437847435474, + 0.011239267885684967, + 0.008838733658194542, + 0.023531878367066383, + 0.017190618440508842, + 0.009651931934058666, + 0.009995833039283752, + 0.012164532206952572, + 0.009653449058532715, + 0.009901260957121849, + 0.011078118346631527, + 0.009914139285683632, + 0.009905662387609482, + 0.012727723456919193, + 0.010237041860818863, + 0.01035221852362156, + 0.008972045965492725, + 0.006992710288614035, + 0.009750774130225182, + 0.010632774792611599, + 0.00996406376361847, + 0.007596265058964491, + 0.008574558421969414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.14276552200317383 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins, the camera tracking smoothly horizontally from left to right, following a lone figure gliding across the vast frozen surface. The figure is a woman, her dark hair pulled back in a tight, elegant bun, and her movements graceful. She wears a deep red, long-sleeved velvet skating dress, the rich fabric striking as it ripples with her movements. Her white skates, adorned with gleaming silver blades, carve silent, clean lines across the ice. She glides effortlessly, her arms extended in a balanced, fluid pose, her expression calm and focused. The camera maintains an eye-level perspective, emphasizing the seamless connection between the skater and her environment: the iconic reflecting pool in front of the Taj Mahal, now a perfect, mirror-smooth expanse of ice. In the background, the majestic white marble mausoleum towers over the monument, its intricate details and magnificent dome softened by the morning mist. The entire scene is bathed in a soft, ethereal morning light, casting long, soft shadows and giving the white marble a warm, pale golden hue, creating an atmosphere of profound tranquility, dreaminess, and majesty.", + "seed": 2513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2513_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": -0.04582745083607733, + "vlm_logprob": -0.04582745083607733, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.04582745083607733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027374165132641792 + ], + "opposite_logprob_tokens": [ + -3.777374267578125 + ], + "logprob_reward": -0.027374165132641792, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027374165132641792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08733748644590378, + 0.030340801924467087, + 0.05362950637936592, + 0.15208935737609863, + 0.054238494485616684, + 0.10265913605690002, + 0.22499814629554749, + 0.13028785586357117, + 0.07946960628032684, + 0.08494958281517029 + ], + "credit_map": [ + 0.01358547993004322, + 0.012400397099554539, + 0.012097598053514957, + 0.013958562165498734, + 0.009420114569365978, + 0.007346447091549635, + 0.010612543672323227, + 0.007916339673101902, + 0.004215035121887922, + 0.003965366166085005, + 0.004167674575001001, + 0.004940250422805548, + 0.0029422042425721884, + 0.0030421148985624313, + 0.003616301342844963, + 0.003451854456216097, + 0.00805272813886404, + 0.013700195588171482, + 0.005221361760050058, + 0.004700596444308758, + 0.0057495092041790485, + 0.007879571989178658, + 0.004160617478191853, + 0.004164927639067173, + 0.02278418280184269, + 0.016402222216129303, + 0.015019705519080162, + 0.015093160793185234, + 0.03804144263267517, + 0.015215077437460423, + 0.018004192039370537, + 0.011529367417097092, + 0.008394456468522549, + 0.006478053983300924, + 0.0066322945058345795, + 0.005869431886821985, + 0.00728681031614542, + 0.007069519720971584, + 0.0064320312812924385, + 0.006075897719711065, + 0.029733937233686447, + 0.009508936665952206, + 0.008113447576761246, + 0.005403432063758373, + 0.008313857018947601, + 0.03188547492027283, + 0.003948313184082508, + 0.005751741584390402, + 0.026496944949030876, + 0.0347999706864357, + 0.031309012323617935, + 0.028942080214619637, + 0.02147061750292778, + 0.026131954044103622, + 0.028446586802601814, + 0.027400994673371315, + 0.01944539323449135, + 0.029872747138142586, + 0.01351549755781889, + 0.019783401861786842, + 0.012491801753640175, + 0.018086126074194908, + 0.00695791793987155, + 0.010134972631931305, + 0.012488022446632385, + 0.011306926608085632, + 0.00740601122379303, + 0.00704722898080945, + 0.010699613951146603, + 0.010427705943584442, + 0.011611741967499256, + 0.008482354693114758, + 0.010178513824939728, + 0.005719494540244341, + 0.012396515347063541, + 0.00780926737934351, + 0.009293273091316223, + 0.011407204903662205, + 0.009643780998885632, + 0.018501535058021545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.027374165132641792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205877669155598 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205877669155598, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205877669155598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07793296873569489, + 0.03029165416955948, + 0.041506364941596985, + 0.06301258504390717, + 0.08459962904453278, + 0.16909344494342804, + 0.17470833659172058, + 0.09293031692504883, + 0.17554546892642975, + 0.09037931263446808 + ], + "credit_map": [ + 0.010137130506336689, + 0.011105043813586235, + 0.009838897734880447, + 0.014195680618286133, + 0.00913587398827076, + 0.009002845734357834, + 0.007328832056373358, + 0.0071886563673615456, + 0.004674480762332678, + 0.004391608759760857, + 0.004026159178465605, + 0.004211005289107561, + 0.003935080952942371, + 0.002823631977662444, + 0.0034555138554424047, + 0.0027741712983697653, + 0.0064985440112650394, + 0.006447099149227142, + 0.00415490148589015, + 0.005571706220507622, + 0.006318835541605949, + 0.004369660746306181, + 0.004102520179003477, + 0.004043095279484987, + 0.00869425293058157, + 0.005788801703602076, + 0.00848124548792839, + 0.006801145151257515, + 0.014559727162122726, + 0.006928565446287394, + 0.007645818404853344, + 0.00411302549764514, + 0.01346613559871912, + 0.012222962453961372, + 0.009281833656132221, + 0.008486506529152393, + 0.014026842080056667, + 0.011258307844400406, + 0.00866004265844822, + 0.007196987513452768, + 0.042265959084033966, + 0.01619063876569271, + 0.009574669413268566, + 0.011747281067073345, + 0.025476956740021706, + 0.043805718421936035, + 0.010989629663527012, + 0.00904258992522955, + 0.0195623729377985, + 0.02603720873594284, + 0.02537418343126774, + 0.02313726767897606, + 0.0142824100330472, + 0.01896892488002777, + 0.024726660922169685, + 0.022619280964136124, + 0.011734502390027046, + 0.022401418536901474, + 0.00960126705467701, + 0.012203468009829521, + 0.008729200810194016, + 0.012853815220296383, + 0.005008041393011808, + 0.010398591868579388, + 0.02662559412419796, + 0.021695367991924286, + 0.014199791476130486, + 0.015079300850629807, + 0.028688646852970123, + 0.021388664841651917, + 0.026196923106908798, + 0.021671155467629433, + 0.008604830130934715, + 0.008757461793720722, + 0.007668588776141405, + 0.010756180621683598, + 0.013166661374270916, + 0.010430047288537025, + 0.016803061589598656, + 0.014192482456564903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.014205877669155598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036178866866976023 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036178866866976023, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036178866866976023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12503494322299957, + 0.1798429787158966, + 0.055840302258729935, + 0.07721062749624252, + 0.08933040499687195, + 0.11235061287879944, + 0.10037047415971756, + 0.1519738733768463, + 0.06761203706264496, + 0.04043377935886383 + ], + "credit_map": [ + 0.016150088980793953, + 0.020543143153190613, + 0.012606346979737282, + 0.032524120062589645, + 0.012776031158864498, + 0.009059996344149113, + 0.013377386145293713, + 0.007997834123671055, + 0.02826850116252899, + 0.033973246812820435, + 0.017088454216718674, + 0.03593415766954422, + 0.01630294695496559, + 0.013016408309340477, + 0.0191658828407526, + 0.01609337143599987, + 0.00733815087005496, + 0.0065181562677025795, + 0.009453821927309036, + 0.007890705950558186, + 0.007975198328495026, + 0.004633315838873386, + 0.006805166136473417, + 0.005225780885666609, + 0.011131899431347847, + 0.010785111226141453, + 0.008769712410867214, + 0.008266609162092209, + 0.01320925448089838, + 0.010792325250804424, + 0.007794531527906656, + 0.006461180280894041, + 0.01267346739768982, + 0.012742273509502411, + 0.01232950296252966, + 0.009657884016633034, + 0.01138761080801487, + 0.009799787774682045, + 0.010069366544485092, + 0.010670511052012444, + 0.02885373868048191, + 0.01135934516787529, + 0.005641350522637367, + 0.007549133617430925, + 0.01607472635805607, + 0.02972903475165367, + 0.007499662693589926, + 0.005643617827445269, + 0.01134514156728983, + 0.017496412619948387, + 0.013913203030824661, + 0.01286568958312273, + 0.009647195227444172, + 0.011189068667590618, + 0.010248344391584396, + 0.013665413483977318, + 0.02154545485973358, + 0.029727837070822716, + 0.015119477175176144, + 0.01925290934741497, + 0.011965577490627766, + 0.020690005272626877, + 0.009042044170200825, + 0.024630576372146606, + 0.00976011622697115, + 0.013455193489789963, + 0.008508632890880108, + 0.007120276335626841, + 0.005120994057506323, + 0.007863092236220837, + 0.005662824492901564, + 0.010120900347828865, + 0.004600722808390856, + 0.004356338642537594, + 0.004702853970229626, + 0.004471709486097097, + 0.005374107975512743, + 0.004853700753301382, + 0.005568838678300381, + 0.006505507044494152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0036178866866976023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092888906598091 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092888906598091, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092888906598091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15293532609939575, + 0.04863468185067177, + 0.03246062994003296, + 0.11842559278011322, + 0.057399701327085495, + 0.19453316926956177, + 0.19856761395931244, + 0.06963047385215759, + 0.06489543616771698, + 0.06251738220453262 + ], + "credit_map": [ + 0.02163856104016304, + 0.01927848346531391, + 0.020028142258524895, + 0.028917662799358368, + 0.017518101260066032, + 0.017728175967931747, + 0.014536614529788494, + 0.013289577327668667, + 0.006382356397807598, + 0.0069343955256044865, + 0.005854183807969093, + 0.008001364767551422, + 0.006103514228016138, + 0.005003499332815409, + 0.005741364322602749, + 0.004614003002643585, + 0.004706802777945995, + 0.005026640370488167, + 0.004944091662764549, + 0.005044823512434959, + 0.00387119734659791, + 0.003949107136577368, + 0.00273556774482131, + 0.002182397525757551, + 0.017180679365992546, + 0.014949952252209187, + 0.011105932295322418, + 0.010595741681754589, + 0.023391934111714363, + 0.019087139517068863, + 0.011928309686481953, + 0.010185900144279003, + 0.007795022800564766, + 0.007449198979884386, + 0.007015167269855738, + 0.008869669400155544, + 0.007559531833976507, + 0.007385019212961197, + 0.005043948069214821, + 0.006282143294811249, + 0.04458233714103699, + 0.02033355087041855, + 0.011969383805990219, + 0.01665528677403927, + 0.02520865947008133, + 0.04643320292234421, + 0.014881527982652187, + 0.014469225890934467, + 0.025698594748973846, + 0.03884617239236832, + 0.019344888627529144, + 0.025223206728696823, + 0.022634001448750496, + 0.023287571966648102, + 0.02213922142982483, + 0.021393954753875732, + 0.006754477042704821, + 0.007625083904713392, + 0.007761276792734861, + 0.007295318879187107, + 0.008016695268452168, + 0.008382054045796394, + 0.0067517138086259365, + 0.01704385131597519, + 0.012080663815140724, + 0.010413842275738716, + 0.006658716592937708, + 0.00647606560960412, + 0.0067746941931545734, + 0.008844268508255482, + 0.006689316127449274, + 0.006957868114113808, + 0.010753827169537544, + 0.011675134301185608, + 0.006871166173368692, + 0.007430927827954292, + 0.006354783661663532, + 0.008983725681900978, + 0.005411611869931221, + 0.005036210175603628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.004092888906598091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17984643578529358 + ], + "opposite_logprob_tokens": [ + -1.8048464059829712 + ], + "logprob_reward": -0.17984643578529358, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17984643578529358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10992380231618881, + 0.06380048394203186, + 0.05579241365194321, + 0.15272784233093262, + 0.04071219265460968, + 0.09418607503175735, + 0.25336095690727234, + 0.10423583537340164, + 0.0640970766544342, + 0.061163339763879776 + ], + "credit_map": [ + 0.014555595815181732, + 0.019053246825933456, + 0.013567214831709862, + 0.01859734021127224, + 0.01434664148837328, + 0.011339843273162842, + 0.009499729610979557, + 0.008964195847511292, + 0.008954409509897232, + 0.010395395569503307, + 0.00731988251209259, + 0.011660274118185043, + 0.0076245409436523914, + 0.005420998204499483, + 0.006943857297301292, + 0.005481125321239233, + 0.00819952692836523, + 0.009357373230159283, + 0.0064824954606592655, + 0.006634411867707968, + 0.00632881885394454, + 0.007233267650008202, + 0.006818675436079502, + 0.004737840034067631, + 0.022897053509950638, + 0.017141398042440414, + 0.013218985870480537, + 0.01620757207274437, + 0.03709753230214119, + 0.018070116639137268, + 0.017554640769958496, + 0.010540551505982876, + 0.006255894433706999, + 0.00593503238633275, + 0.004100622143596411, + 0.006250807549804449, + 0.00577939860522747, + 0.005276676267385483, + 0.0035634408704936504, + 0.0035503224935382605, + 0.015488805249333382, + 0.012071210891008377, + 0.00773495202884078, + 0.0091780386865139, + 0.016761792823672295, + 0.016640929505228996, + 0.009295716881752014, + 0.0070146312937140465, + 0.028371650725603104, + 0.043677136301994324, + 0.03648935630917549, + 0.028533749282360077, + 0.021353069692850113, + 0.029476143419742584, + 0.024680618196725845, + 0.040779244154691696, + 0.011733431369066238, + 0.022233717143535614, + 0.007861047051846981, + 0.01423121802508831, + 0.01089051365852356, + 0.013233923353254795, + 0.01216940674930811, + 0.011882566846907139, + 0.0074816253036260605, + 0.01019152719527483, + 0.006798446178436279, + 0.007424257695674896, + 0.009501125663518906, + 0.008986517786979675, + 0.008387469686567783, + 0.005326106678694487, + 0.008310701698064804, + 0.004951481241732836, + 0.00836313609033823, + 0.004816068336367607, + 0.007358275819569826, + 0.008625769056379795, + 0.01008558738976717, + 0.008652320131659508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.17984643578529358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins, the camera tracking smoothly horizontally from left to right, following a lone figure gliding across the vast frozen surface. The figure is a woman, her dark hair pulled back in a tight, elegant bun, and her movements graceful. She wears a deep red, long-sleeved velvet skating dress, the rich fabric striking as it ripples with her movements. Her white skates, adorned with gleaming silver blades, carve silent, clean lines across the ice. She glides effortlessly, her arms extended in a balanced, fluid pose, her expression calm and focused. The camera maintains an eye-level perspective, emphasizing the seamless connection between the skater and her environment: the iconic reflecting pool in front of the Taj Mahal, now a perfect, mirror-smooth expanse of ice. In the background, the majestic white marble mausoleum towers over the monument, its intricate details and magnificent dome softened by the morning mist. The entire scene is bathed in a soft, ethereal morning light, casting long, soft shadows and giving the white marble a warm, pale golden hue, creating an atmosphere of profound tranquility, dreaminess, and majesty.", + "seed": 2514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2514_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": -0.07128750393167138, + "vlm_logprob": -0.07128750393167138, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.07128750393167138, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01927514374256134 + ], + "opposite_logprob_tokens": [ + -4.394275188446045 + ], + "logprob_reward": -0.01927514374256134, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01927514374256134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12111207842826843, + 0.08116782456636429, + 0.14072099328041077, + 0.08600440621376038, + 0.09439089894294739, + 0.10319694131612778, + 0.09808356314897537, + 0.09092062711715698, + 0.09355461597442627, + 0.09084802865982056 + ], + "credit_map": [ + 0.023142945021390915, + 0.014328968711197376, + 0.02066417969763279, + 0.013735861517488956, + 0.012544144876301289, + 0.01052115298807621, + 0.014911245554685593, + 0.011263575404882431, + 0.018023166805505753, + 0.01142473891377449, + 0.01626531407237053, + 0.007447215262800455, + 0.007403723429888487, + 0.010722564533352852, + 0.0044285631738603115, + 0.005452536977827549, + 0.021320264786481857, + 0.027158189564943314, + 0.01777634024620056, + 0.021058091893792152, + 0.013819900341331959, + 0.019130030646920204, + 0.011195428669452667, + 0.009262746199965477, + 0.009038923308253288, + 0.016529608517885208, + 0.01043378934264183, + 0.011329240165650845, + 0.009507975541055202, + 0.010781015269458294, + 0.009555249474942684, + 0.008828599937260151, + 0.012328104116022587, + 0.011242887936532497, + 0.02464328147470951, + 0.011908803135156631, + 0.006474944297224283, + 0.01004334632307291, + 0.007269634865224361, + 0.010479898191988468, + 0.015456742607057095, + 0.01748577505350113, + 0.011735630221664906, + 0.01584765315055847, + 0.009599617682397366, + 0.012163447216153145, + 0.010204040445387363, + 0.010704037733376026, + 0.011501053348183632, + 0.01615527831017971, + 0.01133031491190195, + 0.011399303562939167, + 0.006743155419826508, + 0.010662012733519077, + 0.017055969685316086, + 0.013236467726528645, + 0.011634530499577522, + 0.01033609826117754, + 0.010364905931055546, + 0.019909802824258804, + 0.010005511343479156, + 0.009614476934075356, + 0.008382529951632023, + 0.010672771371901035, + 0.010379399172961712, + 0.005708687007427216, + 0.01977819949388504, + 0.009463569149374962, + 0.008623408153653145, + 0.010045066475868225, + 0.01935403421521187, + 0.01020225789397955, + 0.011154725216329098, + 0.008296244777739048, + 0.013298263773322105, + 0.018611285835504532, + 0.0073341247625648975, + 0.010290822945535183, + 0.010756542906165123, + 0.011106017045676708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.01927514374256134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553953565657139 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012553953565657139, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553953565657139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12772884964942932, + 0.10222803056240082, + 0.034204237163066864, + 0.0587928481400013, + 0.14060385525226593, + 0.13305968046188354, + 0.056951820850372314, + 0.13955330848693848, + 0.11251065135002136, + 0.09436677396297455 + ], + "credit_map": [ + 0.02076038345694542, + 0.016331171616911888, + 0.017584340646862984, + 0.017012938857078552, + 0.015804169699549675, + 0.01863815449178219, + 0.008521766401827335, + 0.01307592261582613, + 0.022272173315286636, + 0.010842734947800636, + 0.019254416227340698, + 0.011217155493795872, + 0.009343528188765049, + 0.014050873927772045, + 0.006093059200793505, + 0.009154083207249641, + 0.004607671406120062, + 0.0037220020312815905, + 0.004683258011937141, + 0.0032929859589785337, + 0.005914561450481415, + 0.0051018064841628075, + 0.0031617647036910057, + 0.0037201850209385157, + 0.005917517933994532, + 0.010699012316763401, + 0.00696009024977684, + 0.007552792318165302, + 0.007963445037603378, + 0.007256028708070517, + 0.006187708117067814, + 0.006256255321204662, + 0.01612873747944832, + 0.02770286239683628, + 0.014687290415167809, + 0.03375456854701042, + 0.008687163703143597, + 0.013319971971213818, + 0.009955114684998989, + 0.01636815257370472, + 0.016580011695623398, + 0.015197151340544224, + 0.019810354337096214, + 0.02218632772564888, + 0.013688510283827782, + 0.016330070793628693, + 0.015107777900993824, + 0.014159469865262508, + 0.0070728641003370285, + 0.007414829917252064, + 0.005711870733648539, + 0.007127240300178528, + 0.005056903697550297, + 0.005870089400559664, + 0.008558239787817001, + 0.010139781050384045, + 0.014739744365215302, + 0.021366098895668983, + 0.016602296382188797, + 0.029798351228237152, + 0.010664068162441254, + 0.017158055678009987, + 0.013395764864981174, + 0.015828927978873253, + 0.01467287726700306, + 0.012166098691523075, + 0.023682137951254845, + 0.01166109275072813, + 0.0119492057710886, + 0.014912773855030537, + 0.011858620680868626, + 0.011607845313847065, + 0.013417323119938374, + 0.00823194719851017, + 0.02431691437959671, + 0.012392601929605007, + 0.00795230083167553, + 0.011546723544597626, + 0.005956749897450209, + 0.010552218183875084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.012553953565657139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059547401033341885 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.0059547401033341885, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059547401033341885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1750260889530182, + 0.10607700794935226, + 0.09111467003822327, + 0.05902601405978203, + 0.06057457998394966, + 0.06584206223487854, + 0.04622604697942734, + 0.09318151324987411, + 0.10772991925477982, + 0.1952020227909088 + ], + "credit_map": [ + 0.018167313188314438, + 0.024361947551369667, + 0.012598549015820026, + 0.014042680151760578, + 0.029949653893709183, + 0.03955758735537529, + 0.01672583818435669, + 0.01962253265082836, + 0.02415793016552925, + 0.016049867495894432, + 0.01791059970855713, + 0.01059925276786089, + 0.011045623570680618, + 0.015620915219187737, + 0.005174530204385519, + 0.005518305581063032, + 0.011882208287715912, + 0.008488456718623638, + 0.010192873887717724, + 0.010030384175479412, + 0.013311502523720264, + 0.011960597708821297, + 0.015516577288508415, + 0.0097320806235075, + 0.008548140525817871, + 0.008505293168127537, + 0.006637244950979948, + 0.007619486190378666, + 0.006654394790530205, + 0.0077942730858922005, + 0.007206874433904886, + 0.006060314830392599, + 0.00695387739688158, + 0.008703828789293766, + 0.013137795962393284, + 0.007415917236357927, + 0.006225611548870802, + 0.007178470492362976, + 0.004879618063569069, + 0.006079470738768578, + 0.007705367635935545, + 0.008385982364416122, + 0.008675704710185528, + 0.011225167661905289, + 0.007109761703759432, + 0.007293453440070152, + 0.0074834516271948814, + 0.007963184267282486, + 0.005110676400363445, + 0.003946640528738499, + 0.005931383930146694, + 0.005297554191201925, + 0.004450646694749594, + 0.005951062776148319, + 0.006768342573195696, + 0.0087697459384799, + 0.011702675372362137, + 0.01356646977365017, + 0.013370916247367859, + 0.01678280718624592, + 0.008929063566029072, + 0.009917758405208588, + 0.009221374057233334, + 0.009690461680293083, + 0.013257218524813652, + 0.007465766277164221, + 0.014843765646219254, + 0.017455309629440308, + 0.011044084094464779, + 0.013484416529536247, + 0.01843774877488613, + 0.011741620488464832, + 0.028161268681287766, + 0.014317474327981472, + 0.059754401445388794, + 0.029280319809913635, + 0.009847702458500862, + 0.023710452020168304, + 0.012351790443062782, + 0.017778633162379265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0059547401033341885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525260902941227 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.00525260902941227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525260902941227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2175873965024948, + 0.06212618947029114, + 0.06708924472332001, + 0.05594101548194885, + 0.08295807242393494, + 0.06514877080917358, + 0.06525397300720215, + 0.06003790721297264, + 0.18054544925689697, + 0.14331194758415222 + ], + "credit_map": [ + 0.015032568946480751, + 0.03149481117725372, + 0.014833161607384682, + 0.022248033434152603, + 0.030578486621379852, + 0.052611544728279114, + 0.022078262642025948, + 0.028710531070828438, + 0.00827062502503395, + 0.004793722648173571, + 0.00787436030805111, + 0.00802445225417614, + 0.009278824552893639, + 0.00776448380202055, + 0.00674866558983922, + 0.00937105156481266, + 0.00856336671859026, + 0.006394657772034407, + 0.008496686816215515, + 0.007996668107807636, + 0.013170238584280014, + 0.008396903052926064, + 0.009128162637352943, + 0.0049425638280808926, + 0.008192149922251701, + 0.005902659147977829, + 0.006187798921018839, + 0.006570969242602587, + 0.009144922718405724, + 0.007210446521639824, + 0.007355737965553999, + 0.005376326851546764, + 0.01322415005415678, + 0.01292857713997364, + 0.011097511276602745, + 0.011989660561084747, + 0.009149539284408092, + 0.01101284846663475, + 0.007194133475422859, + 0.006361653096973896, + 0.0063648889772593975, + 0.006506005767732859, + 0.013168261386454105, + 0.005784285720437765, + 0.013646556995809078, + 0.006643565837293863, + 0.0072249979712069035, + 0.005810212343931198, + 0.007748184725642204, + 0.008130067959427834, + 0.010554870590567589, + 0.006830750964581966, + 0.006436975672841072, + 0.007532432675361633, + 0.007629066240042448, + 0.010391626507043839, + 0.007170095574110746, + 0.010569513775408268, + 0.008834145963191986, + 0.008047904819250107, + 0.006039525382220745, + 0.006976765114814043, + 0.005092702340334654, + 0.007307251915335655, + 0.026383766904473305, + 0.016096064820885658, + 0.04702577367424965, + 0.015063537284731865, + 0.017971813678741455, + 0.02560943178832531, + 0.013143220916390419, + 0.019251849502325058, + 0.021070919930934906, + 0.014502678997814655, + 0.02762031927704811, + 0.01852354034781456, + 0.01294583547860384, + 0.023237235844135284, + 0.01134207472205162, + 0.014069341123104095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.00525260902941227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31340107321739197 + ], + "opposite_logprob_tokens": [ + -1.3134011030197144 + ], + "logprob_reward": -0.31340107321739197, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31340107321739197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11086714267730713, + 0.040064845234155655, + 0.15975627303123474, + 0.05831686779856682, + 0.09238062798976898, + 0.042339690029621124, + 0.09907756000757217, + 0.11690446734428406, + 0.183890238404274, + 0.09640231728553772 + ], + "credit_map": [ + 0.017616931349039078, + 0.01516471616923809, + 0.013781187124550343, + 0.013835462741553783, + 0.018694574013352394, + 0.013703981414437294, + 0.008556540124118328, + 0.009513746947050095, + 0.006302551832050085, + 0.004950265865772963, + 0.007089454680681229, + 0.004961340222507715, + 0.005068537779152393, + 0.0048265717923641205, + 0.0031447578221559525, + 0.0037213589530438185, + 0.022781772539019585, + 0.022919176146388054, + 0.022664949297904968, + 0.020926583558321, + 0.014230542816221714, + 0.021300652995705605, + 0.02027140185236931, + 0.014661183580756187, + 0.006782206241041422, + 0.01188903208822012, + 0.005940787959843874, + 0.006775816436856985, + 0.007871952839195728, + 0.007820181548595428, + 0.00642919447273016, + 0.004807686898857355, + 0.014097996987402439, + 0.009106292389333248, + 0.011224945075809956, + 0.012035823427140713, + 0.00843733735382557, + 0.012993239797651768, + 0.00780480308458209, + 0.016680177301168442, + 0.0055449120700359344, + 0.006938247941434383, + 0.005652653519064188, + 0.0060961004346609116, + 0.003799454541876912, + 0.004190989304333925, + 0.005587107036262751, + 0.004530221223831177, + 0.012077848426997662, + 0.022603239864110947, + 0.00930024590343237, + 0.01670331321656704, + 0.009092985652387142, + 0.01067015714943409, + 0.009619698859751225, + 0.00901006069034338, + 0.013880839571356773, + 0.014776532538235188, + 0.01151321455836296, + 0.02198966033756733, + 0.01462472788989544, + 0.012666413560509682, + 0.011782591231167316, + 0.01567048206925392, + 0.025187944993376732, + 0.016501791775226593, + 0.042389050126075745, + 0.02342749945819378, + 0.017036547884345055, + 0.024694684892892838, + 0.01719900593161583, + 0.017453711479902267, + 0.013457045890390873, + 0.009705539792776108, + 0.013002496212720871, + 0.01810108684003353, + 0.007677020505070686, + 0.013823306187987328, + 0.008071234449744225, + 0.01256458181887865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.31340107321739197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins, the camera tracking smoothly horizontally from left to right, following a lone figure gliding across the vast frozen surface. The figure is a woman, her dark hair pulled back in a tight, elegant bun, and her movements graceful. She wears a deep red, long-sleeved velvet skating dress, the rich fabric striking as it ripples with her movements. Her white skates, adorned with gleaming silver blades, carve silent, clean lines across the ice. She glides effortlessly, her arms extended in a balanced, fluid pose, her expression calm and focused. The camera maintains an eye-level perspective, emphasizing the seamless connection between the skater and her environment: the iconic reflecting pool in front of the Taj Mahal, now a perfect, mirror-smooth expanse of ice. In the background, the majestic white marble mausoleum towers over the monument, its intricate details and magnificent dome softened by the morning mist. The entire scene is bathed in a soft, ethereal morning light, casting long, soft shadows and giving the white marble a warm, pale golden hue, creating an atmosphere of profound tranquility, dreaminess, and majesty.", + "seed": 2515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_7/index_5812_seed_2515_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": -0.08661275152117014, + "vlm_logprob": -0.08661275152117014, + "vlm_yes_no_margin": 3.424999952316284, + "vlm_token_credit": -0.08661275152117014, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03893212601542473 + ], + "opposite_logprob_tokens": [ + -3.4139320850372314 + ], + "logprob_reward": -0.03893212601542473, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03893212601542473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3080083727836609, + 0.13173052668571472, + 0.09291967749595642, + 0.07841861993074417, + 0.03068036213517189, + 0.04452884569764137, + 0.044550418853759766, + 0.0846552848815918, + 0.09198634326457977, + 0.09252162277698517 + ], + "credit_map": [ + 0.033730048686265945, + 0.03032713569700718, + 0.0367683544754982, + 0.03481823205947876, + 0.03500950708985329, + 0.040794871747493744, + 0.04169749096035957, + 0.054862696677446365, + 0.01848747208714485, + 0.025844253599643707, + 0.009656663052737713, + 0.01437887642532587, + 0.009143047034740448, + 0.03072628565132618, + 0.010085917077958584, + 0.013408005237579346, + 0.015119527466595173, + 0.008847745135426521, + 0.012823181226849556, + 0.01044875755906105, + 0.008501313626766205, + 0.012369568459689617, + 0.01282963901758194, + 0.011979933828115463, + 0.007443649228662252, + 0.008929762989282608, + 0.014571912586688995, + 0.008991950191557407, + 0.009460221976041794, + 0.009078379720449448, + 0.007088548969477415, + 0.01285418402403593, + 0.0031989840790629387, + 0.0040685804560780525, + 0.003088831901550293, + 0.002762798685580492, + 0.00551608856767416, + 0.004426542203873396, + 0.004239786881953478, + 0.0033787472639232874, + 0.004029485397040844, + 0.006484705954790115, + 0.005973820108920336, + 0.0050142561085522175, + 0.0044992160983383656, + 0.005890564993023872, + 0.005908002611249685, + 0.006728787906467915, + 0.003965442534536123, + 0.005555734969675541, + 0.00527985068038106, + 0.00642692856490612, + 0.005035970825701952, + 0.006540697533637285, + 0.0069728754460811615, + 0.004772914573550224, + 0.008056487888097763, + 0.013758890330791473, + 0.0062187435105443, + 0.008552498184144497, + 0.01378743164241314, + 0.012229976244270802, + 0.01472651120275259, + 0.007324742153286934, + 0.005452652927488089, + 0.012040332891047001, + 0.009560039266943932, + 0.020934930071234703, + 0.009761668741703033, + 0.015594270080327988, + 0.010435412637889385, + 0.008207025937736034, + 0.008256054483354092, + 0.012114966288208961, + 0.015578621998429298, + 0.019878217950463295, + 0.007429372053593397, + 0.010740012861788273, + 0.009360088035464287, + 0.009164273738861084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03893212601542473 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060099132359028 + ], + "opposite_logprob_tokens": [ + -3.8956010341644287 + ], + "logprob_reward": -0.02060099132359028, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060099132359028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3244263529777527, + 0.19593088328838348, + 0.0710105150938034, + 0.04984571412205696, + 0.10869565606117249, + 0.031815532594919205, + 0.03620271384716034, + 0.07031939923763275, + 0.02931865304708481, + 0.08243461698293686 + ], + "credit_map": [ + 0.03846070542931557, + 0.03529948368668556, + 0.04625626280903816, + 0.03672442585229874, + 0.03401624783873558, + 0.03345411270856857, + 0.03743769973516464, + 0.06277737021446228, + 0.01770755834877491, + 0.02154363878071308, + 0.01707475446164608, + 0.019594253972172737, + 0.02958099916577339, + 0.03678132966160774, + 0.03977494314312935, + 0.013873395510017872, + 0.008472229354083538, + 0.010228882543742657, + 0.007449722848832607, + 0.008033506572246552, + 0.012823382392525673, + 0.009051229804754257, + 0.00812781322747469, + 0.00682373670861125, + 0.008162949234247208, + 0.0061760107055306435, + 0.008685297332704067, + 0.004438637290149927, + 0.006115337833762169, + 0.005645725876092911, + 0.004104331601411104, + 0.006517416797578335, + 0.02659202553331852, + 0.010714310221374035, + 0.019467473030090332, + 0.006948172114789486, + 0.007365166209638119, + 0.0176884476095438, + 0.007106829900294542, + 0.012813206762075424, + 0.0025586411356925964, + 0.004516826011240482, + 0.005238156765699387, + 0.0042001064866781235, + 0.003921867813915014, + 0.004219376016408205, + 0.004042643588036299, + 0.0031179101206362247, + 0.0040936823934316635, + 0.0052633630111813545, + 0.0025898765306919813, + 0.004126197658479214, + 0.004548394121229649, + 0.006296174600720406, + 0.005607993807643652, + 0.003677023807540536, + 0.004831176251173019, + 0.011717108078300953, + 0.0038806151133030653, + 0.007708752062171698, + 0.01350813265889883, + 0.009972866624593735, + 0.012473064474761486, + 0.006227670703083277, + 0.0021107541397213936, + 0.0037600495852530003, + 0.0029940172098577023, + 0.00640599150210619, + 0.00322741805575788, + 0.005107795353978872, + 0.0036553589161485434, + 0.0020572643261402845, + 0.007278958801180124, + 0.010421596467494965, + 0.01503013912588358, + 0.016763973981142044, + 0.007013808004558086, + 0.008714894764125347, + 0.009345914237201214, + 0.007865315303206444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02060099132359028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741280667483807 + ], + "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": [ + 0.43413007259368896, + 0.031572047621011734, + 0.09740759432315826, + 0.09117691963911057, + 0.052836015820503235, + 0.04068756476044655, + 0.04739172011613846, + 0.04858400672674179, + 0.08490423113107681, + 0.07130984961986542 + ], + "credit_map": [ + 0.0503821074962616, + 0.04075082391500473, + 0.05745533108711243, + 0.04353255406022072, + 0.04658534377813339, + 0.06722314655780792, + 0.04459340497851372, + 0.08360737562179565, + 0.004755991976708174, + 0.0046411133371293545, + 0.003558518597856164, + 0.004184097517281771, + 0.003466906026005745, + 0.00491933012381196, + 0.0030419633258134127, + 0.003004125552251935, + 0.016597967594861984, + 0.012203955091536045, + 0.013818559236824512, + 0.013245750218629837, + 0.006976452190428972, + 0.012910529039800167, + 0.00852877739816904, + 0.013125602155923843, + 0.014810080640017986, + 0.009255693294107914, + 0.02095320075750351, + 0.009185205213725567, + 0.006337644997984171, + 0.011041730642318726, + 0.006602287292480469, + 0.012991080060601234, + 0.009553903713822365, + 0.005635734647512436, + 0.008833571337163448, + 0.003641756484284997, + 0.004557991400361061, + 0.007287784479558468, + 0.006264586467295885, + 0.00706068892031908, + 0.006744953338056803, + 0.005376983899623156, + 0.005859003867954016, + 0.003526626620441675, + 0.004062919411808252, + 0.004894350655376911, + 0.004971198737621307, + 0.005251531023532152, + 0.00395080354064703, + 0.0064329225569963455, + 0.0056161959655582905, + 0.005336070898920298, + 0.00521136075258255, + 0.007093369960784912, + 0.007856390438973904, + 0.005894607864320278, + 0.005395258776843548, + 0.007439278997480869, + 0.0034259967505931854, + 0.005474244710057974, + 0.0076490845531225204, + 0.007035335060209036, + 0.00753316143527627, + 0.004631648305803537, + 0.005635042674839497, + 0.011874774470925331, + 0.008588084951043129, + 0.020144619047641754, + 0.008893025107681751, + 0.014141233637928963, + 0.005944033153355122, + 0.009683418087661266, + 0.012848698534071445, + 0.009180151857435703, + 0.015727082267403603, + 0.0073282052762806416, + 0.006474383175373077, + 0.007545225787907839, + 0.0034248537849634886, + 0.008781250566244125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.006741280667483807 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185783177614212 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185783177614212, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018185783177614212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18361647427082062, + 0.06109727919101715, + 0.07018804550170898, + 0.06713894009590149, + 0.23689895868301392, + 0.04318591579794884, + 0.05275580286979675, + 0.07918612658977509, + 0.12476899474859238, + 0.08116339892148972 + ], + "credit_map": [ + 0.02807907573878765, + 0.02271890826523304, + 0.02263742685317993, + 0.018382715061306953, + 0.014832704328000546, + 0.039178457111120224, + 0.015121222473680973, + 0.022665956988930702, + 0.006023645866662264, + 0.011913913302123547, + 0.006939706392586231, + 0.008497169241309166, + 0.006978667341172695, + 0.00870896503329277, + 0.00511395838111639, + 0.0069212536327540874, + 0.00816434994339943, + 0.008550257422029972, + 0.007959484122693539, + 0.01397737581282854, + 0.006783469580113888, + 0.007675390690565109, + 0.009145354852080345, + 0.007932365871965885, + 0.010050676763057709, + 0.009901510551571846, + 0.00916322786360979, + 0.00684227142482996, + 0.00950125977396965, + 0.007974977605044842, + 0.007814059965312481, + 0.005890960339456797, + 0.05534413829445839, + 0.02728321962058544, + 0.038557276129722595, + 0.013957226648926735, + 0.019818350672721863, + 0.03707871586084366, + 0.015457694418728352, + 0.029402317479252815, + 0.005504765547811985, + 0.00748648913577199, + 0.004716367926448584, + 0.003983271308243275, + 0.004662772174924612, + 0.006285632960498333, + 0.005975309759378433, + 0.004571306984871626, + 0.00492345355451107, + 0.008310855366289616, + 0.004419075790792704, + 0.0066724070347845554, + 0.006105032283812761, + 0.00977453961968422, + 0.005968919955193996, + 0.0065815201960504055, + 0.008566158823668957, + 0.011120915412902832, + 0.00798480212688446, + 0.016658902168273926, + 0.008655397221446037, + 0.010535393841564655, + 0.007021894212812185, + 0.00864266324788332, + 0.007993703708052635, + 0.016036193817853928, + 0.012670546770095825, + 0.02362676151096821, + 0.014995605684816837, + 0.01956397481262684, + 0.019227614626288414, + 0.010654597543179989, + 0.009860294871032238, + 0.009874613955616951, + 0.016057433560490608, + 0.011236971244215965, + 0.006190706510096788, + 0.009813039563596249, + 0.008676811121404171, + 0.00945352390408516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.018185783177614212 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a woman with dark hair pulled back in a tight bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a deep red, long-sleeved velvet skating dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman gliding on white skates with silver blades across a frozen surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background scene the Taj Mahal with its white marble mausoleum and dome?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly horizontally from left to right at an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34860357642173767 + ], + "opposite_logprob_tokens": [ + -1.22360360622406 + ], + "logprob_reward": -0.34860357642173767, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34860357642173767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38707900047302246, + 0.0518948994576931, + 0.023910395801067352, + 0.04949427768588066, + 0.22306303679943085, + 0.029047483578324318, + 0.03158072009682655, + 0.03533298522233963, + 0.14302250742912292, + 0.02557465061545372 + ], + "credit_map": [ + 0.035883985459804535, + 0.050369665026664734, + 0.03796733543276787, + 0.04280788078904152, + 0.05375662446022034, + 0.0636625811457634, + 0.04959181696176529, + 0.05303915590047836, + 0.008649555966258049, + 0.007365746423602104, + 0.004824046976864338, + 0.005270581692457199, + 0.007321021053940058, + 0.007706279866397381, + 0.004982090555131435, + 0.005775580648332834, + 0.002495920518413186, + 0.004097091965377331, + 0.0021543961483985186, + 0.0042833928018808365, + 0.0035581965930759907, + 0.0030900912825018167, + 0.0024193543940782547, + 0.001811952330172062, + 0.004908578470349312, + 0.008388801477849483, + 0.005167216528207064, + 0.005989250261336565, + 0.008367843925952911, + 0.006176961585879326, + 0.006401928141713142, + 0.0040937019512057304, + 0.05253644287586212, + 0.020274203270673752, + 0.04321082681417465, + 0.011946451850235462, + 0.014489078894257545, + 0.036847859621047974, + 0.014320392161607742, + 0.029437808319926262, + 0.0022424322087317705, + 0.0055160364136099815, + 0.003328756196424365, + 0.0033734305761754513, + 0.003539494937285781, + 0.004119927063584328, + 0.0038782700430601835, + 0.003049137769266963, + 0.0034257101360708475, + 0.004335208795964718, + 0.002838333137333393, + 0.004485287703573704, + 0.004089022055268288, + 0.0045160287991166115, + 0.0040956237353384495, + 0.003795509459450841, + 0.0028638767544180155, + 0.005536504555493593, + 0.0021925291512161493, + 0.004116047639399767, + 0.0061904010362923145, + 0.0050649275071918964, + 0.006610445212572813, + 0.0027582566253840923, + 0.008577684871852398, + 0.02077246643602848, + 0.014904018491506577, + 0.03997237607836723, + 0.012965909205377102, + 0.023636173456907272, + 0.009262734092772007, + 0.012931152246892452, + 0.004112839233130217, + 0.003039412200450897, + 0.0025493884459137917, + 0.0037360789719969034, + 0.002978675067424774, + 0.003514448646456003, + 0.002760298317298293, + 0.0028835118282586336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.34860357642173767 + } + ], + "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/0000155/group_8/index_tensor([6803, 6803])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_8/index_tensor([6803, 6803])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..113bc4fb2ba28b6063371daa4d0d6be7799512d3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_8/index_tensor([6803, 6803])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, extreme close-up captures a dense cluster of delicate, bell-shaped inkcap mushrooms sprouting from the damp, moss-covered forest floor. The scene is plunged into dramatic lowlight, with a single focused beam of light piercing the darkness, focusing the fungi and creating a powerful black-and-white contrast. In a mesmerizing time-lapse sequence, the mushrooms begin to dissolve; their pale, grooved caps, initially rimmed with glistening dewdrops, begin to curl upward and inexorably. A thick, inky black substance forms around the edges, dripping viscously onto the decaying leaves below. Each drop catches the light, appearing like a living, dripping globule of black pigment before coalescing into a dark, lustrous puddle that slowly stains the earth. The entire process is a hypnotic spectacle of natural decay, transforming ephemeral fungi into ephemeral works of ink on the forest floor, creating an atmosphere both eerie and beautiful.", + "seed": 30735, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_8/index_6803_seed_30735_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": -0.019672558642923832, + "vlm_logprob": -0.019672558642923832, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.019672558642923832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035898737609386444 + ], + "opposite_logprob_tokens": [ + -3.4108986854553223 + ], + "logprob_reward": -0.035898737609386444, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.035898737609386444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1333189457654953, + 0.31884491443634033, + 0.06176474317908287, + 0.049346670508384705, + 0.07407835125923157, + 0.09239724278450012, + 0.09320173412561417, + 0.058690156787633896, + 0.06397329270839691, + 0.05438397452235222 + ], + "credit_map": [ + 0.012637170031666756, + 0.015556366182863712, + 0.01637152023613453, + 0.02147572860121727, + 0.013630285859107971, + 0.01653359644114971, + 0.017653409391641617, + 0.01946086250245571, + 0.03320390731096268, + 0.03175388649106026, + 0.0449901819229126, + 0.037507314234972, + 0.040481358766555786, + 0.041343964636325836, + 0.042126864194869995, + 0.04743741825222969, + 0.007338234689086676, + 0.005651424173265696, + 0.010171136818826199, + 0.010163122788071632, + 0.006621973123401403, + 0.006219128146767616, + 0.006633785553276539, + 0.008965934626758099, + 0.00587904779240489, + 0.004338405095040798, + 0.0076194037683308125, + 0.009129318408668041, + 0.005941782146692276, + 0.004731603432446718, + 0.004247706383466721, + 0.007459397427737713, + 0.00762898987159133, + 0.007505877409130335, + 0.011555886827409267, + 0.011205832473933697, + 0.008786492049694061, + 0.007308316417038441, + 0.008585730567574501, + 0.011501220054924488, + 0.010197152383625507, + 0.005904202349483967, + 0.014335032552480698, + 0.01658867858350277, + 0.008803896605968475, + 0.01546905841678381, + 0.007400219794362783, + 0.013698998838663101, + 0.012011797167360783, + 0.007808967027813196, + 0.01730937883257866, + 0.01153498888015747, + 0.010857860557734966, + 0.012631440535187721, + 0.00917566753923893, + 0.01187162846326828, + 0.007116312626749277, + 0.0044280607253313065, + 0.008828561753034592, + 0.009977470152080059, + 0.005100550130009651, + 0.010535572655498981, + 0.005379893351346254, + 0.007323732133954763, + 0.009349527768790722, + 0.004150462802499533, + 0.008865492418408394, + 0.010146993212401867, + 0.006869629491120577, + 0.009927758947014809, + 0.006989108398556709, + 0.007674314081668854, + 0.004739320371299982, + 0.0027019274421036243, + 0.0061330972239375114, + 0.008958972059190273, + 0.005227761808782816, + 0.012376300059258938, + 0.00635081622749567, + 0.007895774208009243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.035898737609386444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744596175849438 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744596175849438, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744596175849438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13599540293216705, + 0.2508687973022461, + 0.03749123215675354, + 0.055462222546339035, + 0.13089346885681152, + 0.08271539956331253, + 0.08707261085510254, + 0.09627317637205124, + 0.08582264930009842, + 0.03740504011511803 + ], + "credit_map": [ + 0.010613864287734032, + 0.014189044944941998, + 0.02745787613093853, + 0.015631506219506264, + 0.01001798827201128, + 0.020396189764142036, + 0.015439353883266449, + 0.022249581292271614, + 0.03078141063451767, + 0.023289795964956284, + 0.028957972303032875, + 0.03164960816502571, + 0.0316079817712307, + 0.04076848924160004, + 0.02859063446521759, + 0.035222940146923065, + 0.004078557249158621, + 0.003475476987659931, + 0.007030448876321316, + 0.004500214476138353, + 0.004623082000762224, + 0.004006247967481613, + 0.004574453923851252, + 0.005202747881412506, + 0.006159127224236727, + 0.007434550207108259, + 0.008152014575898647, + 0.007748560979962349, + 0.007248499896377325, + 0.006299067288637161, + 0.0051637133583426476, + 0.007256693206727505, + 0.02020491659641266, + 0.009883551858365536, + 0.021890459582209587, + 0.024833563715219498, + 0.010574289597570896, + 0.01161075197160244, + 0.014523397199809551, + 0.017372535541653633, + 0.008274676278233528, + 0.00556569080799818, + 0.019029861316084862, + 0.0163214523345232, + 0.007382550742477179, + 0.007181897293776274, + 0.008697465993463993, + 0.010261808522045612, + 0.006932549178600311, + 0.01275104470551014, + 0.010335967876017094, + 0.011423091404139996, + 0.013357972726225853, + 0.008736804127693176, + 0.012164653278887272, + 0.0113705238327384, + 0.009501398541033268, + 0.006551130209118128, + 0.01892665959894657, + 0.015760013833642006, + 0.010419771075248718, + 0.011427008546888828, + 0.012364507652819157, + 0.011322682723402977, + 0.00810975767672062, + 0.0052394745871424675, + 0.012506677769124508, + 0.013416511937975883, + 0.00830863881856203, + 0.01069735363125801, + 0.01796245016157627, + 0.009581778198480606, + 0.003051044186577201, + 0.002626606496050954, + 0.005864291451871395, + 0.006461762823164463, + 0.005375789478421211, + 0.0037216017954051495, + 0.0050095790065824986, + 0.005294365342706442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.006744596175849438 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370582312345505 + ], + "opposite_logprob_tokens": [ + -3.408705711364746 + ], + "logprob_reward": -0.03370582312345505, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370582312345505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11750543862581253, + 0.3764193654060364, + 0.032745540142059326, + 0.03306373581290245, + 0.07108502835035324, + 0.04866969585418701, + 0.15174615383148193, + 0.07282523810863495, + 0.044236402958631516, + 0.05170341581106186 + ], + "credit_map": [ + 0.009239265695214272, + 0.01181095466017723, + 0.015329563058912754, + 0.015550522133708, + 0.01279696449637413, + 0.0149072939530015, + 0.019520504400134087, + 0.018350359052419662, + 0.03705630078911781, + 0.03714495524764061, + 0.05112266540527344, + 0.05700500309467316, + 0.04162801429629326, + 0.05185048654675484, + 0.04490318521857262, + 0.05570874735713005, + 0.004212562460452318, + 0.0019927038811147213, + 0.004295023623853922, + 0.003642723895609379, + 0.0050753261893987656, + 0.00463667931035161, + 0.004637646954506636, + 0.0042528728954494, + 0.003575560636818409, + 0.003999648150056601, + 0.005007680039852858, + 0.004258263856172562, + 0.00437705684453249, + 0.005002487450838089, + 0.002616521902382374, + 0.004226512275636196, + 0.00915441196411848, + 0.004958741366863251, + 0.010865864343941212, + 0.013307380490005016, + 0.007359660230576992, + 0.008078462444245815, + 0.008028140291571617, + 0.009332360699772835, + 0.0046592578291893005, + 0.003312267828732729, + 0.008734875358641148, + 0.007973418571054935, + 0.005460929591208696, + 0.00680524343624711, + 0.00555039569735527, + 0.006173305679112673, + 0.01533711701631546, + 0.011723877862095833, + 0.025176160037517548, + 0.018214166164398193, + 0.02679993212223053, + 0.013992791064083576, + 0.023768924176692963, + 0.016733180731534958, + 0.0071724532172083855, + 0.005416423548012972, + 0.013393865898251534, + 0.010300920344889164, + 0.010087594389915466, + 0.008889304473996162, + 0.008842723444104195, + 0.008721948601305485, + 0.006293138954788446, + 0.0023126816377043724, + 0.006754613947123289, + 0.006135866977274418, + 0.005217138212174177, + 0.006488637533038855, + 0.005667215213179588, + 0.0053671058267354965, + 0.0043929871171712875, + 0.00323602301068604, + 0.007530681788921356, + 0.010499313473701477, + 0.006905130576342344, + 0.006409550551325083, + 0.004834407940506935, + 0.00789531972259283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.03370582312345505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606742851436138 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606742851436138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606742851436138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10585444420576096, + 0.15293225646018982, + 0.08138944953680038, + 0.07368849217891693, + 0.09904665499925613, + 0.07435385137796402, + 0.06961304694414139, + 0.06528694182634354, + 0.219429150223732, + 0.05840567499399185 + ], + "credit_map": [ + 0.007929468527436256, + 0.010781415738165379, + 0.014531887136399746, + 0.017346827313303947, + 0.010904931463301182, + 0.009096518158912659, + 0.021036796271800995, + 0.014226594939827919, + 0.011276821605861187, + 0.00885302759706974, + 0.02507428079843521, + 0.023853683844208717, + 0.01381696481257677, + 0.017878418788313866, + 0.028725922107696533, + 0.023453138768672943, + 0.007491128984838724, + 0.0073668439872562885, + 0.011250587180256844, + 0.011948635801672935, + 0.009927881881594658, + 0.00971226580440998, + 0.011575520969927311, + 0.012116581201553345, + 0.007225235924124718, + 0.007603436708450317, + 0.01047712191939354, + 0.01120031625032425, + 0.008855975233018398, + 0.012126216664910316, + 0.00743937399238348, + 0.008760816417634487, + 0.011065950617194176, + 0.006389090325683355, + 0.01825765334069729, + 0.01646808162331581, + 0.009906116873025894, + 0.010844971984624863, + 0.010509910061955452, + 0.015604879707098007, + 0.008952559903264046, + 0.005860751029103994, + 0.012885740026831627, + 0.01380302757024765, + 0.008980341255664825, + 0.006495509762316942, + 0.006168177351355553, + 0.011207747273147106, + 0.008707424625754356, + 0.007795131765305996, + 0.006735437083989382, + 0.008473729714751244, + 0.008043964393436909, + 0.011232024058699608, + 0.010686453431844711, + 0.007938885129988194, + 0.006814454682171345, + 0.00514936912804842, + 0.010995151475071907, + 0.010168797336518764, + 0.007450058590620756, + 0.008227888494729996, + 0.008670826442539692, + 0.007810394745320082, + 0.019154300913214684, + 0.012602049857378006, + 0.02807493694126606, + 0.023563984781503677, + 0.025409908965229988, + 0.030564334243535995, + 0.05151787027716637, + 0.02854178100824356, + 0.006248405668884516, + 0.00560443801805377, + 0.006955234333872795, + 0.008551398292183876, + 0.007301887031644583, + 0.006053433753550053, + 0.010890527628362179, + 0.006800349336117506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.01606742851436138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946207791566849 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946207791566849, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946207791566849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25880077481269836, + 0.13490162789821625, + 0.0705818459391594, + 0.05474521592259407, + 0.06432487070560455, + 0.06205328553915024, + 0.07156483083963394, + 0.05492938309907913, + 0.15225672721862793, + 0.07584143429994583 + ], + "credit_map": [ + 0.02174985036253929, + 0.022253191098570824, + 0.033745694905519485, + 0.047818977385759354, + 0.021343795582652092, + 0.03902459889650345, + 0.03713193163275719, + 0.035732757300138474, + 0.008852649480104446, + 0.00844295509159565, + 0.02093256637454033, + 0.022686460986733437, + 0.012393489480018616, + 0.015484233386814594, + 0.02604292705655098, + 0.020066343247890472, + 0.007679745089262724, + 0.006223592907190323, + 0.00952946301549673, + 0.011282520368695259, + 0.008261226117610931, + 0.007588593754917383, + 0.009694013744592667, + 0.010322687216103077, + 0.006694696377962828, + 0.005795727018266916, + 0.00820409320294857, + 0.006427240092307329, + 0.006798314396291971, + 0.0075986068695783615, + 0.005289945751428604, + 0.007936595939099789, + 0.0060781375505030155, + 0.007247183471918106, + 0.01052532996982336, + 0.009555227123200893, + 0.007388037163764238, + 0.0066015287302434444, + 0.007571506313979626, + 0.00935792364180088, + 0.007469226140528917, + 0.006027047522366047, + 0.0077369785867631435, + 0.011212208308279514, + 0.009154552593827248, + 0.0068275462836027145, + 0.00555112399160862, + 0.008074598386883736, + 0.00822830107063055, + 0.006044756155461073, + 0.014491967856884003, + 0.0074921539053320885, + 0.011276356875896454, + 0.008798085153102875, + 0.006537333596497774, + 0.00869587529450655, + 0.0078103747218847275, + 0.0038729191292077303, + 0.006503969896584749, + 0.0069368439726531506, + 0.005463078618049622, + 0.0116721261292696, + 0.00599242839962244, + 0.006677638739347458, + 0.01911686360836029, + 0.0091717429459095, + 0.02120360918343067, + 0.015634149312973022, + 0.01307468581944704, + 0.01754169538617134, + 0.03648823872208595, + 0.02002573199570179, + 0.008039023727178574, + 0.006345709320157766, + 0.01071212999522686, + 0.011104252189397812, + 0.009587855078279972, + 0.009920450858771801, + 0.012051374651491642, + 0.008080640807747841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005946207791566849 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme close-up captures a dense cluster of delicate, bell-shaped inkcap mushrooms sprouting from the damp, moss-covered forest floor. The scene is plunged into dramatic lowlight, with a single focused beam of light piercing the darkness, focusing the fungi and creating a powerful black-and-white contrast. In a mesmerizing time-lapse sequence, the mushrooms begin to dissolve; their pale, grooved caps, initially rimmed with glistening dewdrops, begin to curl upward and inexorably. A thick, inky black substance forms around the edges, dripping viscously onto the decaying leaves below. Each drop catches the light, appearing like a living, dripping globule of black pigment before coalescing into a dark, lustrous puddle that slowly stains the earth. The entire process is a hypnotic spectacle of natural decay, transforming ephemeral fungi into ephemeral works of ink on the forest floor, creating an atmosphere both eerie and beautiful.", + "seed": 30736, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_8/index_6803_seed_30736_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": -0.016670597577467562, + "vlm_logprob": -0.016670597577467562, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.016670597577467562, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010367939248681068 + ], + "opposite_logprob_tokens": [ + -5.0103678703308105 + ], + "logprob_reward": -0.010367939248681068, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010367939248681068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1784486174583435, + 0.03292352706193924, + 0.033869750797748566, + 0.07650083303451538, + 0.026815202087163925, + 0.49470818042755127, + 0.028527757152915, + 0.031294532120227814, + 0.029792094603180885, + 0.0671195313334465 + ], + "credit_map": [ + 0.025517618283629417, + 0.01780611462891102, + 0.01944190450012684, + 0.01655404642224312, + 0.02244892343878746, + 0.03881562128663063, + 0.02439354546368122, + 0.013470823876559734, + 0.003852378809824586, + 0.0048076314851641655, + 0.004104985855519772, + 0.004155371338129044, + 0.0028585607651621103, + 0.004764196462929249, + 0.0048071881756186485, + 0.003573216963559389, + 0.0044482252560555935, + 0.0036285920068621635, + 0.006286174990236759, + 0.0052846018224954605, + 0.002356918528676033, + 0.0043596141040325165, + 0.003372632199898362, + 0.004132992122322321, + 0.004953010007739067, + 0.006705268286168575, + 0.018104728311300278, + 0.023538537323474884, + 0.004704867955297232, + 0.0073100202716887, + 0.005173690617084503, + 0.006010712590068579, + 0.0028791548684239388, + 0.00441947765648365, + 0.003080215072259307, + 0.0034695074427872896, + 0.0030166783835738897, + 0.0033561468590050936, + 0.003115898696705699, + 0.003478122176602483, + 0.01149053405970335, + 0.011564360931515694, + 0.0168163999915123, + 0.053105682134628296, + 0.01619548909366131, + 0.019992148503661156, + 0.20763571560382843, + 0.1579078584909439, + 0.002271684817969799, + 0.00606492767110467, + 0.0023908752482384443, + 0.003382815048098564, + 0.002247192896902561, + 0.0027104546315968037, + 0.00404716469347477, + 0.005412642378360033, + 0.003144467482343316, + 0.0060060713440179825, + 0.002874023513868451, + 0.003787179011851549, + 0.002556615276262164, + 0.0037213389296084642, + 0.004850966855883598, + 0.0043538701720535755, + 0.0038496763445436954, + 0.002907605143263936, + 0.003821789985522628, + 0.0053769745863974094, + 0.0025084211956709623, + 0.003441654611378908, + 0.004428635351359844, + 0.00345733854919672, + 0.008747333660721779, + 0.01063229888677597, + 0.007360528688877821, + 0.010313963517546654, + 0.005350157152861357, + 0.0073819407261908054, + 0.008100531995296478, + 0.009232775308191776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.010367939248681068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646219778805971 + ], + "opposite_logprob_tokens": [ + -5.379646301269531 + ], + "logprob_reward": -0.004646219778805971, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646219778805971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11438610404729843, + 0.10128137469291687, + 0.052543848752975464, + 0.07059422880411148, + 0.04180731996893883, + 0.4671211540699005, + 0.034510526806116104, + 0.04265233129262924, + 0.020401623100042343, + 0.054701462388038635 + ], + "credit_map": [ + 0.016534743830561638, + 0.013310212641954422, + 0.017193038016557693, + 0.01489404309540987, + 0.011338128708302975, + 0.01927248015999794, + 0.009829950518906116, + 0.012013504281640053, + 0.012368964031338692, + 0.011180086992681026, + 0.013011808507144451, + 0.01263541541993618, + 0.012808572500944138, + 0.015501918271183968, + 0.01256803423166275, + 0.011206574738025665, + 0.005804707761853933, + 0.005746255163103342, + 0.00930717121809721, + 0.009333420544862747, + 0.004171030130237341, + 0.005636473186314106, + 0.005948401987552643, + 0.006596387829631567, + 0.0052888779900968075, + 0.008754801005125046, + 0.01260333601385355, + 0.01626560091972351, + 0.00430010911077261, + 0.0075741056352853775, + 0.006109994370490313, + 0.009697403758764267, + 0.0023645961191505194, + 0.005753870587795973, + 0.007057401351630688, + 0.012700626626610756, + 0.0023312619887292385, + 0.004560711793601513, + 0.003462312975898385, + 0.0035765382926911116, + 0.01281096413731575, + 0.018327247351408005, + 0.020832667127251625, + 0.05357282608747482, + 0.01708691380918026, + 0.015589072369039059, + 0.18956537544727325, + 0.13933609426021576, + 0.0037045262288302183, + 0.005180638283491135, + 0.004843186587095261, + 0.0042409515008330345, + 0.003243346232920885, + 0.003913558088243008, + 0.004563960712403059, + 0.004820357542484999, + 0.005957080516964197, + 0.004862063564360142, + 0.004737803712487221, + 0.0053573111072182655, + 0.004143018275499344, + 0.004029174335300922, + 0.006710751447826624, + 0.006855127401649952, + 0.002943983068689704, + 0.0024547413922846317, + 0.00235298671759665, + 0.0027461734134703875, + 0.0017177755944430828, + 0.0021000343840569258, + 0.003746805479750037, + 0.002339123049750924, + 0.006044700276106596, + 0.006391413975507021, + 0.010584456846117973, + 0.008014746010303497, + 0.004010043106973171, + 0.006428481545299292, + 0.006639560218900442, + 0.0065880571492016315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.004646219778805971 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865977540612221 + ], + "opposite_logprob_tokens": [ + -3.0486598014831543 + ], + "logprob_reward": -0.04865977540612221, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865977540612221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08030979335308075, + 0.03743685409426689, + 0.05933915823698044, + 0.04421191290020943, + 0.03728153929114342, + 0.577302098274231, + 0.02089289203286171, + 0.0440492257475853, + 0.0404561422765255, + 0.0587204247713089 + ], + "credit_map": [ + 0.011074408888816833, + 0.011557755060493946, + 0.014855891466140747, + 0.009801213629543781, + 0.007930480875074863, + 0.009516641497612, + 0.007759764790534973, + 0.007813639007508755, + 0.0051187933422625065, + 0.004404350183904171, + 0.004276798106729984, + 0.005303347483277321, + 0.004322250839322805, + 0.004410363733768463, + 0.005550709553062916, + 0.004050238989293575, + 0.007201756350696087, + 0.006922393571585417, + 0.00916772335767746, + 0.009436764754354954, + 0.005786187015473843, + 0.007378065027296543, + 0.007047215010970831, + 0.0063990517519414425, + 0.004168136045336723, + 0.0036283922381699085, + 0.010203427635133266, + 0.013449459336698055, + 0.002442833036184311, + 0.0033284032251685858, + 0.002781851217150688, + 0.004209411330521107, + 0.0019669104367494583, + 0.006441635079681873, + 0.006342921406030655, + 0.010186171159148216, + 0.0024187457747757435, + 0.0037215647753328085, + 0.002984754042699933, + 0.0032188373152166605, + 0.01634857803583145, + 0.017661202698946, + 0.023523865267634392, + 0.06660904735326767, + 0.02037455514073372, + 0.01798972114920616, + 0.23805281519889832, + 0.176742285490036, + 0.0020274813286960125, + 0.004563868045806885, + 0.0020085908472537994, + 0.0026991821359843016, + 0.001419359934516251, + 0.002453436143696308, + 0.0027848309837281704, + 0.002936143195256591, + 0.005671374034136534, + 0.005939303897321224, + 0.00513877160847187, + 0.006059631239622831, + 0.003189929062500596, + 0.005698873661458492, + 0.007138218265026808, + 0.005213121883571148, + 0.006362724117934704, + 0.004687917418777943, + 0.003866841085255146, + 0.006610702723264694, + 0.003100717905908823, + 0.0038270559161901474, + 0.007067754864692688, + 0.004932427778840065, + 0.00677773542702198, + 0.007976670749485493, + 0.01317931804805994, + 0.007876372896134853, + 0.004312816541641951, + 0.006194170564413071, + 0.00526124844327569, + 0.007142090704292059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.04865977540612221 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064848750829697 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064848750829697, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064848750829697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03101908043026924, + 0.05196176841855049, + 0.02978474460542202, + 0.03508199006319046, + 0.027360109612345695, + 0.7365580201148987, + 0.02637152373790741, + 0.030261972919106483, + 0.01480163261294365, + 0.016799136996269226 + ], + "credit_map": [ + 0.004907749127596617, + 0.0029944174457341433, + 0.00345033360645175, + 0.003753621131181717, + 0.0037012766115367413, + 0.0049050068482756615, + 0.004087419714778662, + 0.003219258738681674, + 0.007009671535342932, + 0.00650166068226099, + 0.00770532013848424, + 0.005434815771877766, + 0.006782789248973131, + 0.0076504661701619625, + 0.0059236339293420315, + 0.004953411407768726, + 0.004070247057825327, + 0.0031173215247690678, + 0.0040338472463190556, + 0.004426150117069483, + 0.00315794232301414, + 0.0033738084603101015, + 0.004434973932802677, + 0.0031704565044492483, + 0.0036575784906744957, + 0.0034954252187162638, + 0.007001918274909258, + 0.009968637488782406, + 0.0017441073432564735, + 0.0031269430182874203, + 0.0025972817093133926, + 0.0034901017788797617, + 0.0024023556616157293, + 0.004197222646325827, + 0.003058505477383733, + 0.0053824796341359615, + 0.00198451429605484, + 0.0038838540203869343, + 0.0026553173083812, + 0.003795862663537264, + 0.019662806764245033, + 0.024926548823714256, + 0.029957693070173264, + 0.07985629886388779, + 0.025443241000175476, + 0.030411895364522934, + 0.300008624792099, + 0.22629088163375854, + 0.0019929870031774044, + 0.0051792217418551445, + 0.0030614626593887806, + 0.003487786278128624, + 0.0018356944201514125, + 0.002873981837183237, + 0.00351731083355844, + 0.004423083737492561, + 0.003008734667673707, + 0.0038985218852758408, + 0.0030826879665255547, + 0.003964447416365147, + 0.0022095933090895414, + 0.0036254767328500748, + 0.005581486038863659, + 0.004891027230769396, + 0.0017944948049262166, + 0.0013149171136319637, + 0.0025355981197208166, + 0.001837396644987166, + 0.0013026841916143894, + 0.0012529530795291066, + 0.002905936911702156, + 0.0018576522124931216, + 0.0020264512859284878, + 0.0032360025215893984, + 0.0018455958925187588, + 0.0026638666167855263, + 0.0010906567331403494, + 0.0014441987732425332, + 0.002007672330364585, + 0.0024846945889294147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.016064848750829697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11865061521530151, + 0.07813692092895508, + 0.03861292079091072, + 0.07750590890645981, + 0.02859245426952839, + 0.47972944378852844, + 0.03736412152647972, + 0.0675303265452385, + 0.023376913741230965, + 0.05050037056207657 + ], + "credit_map": [ + 0.01619965024292469, + 0.014986924827098846, + 0.016440972685813904, + 0.01266960334032774, + 0.014688635244965553, + 0.021314337849617004, + 0.011249284259974957, + 0.011101209558546543, + 0.011777312494814396, + 0.009348244406282902, + 0.0110814543440938, + 0.008229360915720463, + 0.009592185728251934, + 0.011203175410628319, + 0.009069878607988358, + 0.007835307158529758, + 0.005211450159549713, + 0.004102557431906462, + 0.005720602814108133, + 0.006515847519040108, + 0.003525680862367153, + 0.00397677393630147, + 0.005092512350529432, + 0.004467493388801813, + 0.006816016044467688, + 0.006648309528827667, + 0.01463403645902872, + 0.020558131858706474, + 0.005384424701333046, + 0.00862234365195036, + 0.0060754939913749695, + 0.008767153136432171, + 0.002387581393122673, + 0.0024249781854450703, + 0.004173628985881805, + 0.003827410750091076, + 0.0021966882050037384, + 0.0047933668829500675, + 0.0032183777075260878, + 0.005570422857999802, + 0.014336521737277508, + 0.016774481162428856, + 0.0192385446280241, + 0.04772910103201866, + 0.017171047627925873, + 0.024203458800911903, + 0.19032135605812073, + 0.1499549299478531, + 0.002995863789692521, + 0.006275983992964029, + 0.005002704914659262, + 0.005473143886774778, + 0.0026809563860297203, + 0.004949310328811407, + 0.004275939892977476, + 0.00571022043004632, + 0.006173549685627222, + 0.005595546681433916, + 0.007368066813796759, + 0.007532465737313032, + 0.005560960620641708, + 0.009518875740468502, + 0.015790622681379318, + 0.009990244172513485, + 0.0033210229594260454, + 0.0027467203326523304, + 0.0026551117189228535, + 0.003633993212133646, + 0.002623702632263303, + 0.002294340403750539, + 0.002996585099026561, + 0.0031054371502250433, + 0.005320193246006966, + 0.0061835781671106815, + 0.006845065392553806, + 0.006388901267200708, + 0.004298183601349592, + 0.008204213343560696, + 0.006802908610552549, + 0.006457326002418995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.00361420470289886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme close-up captures a dense cluster of delicate, bell-shaped inkcap mushrooms sprouting from the damp, moss-covered forest floor. The scene is plunged into dramatic lowlight, with a single focused beam of light piercing the darkness, focusing the fungi and creating a powerful black-and-white contrast. In a mesmerizing time-lapse sequence, the mushrooms begin to dissolve; their pale, grooved caps, initially rimmed with glistening dewdrops, begin to curl upward and inexorably. A thick, inky black substance forms around the edges, dripping viscously onto the decaying leaves below. Each drop catches the light, appearing like a living, dripping globule of black pigment before coalescing into a dark, lustrous puddle that slowly stains the earth. The entire process is a hypnotic spectacle of natural decay, transforming ephemeral fungi into ephemeral works of ink on the forest floor, creating an atmosphere both eerie and beautiful.", + "seed": 30737, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_8/index_6803_seed_30737_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": -0.015064917132258415, + "vlm_logprob": -0.015064917132258415, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.015064917132258415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012286453507840633 + ], + "opposite_logprob_tokens": [ + -4.88728666305542 + ], + "logprob_reward": -0.012286453507840633, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012286453507840633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16659900546073914, + 0.057702213525772095, + 0.06371522694826126, + 0.09773954749107361, + 0.05119074508547783, + 0.07053676247596741, + 0.04650329798460007, + 0.24561496078968048, + 0.1473838835954666, + 0.053014349192380905 + ], + "credit_map": [ + 0.023856481537222862, + 0.020971754565835, + 0.0160573311150074, + 0.016232233494520187, + 0.021108003333210945, + 0.023750193417072296, + 0.02200992777943611, + 0.02261306717991829, + 0.0075555602088570595, + 0.006924430374056101, + 0.008235282264649868, + 0.0060517932288348675, + 0.006711827125400305, + 0.008289777673780918, + 0.0057411715388298035, + 0.008192367851734161, + 0.006593502592295408, + 0.00576696777716279, + 0.009761965833604336, + 0.009457012638449669, + 0.005667647812515497, + 0.009383331052958965, + 0.008858118206262589, + 0.008226682431995869, + 0.010718443430960178, + 0.010469203814864159, + 0.013242908753454685, + 0.011598514392971992, + 0.010958298109471798, + 0.01418537087738514, + 0.01255795918405056, + 0.014008847065269947, + 0.005498664453625679, + 0.0059451148845255375, + 0.005481433589011431, + 0.006836376152932644, + 0.006409157533198595, + 0.006222549360245466, + 0.006275867577642202, + 0.008521579205989838, + 0.00618393998593092, + 0.004475725349038839, + 0.009159592911601067, + 0.010383492335677147, + 0.0066607194021344185, + 0.0126337306573987, + 0.010845813900232315, + 0.010193743743002415, + 0.0035411592107266188, + 0.003134586149826646, + 0.008833266794681549, + 0.006995214149355888, + 0.003106264164671302, + 0.010016870684921741, + 0.004621050786226988, + 0.006254886277019978, + 0.01660284586250782, + 0.013954238034784794, + 0.05235613137483597, + 0.03181599825620651, + 0.027132142335176468, + 0.028341736644506454, + 0.036695171147584915, + 0.038716696202754974, + 0.009071186184883118, + 0.008311939425766468, + 0.05086599290370941, + 0.014671111479401588, + 0.00692066689953208, + 0.01175648346543312, + 0.020448720082640648, + 0.02533777803182602, + 0.004148981533944607, + 0.003921655472368002, + 0.010723674669861794, + 0.005401875823736191, + 0.0051026539877057076, + 0.006414731964468956, + 0.009111943654716015, + 0.008188831619918346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.012286453507840633 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794510900974274 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794510900974274, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794510900974274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12361343950033188, + 0.054910000413656235, + 0.08828747272491455, + 0.09983556717634201, + 0.06324026733636856, + 0.06835811585187912, + 0.04981163889169693, + 0.06015145033597946, + 0.2677823007106781, + 0.12400975823402405 + ], + "credit_map": [ + 0.021361852064728737, + 0.010294351726770401, + 0.013685720041394234, + 0.01001146249473095, + 0.011802555061876774, + 0.01949298195540905, + 0.011840840801596642, + 0.025123674422502518, + 0.007523923180997372, + 0.007930673658847809, + 0.005615263245999813, + 0.006029559765011072, + 0.0054928334429860115, + 0.007433209102600813, + 0.007948677986860275, + 0.006935862824320793, + 0.010180734097957611, + 0.010971727780997753, + 0.010480357334017754, + 0.011909970082342625, + 0.00652712769806385, + 0.014692585915327072, + 0.010089309886097908, + 0.013435662724077702, + 0.012857384979724884, + 0.008764390833675861, + 0.0123807517811656, + 0.012025645934045315, + 0.009717418812215328, + 0.018532995134592056, + 0.011683405376970768, + 0.013873573392629623, + 0.006483284290879965, + 0.006055818870663643, + 0.00937590841203928, + 0.009659209288656712, + 0.006762208882719278, + 0.007887096144258976, + 0.007704272400587797, + 0.009312469512224197, + 0.005775697994977236, + 0.004108860157430172, + 0.009800359606742859, + 0.008087973110377789, + 0.006869800854474306, + 0.016356732696294785, + 0.0076282937079668045, + 0.009730398654937744, + 0.003919071983546019, + 0.0023436567280441523, + 0.01576886884868145, + 0.004716694820672274, + 0.0032039883080869913, + 0.007171939592808485, + 0.0054685259237885475, + 0.00721889128908515, + 0.005291219335049391, + 0.004329773131757975, + 0.008097740821540356, + 0.007153554819524288, + 0.008135420270264149, + 0.009850124828517437, + 0.009308569133281708, + 0.007985051721334457, + 0.023330094292759895, + 0.016815172508358955, + 0.07009445875883102, + 0.027239922434091568, + 0.02304886281490326, + 0.04083981364965439, + 0.03159281238913536, + 0.03482116386294365, + 0.012152944691479206, + 0.009399889037013054, + 0.02853052318096161, + 0.016984418034553528, + 0.01258946768939495, + 0.011533476412296295, + 0.014966151677072048, + 0.017852889373898506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.009794510900974274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807578772306442 + ], + "opposite_logprob_tokens": [ + -3.5298075675964355 + ], + "logprob_reward": -0.029807578772306442, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807578772306442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10827763378620148, + 0.12274084985256195, + 0.13893108069896698, + 0.08426928520202637, + 0.05683422088623047, + 0.12433251738548279, + 0.04681214690208435, + 0.08233842998743057, + 0.0517520010471344, + 0.18371187150478363 + ], + "credit_map": [ + 0.01955926977097988, + 0.011505679227411747, + 0.015005899593234062, + 0.012460348196327686, + 0.008077061735093594, + 0.015919089317321777, + 0.008331825025379658, + 0.017418455332517624, + 0.015854232013225555, + 0.01600165106356144, + 0.013287123292684555, + 0.014702506363391876, + 0.01422247476875782, + 0.01728084124624729, + 0.014156954362988472, + 0.017235063016414642, + 0.015180771239101887, + 0.01181178167462349, + 0.03796304389834404, + 0.015039111487567425, + 0.009446653537452221, + 0.018301744014024734, + 0.01336587406694889, + 0.01782209426164627, + 0.011871570721268654, + 0.007986496202647686, + 0.011313175782561302, + 0.009925334714353085, + 0.008386339992284775, + 0.012266885489225388, + 0.010197872295975685, + 0.01232160348445177, + 0.005180520471185446, + 0.0055572642013430595, + 0.007047631777822971, + 0.006943260785192251, + 0.00632850918918848, + 0.009258390404284, + 0.006377738900482655, + 0.010140904225409031, + 0.013237758539617062, + 0.00839139148592949, + 0.024736791849136353, + 0.015987833961844444, + 0.011928169056773186, + 0.026114946231245995, + 0.009393186308443546, + 0.014542430639266968, + 0.004230262245982885, + 0.002955404343083501, + 0.010323614813387394, + 0.007350990083068609, + 0.0033208720851689577, + 0.006048049312084913, + 0.006105853710323572, + 0.0064770993776619434, + 0.00597290089353919, + 0.004382146988064051, + 0.02234860137104988, + 0.009141628630459309, + 0.006801254115998745, + 0.010174188762903214, + 0.010277693159878254, + 0.013240008614957333, + 0.004069807473570108, + 0.005526005756109953, + 0.009930063039064407, + 0.00796252116560936, + 0.003323906334117055, + 0.008149783127009869, + 0.005628993269056082, + 0.007160920184105635, + 0.02531667798757553, + 0.01304610725492239, + 0.046318694949150085, + 0.0222385935485363, + 0.012790638022124767, + 0.01962360367178917, + 0.019178254529833794, + 0.025199294090270996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.029807578772306442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.15926988422870636, + 0.08590307831764221, + 0.08135928958654404, + 0.06055133789777756, + 0.050675250589847565, + 0.09485339373350143, + 0.06335548311471939, + 0.21114152669906616, + 0.09657313674688339, + 0.09631770849227905 + ], + "credit_map": [ + 0.021146243438124657, + 0.01881426014006138, + 0.02108073979616165, + 0.019138263538479805, + 0.022652724757790565, + 0.020976223051548004, + 0.016720997169613838, + 0.018740413710474968, + 0.009221237152814865, + 0.010439105331897736, + 0.011024557054042816, + 0.009852122515439987, + 0.009773071855306625, + 0.013060787692666054, + 0.010040744207799435, + 0.0124914376065135, + 0.008423449471592903, + 0.00876672938466072, + 0.014186904765665531, + 0.009420185349881649, + 0.00877263955771923, + 0.010228468105196953, + 0.009618774056434631, + 0.011942131444811821, + 0.007305552251636982, + 0.006091955117881298, + 0.007807170506566763, + 0.007520168554037809, + 0.005521283019334078, + 0.01086537167429924, + 0.008174643851816654, + 0.0072651877999305725, + 0.005090184044092894, + 0.005683740135282278, + 0.005968406330794096, + 0.007420344743877649, + 0.0051195453852415085, + 0.00690483720973134, + 0.004386545158922672, + 0.01010164339095354, + 0.008859620429575443, + 0.007211461663246155, + 0.015878232195973396, + 0.012771117500960827, + 0.009935157373547554, + 0.01969638094305992, + 0.009403901174664497, + 0.011097506619989872, + 0.004952436313033104, + 0.005227392073720694, + 0.009963739663362503, + 0.010747876949608326, + 0.0061154901050031185, + 0.008179078809916973, + 0.009914577007293701, + 0.008254887536168098, + 0.013958636671304703, + 0.01330984104424715, + 0.05795267969369888, + 0.020219754427671432, + 0.01922418177127838, + 0.023180313408374786, + 0.029680797830224037, + 0.03361527994275093, + 0.007277848199009895, + 0.006227992475032806, + 0.02555221877992153, + 0.012416123412549496, + 0.005777495913207531, + 0.009247347712516785, + 0.014026240445673466, + 0.01604786328971386, + 0.011509218253195286, + 0.01055979635566473, + 0.014563027769327164, + 0.009887232445180416, + 0.010343112051486969, + 0.013859921135008335, + 0.013523155823349953, + 0.012072233483195305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "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 shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18481257557868958, + 0.07226172089576721, + 0.14208577573299408, + 0.06807462126016617, + 0.06014952436089516, + 0.08350169658660889, + 0.05534178018569946, + 0.11448706686496735, + 0.0801793709397316, + 0.139105886220932 + ], + "credit_map": [ + 0.03079700469970703, + 0.017743421718478203, + 0.02628118172287941, + 0.017315566539764404, + 0.01647099293768406, + 0.026985708624124527, + 0.021244708448648453, + 0.027973981574177742, + 0.009416110813617706, + 0.0094709862023592, + 0.010402397252619267, + 0.008743560872972012, + 0.007104056421667337, + 0.00973266176879406, + 0.007897450588643551, + 0.009494493715465069, + 0.014120394363999367, + 0.013909922912716866, + 0.03271862864494324, + 0.015650929883122444, + 0.013432547450065613, + 0.01707891747355461, + 0.015301947481930256, + 0.019872481003403664, + 0.008012196980416775, + 0.008661198429763317, + 0.0075795771554112434, + 0.00997159630060196, + 0.007589271757751703, + 0.007836420089006424, + 0.00962757971137762, + 0.008796773850917816, + 0.005127998534590006, + 0.009082362055778503, + 0.006930878851562738, + 0.0076590352691709995, + 0.007360589224845171, + 0.006759285926818848, + 0.006190736312419176, + 0.011038634926080704, + 0.007129081059247255, + 0.005571715999394655, + 0.014027982018887997, + 0.009591340087354183, + 0.006774894893169403, + 0.021656673401594162, + 0.008480742573738098, + 0.010269266553223133, + 0.003968737553805113, + 0.003785155713558197, + 0.010849605314433575, + 0.007001557387411594, + 0.004504587966948748, + 0.010092919692397118, + 0.007196852471679449, + 0.00794235523790121, + 0.008998660370707512, + 0.007197272498160601, + 0.030225718393921852, + 0.011459470726549625, + 0.0103593859821558, + 0.015470202080905437, + 0.011376282200217247, + 0.019400063902139664, + 0.008219529874622822, + 0.008516088128089905, + 0.017361773177981377, + 0.008527172729372978, + 0.006232839077711105, + 0.012261253781616688, + 0.00852136965841055, + 0.010539336130023003, + 0.014862691052258015, + 0.01197169441729784, + 0.03032994642853737, + 0.014276610687375069, + 0.01459901686757803, + 0.014927440322935581, + 0.017260819673538208, + 0.0208776593208313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0052520157769322395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, extreme close-up captures a dense cluster of delicate, bell-shaped inkcap mushrooms sprouting from the damp, moss-covered forest floor. The scene is plunged into dramatic lowlight, with a single focused beam of light piercing the darkness, focusing the fungi and creating a powerful black-and-white contrast. In a mesmerizing time-lapse sequence, the mushrooms begin to dissolve; their pale, grooved caps, initially rimmed with glistening dewdrops, begin to curl upward and inexorably. A thick, inky black substance forms around the edges, dripping viscously onto the decaying leaves below. Each drop catches the light, appearing like a living, dripping globule of black pigment before coalescing into a dark, lustrous puddle that slowly stains the earth. The entire process is a hypnotic spectacle of natural decay, transforming ephemeral fungi into ephemeral works of ink on the forest floor, creating an atmosphere both eerie and beautiful.", + "seed": 30738, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_8/index_6803_seed_30738_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": -0.011481831967830657, + "vlm_logprob": -0.011481831967830657, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.011481831967830657, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009579760022461414 + ], + "opposite_logprob_tokens": [ + -5.134579658508301 + ], + "logprob_reward": -0.009579760022461414, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009579760022461414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07089122384786606, + 0.11996766924858093, + 0.11271312087774277, + 0.07098402082920074, + 0.49990421533584595, + 0.022620126605033875, + 0.03622745722532272, + 0.02669677883386612, + 0.014220360666513443, + 0.02577507309615612 + ], + "credit_map": [ + 0.00577639089897275, + 0.005006362684071064, + 0.009918305091559887, + 0.011428671889007092, + 0.006707162596285343, + 0.010482095181941986, + 0.008422567509114742, + 0.013149668462574482, + 0.008250869810581207, + 0.008962228894233704, + 0.008912544697523117, + 0.021741990000009537, + 0.012608274817466736, + 0.027731135487556458, + 0.013449649326503277, + 0.018310973420739174, + 0.008956506848335266, + 0.008367941714823246, + 0.0123013099655509, + 0.016686154529452324, + 0.011202716268599033, + 0.021380940452218056, + 0.014094102196395397, + 0.019723450765013695, + 0.0035086164716631174, + 0.002393411472439766, + 0.008668939583003521, + 0.008838311769068241, + 0.002512207953259349, + 0.020848069339990616, + 0.004055467899888754, + 0.02015899494290352, + 0.011703982017934322, + 0.012758786790072918, + 0.012864314019680023, + 0.05213693156838417, + 0.010286103002727032, + 0.18130946159362793, + 0.023997388780117035, + 0.1948472559452057, + 0.002802850678563118, + 0.0017325286753475666, + 0.004459989257156849, + 0.003569811349734664, + 0.0013952269218862057, + 0.0026025502011179924, + 0.0029547037556767464, + 0.0031024639029055834, + 0.0024075729306787252, + 0.0022682177368551493, + 0.003855277318507433, + 0.004266741219907999, + 0.0015503481263294816, + 0.008604029193520546, + 0.0028195769991725683, + 0.010455694980919361, + 0.0030685996171087027, + 0.0020791732240468264, + 0.005427840631455183, + 0.0035231278743594885, + 0.0013097422197461128, + 0.00240168790332973, + 0.003355399239808321, + 0.005531209986656904, + 0.001605883939191699, + 0.0011732254642993212, + 0.0020462076645344496, + 0.0019479672191664577, + 0.0011891295434907079, + 0.0028698991518467665, + 0.0018024117453023791, + 0.0015856357058510184, + 0.002963636303320527, + 0.0026286200154572725, + 0.005291777662932873, + 0.0038669919595122337, + 0.0015224413946270943, + 0.0019997062627226114, + 0.0036447367165237665, + 0.003857162082567811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.009579760022461414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746372207999229 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746372207999229, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746372207999229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09773281961679459, + 0.06997735053300858, + 0.1567607969045639, + 0.09702792018651962, + 0.17562681436538696, + 0.06723562628030777, + 0.048065755516290665, + 0.17625631392002106, + 0.03948004171252251, + 0.07183656096458435 + ], + "credit_map": [ + 0.009687290526926517, + 0.004918952006846666, + 0.015035085380077362, + 0.018584951758384705, + 0.006822448223829269, + 0.015785792842507362, + 0.009947679936885834, + 0.016950614750385284, + 0.005034696310758591, + 0.0037733183708041906, + 0.005314686801284552, + 0.013730852864682674, + 0.004898125771433115, + 0.0194733627140522, + 0.007485885638743639, + 0.010266422294080257, + 0.014183949679136276, + 0.014866895973682404, + 0.019913263618946075, + 0.022962624207139015, + 0.01731199584901333, + 0.019555766135454178, + 0.02562950737774372, + 0.022336794063448906, + 0.006993395276367664, + 0.006804364733397961, + 0.018134422600269318, + 0.01563437469303608, + 0.009618200361728668, + 0.014490198343992233, + 0.014267461374402046, + 0.011085502803325653, + 0.005827257409691811, + 0.005529864225536585, + 0.005362326744943857, + 0.018450269475579262, + 0.004812211729586124, + 0.061016976833343506, + 0.01004763226956129, + 0.0645802691578865, + 0.00757436640560627, + 0.004996809177100658, + 0.009158671833574772, + 0.00926055945456028, + 0.006314978934824467, + 0.006957253906875849, + 0.014052811078727245, + 0.008920174092054367, + 0.0035653586965054274, + 0.0024653791915625334, + 0.0052411966025829315, + 0.005760095082223415, + 0.002803995506837964, + 0.010244918055832386, + 0.0036197754088789225, + 0.014365037903189659, + 0.014032329432666302, + 0.010453610680997372, + 0.0314454659819603, + 0.03301054984331131, + 0.009388581849634647, + 0.017517531290650368, + 0.02077188529074192, + 0.03963635861873627, + 0.008428067900240421, + 0.0030712264124304056, + 0.004293177276849747, + 0.005355104338377714, + 0.00354822282679379, + 0.00576337706297636, + 0.0038873280864208937, + 0.005133537575602531, + 0.009059960953891277, + 0.007170833181589842, + 0.014976845122873783, + 0.012242404744029045, + 0.003797316923737526, + 0.005128901451826096, + 0.008621136657893658, + 0.010839160531759262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.006746372207999229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294012993574142 + ], + "opposite_logprob_tokens": [ + -3.773293972015381 + ], + "logprob_reward": -0.023294012993574142, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294012993574142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03344274312257767, + 0.042579587548971176, + 0.15247085690498352, + 0.12050881236791611, + 0.3779059648513794, + 0.038179975003004074, + 0.06063573434948921, + 0.07980800420045853, + 0.03362260386347771, + 0.0608457550406456 + ], + "credit_map": [ + 0.0035718674771487713, + 0.0020986064337193966, + 0.006595887243747711, + 0.0059504457749426365, + 0.002611566102132201, + 0.004008684307336807, + 0.0029930404853075743, + 0.005612644832581282, + 0.004349102731794119, + 0.0023975837975740433, + 0.004272113088518381, + 0.006797639187425375, + 0.003284326521679759, + 0.009049449115991592, + 0.005536121781915426, + 0.006893248297274113, + 0.009496604092419147, + 0.012627764604985714, + 0.017513174563646317, + 0.024102194234728813, + 0.01631583459675312, + 0.024196956306695938, + 0.022699987515807152, + 0.025518346577882767, + 0.003747838083654642, + 0.0035744616761803627, + 0.00806584395468235, + 0.0143795320764184, + 0.005111494567245245, + 0.0371326208114624, + 0.009104444645345211, + 0.03939256817102432, + 0.010369963012635708, + 0.010813853703439236, + 0.011802793480455875, + 0.03837476298213005, + 0.008332693949341774, + 0.13630244135856628, + 0.018953602761030197, + 0.14295582473278046, + 0.005029197316616774, + 0.002882518107071519, + 0.006772512104362249, + 0.004454741254448891, + 0.002596469596028328, + 0.003933344967663288, + 0.005046263802796602, + 0.007464922033250332, + 0.0030757077038288116, + 0.003247722052037716, + 0.004868645220994949, + 0.007889670319855213, + 0.0032174717634916306, + 0.013636313378810883, + 0.004405120387673378, + 0.020295079797506332, + 0.00598363159224391, + 0.007879232987761497, + 0.011653948575258255, + 0.01598740741610527, + 0.005324255209416151, + 0.01056048832833767, + 0.00803064089268446, + 0.01438839454203844, + 0.004060717765241861, + 0.002873198129236698, + 0.006313553545624018, + 0.006889226846396923, + 0.0018167855450883508, + 0.0021768712904304266, + 0.004368983209133148, + 0.005123264156281948, + 0.008470963686704636, + 0.004731065593659878, + 0.011414330452680588, + 0.011304665356874466, + 0.002632844028994441, + 0.0045448848977684975, + 0.0073308441787958145, + 0.010416155681014061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.023294012993574142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537354603409767 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537354603409767, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537354603409767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031007789075374603, + 0.052119240164756775, + 0.11008758097887039, + 0.05311845615506172, + 0.47570347785949707, + 0.08707151561975479, + 0.05839679390192032, + 0.05212580785155296, + 0.04806740954518318, + 0.032301995903253555 + ], + "credit_map": [ + 0.0032957594376057386, + 0.0023948815651237965, + 0.006067395210266113, + 0.0037768797483295202, + 0.0030516479164361954, + 0.005005097948014736, + 0.00275116297416389, + 0.004664962645620108, + 0.005517489276826382, + 0.004245868884027004, + 0.0067209345288574696, + 0.009091343730688095, + 0.004164570011198521, + 0.010618102736771107, + 0.004956244491040707, + 0.006804686039686203, + 0.005249822046607733, + 0.00849702674895525, + 0.008843737654387951, + 0.018585901707410812, + 0.010986289009451866, + 0.019159331917762756, + 0.021388042718172073, + 0.01737743616104126, + 0.0030158362351357937, + 0.00182135752402246, + 0.004781777039170265, + 0.006275413557887077, + 0.0020706544164568186, + 0.015262661501765251, + 0.00400504469871521, + 0.015885712578892708, + 0.017335228621959686, + 0.019313791766762733, + 0.017566850408911705, + 0.047045446932315826, + 0.017932103946805, + 0.15808649361133575, + 0.02716677449643612, + 0.17125678062438965, + 0.007305482402443886, + 0.007605746854096651, + 0.01183152012526989, + 0.010944029316306114, + 0.009111879393458366, + 0.011720532551407814, + 0.014913703314960003, + 0.01363862119615078, + 0.0038619220722466707, + 0.00441936356946826, + 0.00908614695072174, + 0.008215565234422684, + 0.004048306494951248, + 0.00994325801730156, + 0.005526029504835606, + 0.013296201825141907, + 0.004705139901489019, + 0.003955455031245947, + 0.00852230004966259, + 0.009314639493823051, + 0.003964945673942566, + 0.0053457459434866905, + 0.006761305965483189, + 0.00955627579241991, + 0.0043917871080338955, + 0.005189752671867609, + 0.007361577358096838, + 0.009769418276846409, + 0.0031821192242205143, + 0.0048116883262991905, + 0.006115989293903112, + 0.0072450777515769005, + 0.0020769790280610323, + 0.003467866685241461, + 0.005414154846221209, + 0.0037784255109727383, + 0.0028809718787670135, + 0.005274645984172821, + 0.004400324542075396, + 0.005008627660572529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.012537354603409767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, low-angle extreme close-up of a dense cluster of bell-shaped inkcap mushrooms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a damp, moss-covered forest floor with dramatic lowlight and a single focused beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video presented in black-and-white contrast with pale, grooved mushroom caps rimmed with glistening dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a time-lapse sequence where the mushrooms dissolve, curl upward, and are covered by a thick, inky black substance dripping onto decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the dripping black drops coalesce into a dark, lustrous puddle that slowly stains the earth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13390150666236877, + 0.07116660475730896, + 0.054456934332847595, + 0.10308480262756348, + 0.10922393947839737, + 0.21681183576583862, + 0.0473029650747776, + 0.05632185935974121, + 0.0952962264418602, + 0.1124332919716835 + ], + "credit_map": [ + 0.013292180374264717, + 0.00547405518591404, + 0.02739018201828003, + 0.028234634548425674, + 0.00788670964539051, + 0.019259581342339516, + 0.01645166054368019, + 0.015912510454654694, + 0.005915268789976835, + 0.005177738144993782, + 0.007775524631142616, + 0.009101632051169872, + 0.009004858322441578, + 0.01318326499313116, + 0.009045906364917755, + 0.011962419375777245, + 0.003667254466563463, + 0.004059947095811367, + 0.005855139810591936, + 0.00851510651409626, + 0.005401939153671265, + 0.006748327519744635, + 0.010235613211989403, + 0.00997360609471798, + 0.004937136545777321, + 0.0027946820482611656, + 0.006800949573516846, + 0.010740429162979126, + 0.0032983075361698866, + 0.03559008985757828, + 0.007709803059697151, + 0.03121340647339821, + 0.00580058665946126, + 0.005496490281075239, + 0.0086003877222538, + 0.011008615605533123, + 0.00528306420892477, + 0.030674666166305542, + 0.007878735661506653, + 0.03448139503598213, + 0.017028026282787323, + 0.020007813349366188, + 0.028286227956414223, + 0.02916024439036846, + 0.022309435531497, + 0.029528915882110596, + 0.0345180444419384, + 0.03597312048077583, + 0.004746186081320047, + 0.0063977851532399654, + 0.007222347427159548, + 0.005329178646206856, + 0.0041711051017045975, + 0.005887740291655064, + 0.004729802720248699, + 0.008818820118904114, + 0.004290230106562376, + 0.004964857362210751, + 0.010077722370624542, + 0.0072460477240383625, + 0.0034451670944690704, + 0.005005149636417627, + 0.00788920372724533, + 0.013403486460447311, + 0.010901532135903835, + 0.011162697337567806, + 0.014034897089004517, + 0.01788320392370224, + 0.006541089154779911, + 0.009177998639643192, + 0.013461708091199398, + 0.012133094482123852, + 0.007946320809423923, + 0.014287956058979034, + 0.013259295374155045, + 0.014000872150063515, + 0.010003771632909775, + 0.0181924756616354, + 0.01526294369250536, + 0.019479651004076004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.005251660011708736 + } + ], + "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/0000155/group_9/index_tensor([1264, 1264])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_9/index_tensor([1264, 1264])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e89ca1bb3a370754b7b1e1a44abd599ff30a7c12 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_9/index_tensor([1264, 1264])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera begins in a mid-level shot, focusing on a driverless vehicle constructed from smooth white voxels. It glides smoothly along clear light-gray voxel guides on the floor within a remarkably clean and bright automated piggery. Securely seated within the vehicle are several adorable voxel piglets, composed entirely of pink blocks. They possess simple black voxel eyes, slightly protruding square snouts, and a simple, innocent demeanor. The camera then steadily and deliberately ascends, gradually expanding from this cozy corner. As it rises higher, the imposing structure of the entire facility unfolds: countless uniform compartments, separated by translucent voxel walls, are arranged like a matrix, stretching into the distance to form a multi-layered, bottomless, vast space. Other autonomous vehicles and robotic arms work in an orderly fashion on their respective tracks. The entire scene is illuminated by a uniform, soft white light emitted by luminous voxel panels on the ceiling, creating a near-absence of shadows and a high-tech, sterile atmosphere. Ultimately, the camera hovers in a spectacular bird's-eye view, revealing a stunning voxel world that blends cute aesthetics with futuristic industrial efficiency, creating a tranquil and surreal atmosphere.", + "seed": 3564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_9/index_1264_seed_3564_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": -0.04407508887816221, + "vlm_logprob": -0.04407508887816221, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.04407508887816221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.18633891642093658 + ], + "opposite_logprob_tokens": [ + -1.8113389015197754 + ], + "logprob_reward": -0.18633891642093658, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18633891642093658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044483739882707596, + 0.2504754364490509, + 0.014085053466260433, + 0.031888723373413086, + 0.01086821686476469, + 0.0947151929140091, + 0.022732878103852272, + 0.04419699311256409, + 0.4411624073982239, + 0.045391347259283066 + ], + "credit_map": [ + 0.0057638101279735565, + 0.004850556142628193, + 0.005437792744487524, + 0.004729855805635452, + 0.005416186060756445, + 0.007953324355185032, + 0.0050633070059120655, + 0.005268909968435764, + 0.01673068478703499, + 0.009204600006341934, + 0.08700744062662125, + 0.014254887588322163, + 0.00883051473647356, + 0.0146492263302207, + 0.08278200030326843, + 0.017016083002090454, + 0.0015493808314204216, + 0.0020106430165469646, + 0.0021970090456306934, + 0.001382629037834704, + 0.0011793124722316861, + 0.0024044837336987257, + 0.0013084466336295009, + 0.002053148578852415, + 0.0038878575433045626, + 0.0036738577764481306, + 0.0028010746464133263, + 0.004401050508022308, + 0.0034085430670529604, + 0.006083832588046789, + 0.003342360956594348, + 0.004290145821869373, + 0.001427313545718789, + 0.0017849348951131105, + 0.0009167970856651664, + 0.0016474459553137422, + 0.0009712781757116318, + 0.000949317414779216, + 0.0015558250015601516, + 0.0016153042670339346, + 0.009811061434447765, + 0.013741758652031422, + 0.009642589837312698, + 0.01137935183942318, + 0.010800733231008053, + 0.013805253431200981, + 0.013837876729667187, + 0.011696564964950085, + 0.0033133765682578087, + 0.0031904291827231646, + 0.0026323185302317142, + 0.002528376877307892, + 0.00338759645819664, + 0.0026098110247403383, + 0.003147704293951392, + 0.0019232649356126785, + 0.0042100087739527225, + 0.005915626417845488, + 0.004785226657986641, + 0.005461731925606728, + 0.006321747321635485, + 0.005517508368939161, + 0.006434908136725426, + 0.00555023830384016, + 0.019970353692770004, + 0.17064468562602997, + 0.007472337689250708, + 0.011352350004017353, + 0.17837393283843994, + 0.03456077724695206, + 0.010769499465823174, + 0.008018444292247295, + 0.007024527061730623, + 0.004199435468763113, + 0.0052299196831882, + 0.004061294719576836, + 0.00671218940988183, + 0.006218588910996914, + 0.005423611029982567, + 0.006521780043840408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.18633891642093658 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.001722282380796969 + ], + "opposite_logprob_tokens": [ + -6.37672233581543 + ], + "logprob_reward": -0.001722282380796969, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001722282380796969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2046635001897812, + 0.1622030884027481, + 0.14451515674591064, + 0.052186883985996246, + 0.07134461402893066, + 0.033872101455926895, + 0.031022176146507263, + 0.031280823051929474, + 0.20535771548748016, + 0.06355396658182144 + ], + "credit_map": [ + 0.038365643471479416, + 0.018661124631762505, + 0.036127638071775436, + 0.02354838140308857, + 0.01707850955426693, + 0.019242359325289726, + 0.026948237791657448, + 0.024691585451364517, + 0.011374629102647305, + 0.011028535664081573, + 0.049970775842666626, + 0.009391723200678825, + 0.013878853060305119, + 0.011398354545235634, + 0.042065415531396866, + 0.013094776310026646, + 0.014873302541673183, + 0.021571483463048935, + 0.021832851693034172, + 0.01648261956870556, + 0.014738188125193119, + 0.02334510162472725, + 0.012683241628110409, + 0.0189883504062891, + 0.00755086075514555, + 0.00547786196693778, + 0.0049071237444877625, + 0.007972517050802708, + 0.006156068295240402, + 0.006262431852519512, + 0.0061742886900901794, + 0.007685726508498192, + 0.008884169161319733, + 0.009862625040113926, + 0.006749799475073814, + 0.010988352820277214, + 0.006917674094438553, + 0.0107314083725214, + 0.010101839900016785, + 0.007108737248927355, + 0.0032478515058755875, + 0.005024970509111881, + 0.0034368319902569056, + 0.004877910949289799, + 0.0043189870193600655, + 0.0049920761957764626, + 0.004364251159131527, + 0.0036092186346650124, + 0.003547501750290394, + 0.003909475170075893, + 0.003982853144407272, + 0.004606249276548624, + 0.004429186228662729, + 0.0028798601124435663, + 0.004351010546088219, + 0.0033160352613776922, + 0.0033407998271286488, + 0.004394951276481152, + 0.005635398440063, + 0.004171125125139952, + 0.003558909986168146, + 0.0033459141850471497, + 0.004188600927591324, + 0.002645120257511735, + 0.009667806327342987, + 0.07959022372961044, + 0.004176867660135031, + 0.005369723308831453, + 0.0824471265077591, + 0.01494919415563345, + 0.004989256151020527, + 0.004167524166405201, + 0.008244039490818977, + 0.006150432862341404, + 0.009443135932087898, + 0.00873430073261261, + 0.0070178634487092495, + 0.006465977057814598, + 0.009193683043122292, + 0.008304527960717678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.001722282380796969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.01606144569814205 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606144569814205, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606144569814205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04451463371515274, + 0.17325793206691742, + 0.03025893308222294, + 0.025243857875466347, + 0.06495552510023117, + 0.04193946346640587, + 0.13338442146778107, + 0.1925092339515686, + 0.23881796002388, + 0.05511792376637459 + ], + "credit_map": [ + 0.005211718380451202, + 0.005211878567934036, + 0.004309672396630049, + 0.005745306145399809, + 0.005781993269920349, + 0.005258642602711916, + 0.007146391086280346, + 0.005849034059792757, + 0.012440633028745651, + 0.006757835391908884, + 0.06127599626779556, + 0.009332935325801373, + 0.007959675043821335, + 0.013018779456615448, + 0.05194258317351341, + 0.010529524646699429, + 0.0031325763557106256, + 0.004759540781378746, + 0.005116238724440336, + 0.003206592984497547, + 0.0034129293635487556, + 0.004122464917600155, + 0.002746239770203829, + 0.0037623520474880934, + 0.003441470442339778, + 0.003292833687737584, + 0.002595545956864953, + 0.0038178435061126947, + 0.003140058135613799, + 0.002583020133897662, + 0.0028486396186053753, + 0.0035244496539235115, + 0.006897307001054287, + 0.0161996241658926, + 0.004413257818669081, + 0.006081112194806337, + 0.004345996771007776, + 0.006088895723223686, + 0.007497663609683514, + 0.013431682251393795, + 0.005475883837789297, + 0.004298269748687744, + 0.004234424326568842, + 0.005709843710064888, + 0.006636406295001507, + 0.004531340207904577, + 0.005716007202863693, + 0.005337290000170469, + 0.020586499944329262, + 0.01997988112270832, + 0.013165131211280823, + 0.01269819401204586, + 0.02248159795999527, + 0.014527020044624805, + 0.015904759988188744, + 0.01404134277254343, + 0.009117607027292252, + 0.06569430977106094, + 0.007066924124956131, + 0.008993146941065788, + 0.06586231291294098, + 0.011492975056171417, + 0.014709357172250748, + 0.009572619572281837, + 0.010667537339031696, + 0.0918346419930458, + 0.005470694974064827, + 0.006624327506870031, + 0.09582798182964325, + 0.017937276512384415, + 0.005403089802712202, + 0.005052434746176004, + 0.004811668768525124, + 0.006106891203671694, + 0.0077668605372309685, + 0.009750652126967907, + 0.006675230339169502, + 0.007014130707830191, + 0.007222754415124655, + 0.005769739393144846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.01606144569814205 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.007620786316692829 + ], + "opposite_logprob_tokens": [ + -4.882620811462402 + ], + "logprob_reward": -0.007620786316692829, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007620786316692829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08137530833482742, + 0.07897786051034927, + 0.022978859022259712, + 0.043235763907432556, + 0.07098270952701569, + 0.12876038253307343, + 0.026762016117572784, + 0.1994645595550537, + 0.2807002365589142, + 0.06676223129034042 + ], + "credit_map": [ + 0.012743178755044937, + 0.006417472846806049, + 0.01153513789176941, + 0.007936862297356129, + 0.006842251867055893, + 0.01101869810372591, + 0.010188995860517025, + 0.014692705124616623, + 0.006448337342590094, + 0.006326751317828894, + 0.016256416216492653, + 0.007109638303518295, + 0.007151071913540363, + 0.016680842265486717, + 0.010370706208050251, + 0.008634096011519432, + 0.0029143504798412323, + 0.00278591550886631, + 0.003691496327519417, + 0.0025470866821706295, + 0.0025027627125382423, + 0.003483957378193736, + 0.002447618404403329, + 0.0026056705974042416, + 0.006354945246130228, + 0.004640345927327871, + 0.003841696074232459, + 0.005000769160687923, + 0.00459144776687026, + 0.008589357137680054, + 0.00402339082211256, + 0.006193812470883131, + 0.007755616679787636, + 0.008006792515516281, + 0.007539382670074701, + 0.012851053848862648, + 0.006722751539200544, + 0.009604142047464848, + 0.01223465334624052, + 0.006268314551562071, + 0.01252216100692749, + 0.01997605711221695, + 0.01350885909050703, + 0.016747042536735535, + 0.015274109318852425, + 0.018625935539603233, + 0.017062008380889893, + 0.015044212341308594, + 0.002600595122203231, + 0.0036948679480701685, + 0.0038045800756663084, + 0.002910810289904475, + 0.003667686367407441, + 0.002955221338197589, + 0.004336581099778414, + 0.0027916706167161465, + 0.013240964151918888, + 0.06297339498996735, + 0.011709761805832386, + 0.010592592880129814, + 0.05826617404818535, + 0.016947224736213684, + 0.013795409351587296, + 0.011939042247831821, + 0.012292356230318546, + 0.10241811722517014, + 0.006538176443427801, + 0.006499270908534527, + 0.1149177998304367, + 0.025489261373877525, + 0.006929033901542425, + 0.0056162322871387005, + 0.007525704801082611, + 0.006230778526514769, + 0.00802026130259037, + 0.008322195149958134, + 0.011031482368707657, + 0.007494586054235697, + 0.008120100013911724, + 0.010017121210694313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.007620786316692829 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.008632013574242592 + ], + "opposite_logprob_tokens": [ + -4.758632183074951 + ], + "logprob_reward": -0.008632013574242592, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632013574242592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07704035192728043, + 0.13424549996852875, + 0.02259679138660431, + 0.026926420629024506, + 0.02441335655748844, + 0.06219906732439995, + 0.03705371916294098, + 0.06542900949716568, + 0.43069663643836975, + 0.11939909309148788 + ], + "credit_map": [ + 0.010220815427601337, + 0.007131911814212799, + 0.005992593243718147, + 0.006647544912993908, + 0.008457127027213573, + 0.009977246634662151, + 0.014038493856787682, + 0.014574620872735977, + 0.009451068937778473, + 0.006049011368304491, + 0.045222360640764236, + 0.0072469389997422695, + 0.005845970474183559, + 0.007858839817345142, + 0.04272603616118431, + 0.009845268912613392, + 0.0022414131090044975, + 0.003722903085872531, + 0.0029859289061278105, + 0.003136666025966406, + 0.002625431399792433, + 0.002443950856104493, + 0.0021034067030996084, + 0.0033370894379913807, + 0.003403046168386936, + 0.0038122141268104315, + 0.0020934680942445993, + 0.0037873433902859688, + 0.003021653275936842, + 0.0044030663557350636, + 0.002688090782612562, + 0.003717539831995964, + 0.002987730083987117, + 0.005519963335245848, + 0.0020379831548780203, + 0.0025267447344958782, + 0.0018515997799113393, + 0.001978909829631448, + 0.0028969664126634598, + 0.0046134586445987225, + 0.007212652359157801, + 0.009557968936860561, + 0.005780280102044344, + 0.005777562037110329, + 0.007517711259424686, + 0.00877191312611103, + 0.009301646612584591, + 0.008279336616396904, + 0.0057471198961138725, + 0.0056244442239403725, + 0.003001321339979768, + 0.003731146454811096, + 0.007850606925785542, + 0.003738021943718195, + 0.0045836507342755795, + 0.002777406480163336, + 0.0038100050296634436, + 0.022327706217765808, + 0.0034555040765553713, + 0.002797270892187953, + 0.022128663957118988, + 0.00379790086299181, + 0.004381812643259764, + 0.0027301537338644266, + 0.018542630597949028, + 0.1637144684791565, + 0.006436518859118223, + 0.008848554454743862, + 0.17990772426128387, + 0.0377722904086113, + 0.008173855021595955, + 0.007300585508346558, + 0.013127309270203114, + 0.011158338747918606, + 0.011807751841843128, + 0.011070407927036285, + 0.020560694858431816, + 0.021217359229922295, + 0.012747475877404213, + 0.0177097599953413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.008632013574242592 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins in a mid-level shot, focusing on a driverless vehicle constructed from smooth white voxels. It glides smoothly along clear light-gray voxel guides on the floor within a remarkably clean and bright automated piggery. Securely seated within the vehicle are several adorable voxel piglets, composed entirely of pink blocks. They possess simple black voxel eyes, slightly protruding square snouts, and a simple, innocent demeanor. The camera then steadily and deliberately ascends, gradually expanding from this cozy corner. As it rises higher, the imposing structure of the entire facility unfolds: countless uniform compartments, separated by translucent voxel walls, are arranged like a matrix, stretching into the distance to form a multi-layered, bottomless, vast space. Other autonomous vehicles and robotic arms work in an orderly fashion on their respective tracks. The entire scene is illuminated by a uniform, soft white light emitted by luminous voxel panels on the ceiling, creating a near-absence of shadows and a high-tech, sterile atmosphere. Ultimately, the camera hovers in a spectacular bird's-eye view, revealing a stunning voxel world that blends cute aesthetics with futuristic industrial efficiency, creating a tranquil and surreal atmosphere.", + "seed": 3565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_9/index_1264_seed_3565_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": -0.011996979522518814, + "vlm_logprob": -0.011996979522518814, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.011996979522518814, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.02883179672062397 + ], + "opposite_logprob_tokens": [ + -3.653831720352173 + ], + "logprob_reward": -0.02883179672062397, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02883179672062397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0209678765386343, + 0.034936580806970596, + 0.030003869906067848, + 0.15123391151428223, + 0.05821319669485092, + 0.018875814974308014, + 0.06303617358207703, + 0.04189038276672363, + 0.03119097277522087, + 0.5496512055397034 + ], + "credit_map": [ + 0.002073949435725808, + 0.0018549623200669885, + 0.0025154510512948036, + 0.00279663666151464, + 0.002363705076277256, + 0.003871261142194271, + 0.0027646124362945557, + 0.0027272978331893682, + 0.0031239790841937065, + 0.0028484228532761335, + 0.004361062776297331, + 0.002568827709183097, + 0.0031211713794618845, + 0.004557841457426548, + 0.004392062779515982, + 0.009963211603462696, + 0.00448869401589036, + 0.003781751962378621, + 0.0044639501720666885, + 0.0031552473083138466, + 0.00409674970433116, + 0.0037370375357568264, + 0.002804931253194809, + 0.0034755084197968245, + 0.01130323950201273, + 0.003822064260020852, + 0.04609470069408417, + 0.009351770393550396, + 0.003217593766748905, + 0.04920491576194763, + 0.00885490607470274, + 0.01938471570611, + 0.008255843073129654, + 0.004178194794803858, + 0.022685682401061058, + 0.0030681928619742393, + 0.00398894352838397, + 0.007484850939363241, + 0.005664696916937828, + 0.0028867912478744984, + 0.0034784593153744936, + 0.0030215238220989704, + 0.0021412961650639772, + 0.0026277077849954367, + 0.0026652785018086433, + 0.0016266165766865015, + 0.0017198602436110377, + 0.0015950730303302407, + 0.006079444196075201, + 0.007295140530914068, + 0.0062090735882520676, + 0.004146711900830269, + 0.021107055246829987, + 0.0053379228338599205, + 0.0073543754406273365, + 0.005506447982043028, + 0.008917460218071938, + 0.0034708252642303705, + 0.008447867818176746, + 0.002170835854485631, + 0.0037475985009223223, + 0.00688851997256279, + 0.004875643644481897, + 0.003371631260961294, + 0.008039573207497597, + 0.0025241340044885874, + 0.0032761553302407265, + 0.001785546774044633, + 0.0018401062116026878, + 0.009133041836321354, + 0.0017685224302113056, + 0.0028238920494914055, + 0.21098464727401733, + 0.030135905370116234, + 0.026575032621622086, + 0.007765693590044975, + 0.016671961173415184, + 0.22478611767292023, + 0.020549915730953217, + 0.0121819619089365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.02883179672062397 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.0019432486733421683 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.0019432486733421683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019432486733421683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039345867931842804, + 0.029303675517439842, + 0.031028615310788155, + 0.13088127970695496, + 0.04109000042080879, + 0.013852473348379135, + 0.010667413473129272, + 0.016586491838097572, + 0.031109366565942764, + 0.6561347246170044 + ], + "credit_map": [ + 0.0049172076396644115, + 0.00435143755748868, + 0.005669836420565844, + 0.005428819451481104, + 0.004700074903666973, + 0.006382982712239027, + 0.003738234518095851, + 0.004157275427132845, + 0.003925869707018137, + 0.004243561998009682, + 0.002863712841644883, + 0.002941177925094962, + 0.004076744895428419, + 0.0035429203417152166, + 0.0025880385655909777, + 0.005121649242937565, + 0.005800280719995499, + 0.00247780024074018, + 0.0044133602641522884, + 0.003342114621773362, + 0.003280757460743189, + 0.005365836899727583, + 0.0024296948686242104, + 0.003918773028999567, + 0.010981212370097637, + 0.0033724606037139893, + 0.04608979821205139, + 0.008369319140911102, + 0.0028078246396034956, + 0.041361261159181595, + 0.00740791717544198, + 0.010491493158042431, + 0.0031327554024755955, + 0.0028188617434352636, + 0.01024834718555212, + 0.0018484251340851188, + 0.004558539018034935, + 0.002813076600432396, + 0.013516252860426903, + 0.002153747482225299, + 0.0014863567193970084, + 0.001986897550523281, + 0.001038291840814054, + 0.0017806628020480275, + 0.0035432965960353613, + 0.00121324323117733, + 0.0013682497665286064, + 0.0014354761224240065, + 0.0010586921125650406, + 0.001094354665838182, + 0.0017181074945256114, + 0.0015354619827121496, + 0.002132896799594164, + 0.0009245482506230474, + 0.001128595438785851, + 0.0010747587075456977, + 0.002742467448115349, + 0.0015260373475030065, + 0.0024145387578755617, + 0.00103734340518713, + 0.0016755912220105529, + 0.003487110137939453, + 0.0019363726023584604, + 0.0017670317320153117, + 0.0094726188108325, + 0.0017241999739781022, + 0.0022083271760493517, + 0.0011383662931621075, + 0.0017096756491810083, + 0.010581397451460361, + 0.0018056168919429183, + 0.0024691668804734945, + 0.25181835889816284, + 0.03566976636648178, + 0.03179500624537468, + 0.009440143592655659, + 0.0201280377805233, + 0.26821717619895935, + 0.024258557707071304, + 0.014807738363742828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.0019432486733421683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.0052490513771772385 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.0052490513771772385, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052490513771772385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02272181026637554, + 0.020715033635497093, + 0.025243418291211128, + 0.24258509278297424, + 0.13760395348072052, + 0.2629387080669403, + 0.02151254564523697, + 0.017507776618003845, + 0.017145982012152672, + 0.23202571272850037 + ], + "credit_map": [ + 0.002273635705932975, + 0.0030049977358430624, + 0.002808646997436881, + 0.0031699298415333033, + 0.003601409262046218, + 0.0040291291661560535, + 0.0018883593147620559, + 0.001945702824741602, + 0.002035679994150996, + 0.0016237198142334819, + 0.0028793588280677795, + 0.0016287685139104724, + 0.0019136545015498996, + 0.0029995907098054886, + 0.0022388468496501446, + 0.005395414773374796, + 0.004249336663633585, + 0.0024928117636591196, + 0.003587018232792616, + 0.0021965859923511744, + 0.0027647940441966057, + 0.005000725854188204, + 0.002041611121967435, + 0.002910534618422389, + 0.02114393562078476, + 0.005863802507519722, + 0.08962813764810562, + 0.015486297197639942, + 0.005182353779673576, + 0.07358228415250778, + 0.014239992946386337, + 0.017458297312259674, + 0.03519607335329056, + 0.01143364142626524, + 0.05358133465051651, + 0.007631010375916958, + 0.006680562160909176, + 0.010699075646698475, + 0.006632066797465086, + 0.005750196520239115, + 0.019623449072241783, + 0.108141228556633, + 0.00712310615926981, + 0.0850558802485466, + 0.014829587191343307, + 0.011075346730649471, + 0.008208226412534714, + 0.008881890214979649, + 0.002698388183489442, + 0.0026578842662274837, + 0.0032934483606368303, + 0.0019092783331871033, + 0.0039389654994010925, + 0.002171314787119627, + 0.0023077039513736963, + 0.002535561565309763, + 0.002679449040442705, + 0.0017030207673087716, + 0.002461283467710018, + 0.0011871586320921779, + 0.0019990145228803158, + 0.003040896961465478, + 0.0024466703180223703, + 0.001990283839404583, + 0.003627021098509431, + 0.0013105102116242051, + 0.002036087680608034, + 0.00102415110450238, + 0.0012116360012441874, + 0.0046869260258972645, + 0.0011787492549046874, + 0.0020708998199552298, + 0.08891890197992325, + 0.012638277374207973, + 0.011145101860165596, + 0.0033251747954636812, + 0.007126127369701862, + 0.09488627314567566, + 0.008809845894575119, + 0.005175996571779251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.0052490513771772385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.009772200137376785 + ], + "opposite_logprob_tokens": [ + -4.634772300720215 + ], + "logprob_reward": -0.009772200137376785, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772200137376785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06381304562091827, + 0.05875494331121445, + 0.06702596694231033, + 0.0804843008518219, + 0.07779422402381897, + 0.0574321523308754, + 0.03804474696516991, + 0.05159050226211548, + 0.034756582230329514, + 0.47030359506607056 + ], + "credit_map": [ + 0.006514688953757286, + 0.004923361819237471, + 0.01042742095887661, + 0.00789128988981247, + 0.007316635921597481, + 0.008907867595553398, + 0.01037505827844143, + 0.007456716615706682, + 0.006682687439024448, + 0.007601834833621979, + 0.006404771003872156, + 0.007003852631896734, + 0.01001064758747816, + 0.007846225053071976, + 0.0060084122233092785, + 0.007196506951004267, + 0.013690467923879623, + 0.004245374351739883, + 0.010148923844099045, + 0.006304231472313404, + 0.004999767057597637, + 0.012200175784528255, + 0.005842705722898245, + 0.009594324044883251, + 0.006646492052823305, + 0.0035583137068897486, + 0.02424275316298008, + 0.005698372144252062, + 0.0025549631100147963, + 0.024416398257017136, + 0.00472680339589715, + 0.008640202693641186, + 0.0137192877009511, + 0.006880003027617931, + 0.01710573397576809, + 0.005223378073424101, + 0.004678019322454929, + 0.012939813546836376, + 0.009949604980647564, + 0.007298386190086603, + 0.004009239375591278, + 0.02076124958693981, + 0.0020112493075430393, + 0.0177485141903162, + 0.004506195895373821, + 0.0025426726788282394, + 0.0029606318566948175, + 0.0028924026992172003, + 0.005091121420264244, + 0.004000059794634581, + 0.007624673657119274, + 0.00534706050530076, + 0.003930715378373861, + 0.004213407635688782, + 0.0040876008570194244, + 0.0037501079495996237, + 0.009880750440061092, + 0.004217203240841627, + 0.011038822121918201, + 0.0033626160584390163, + 0.003471672534942627, + 0.01127493754029274, + 0.003833198454231024, + 0.004511302802711725, + 0.007500495761632919, + 0.003037324408069253, + 0.004917680751532316, + 0.002280052751302719, + 0.0030327013228088617, + 0.00959162600338459, + 0.0020744206849485636, + 0.0023222824092954397, + 0.18037371337413788, + 0.02541871927678585, + 0.022459618747234344, + 0.006905037444084883, + 0.014494791626930237, + 0.19231122732162476, + 0.01764063350856304, + 0.010699848644435406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.009772200137376785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.014188600704073906 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188600704073906, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188600704073906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038315191864967346, + 0.016360202804207802, + 0.025085359811782837, + 0.01896517723798752, + 0.08393455296754837, + 0.039695631712675095, + 0.03308425843715668, + 0.07010740041732788, + 0.047453831881284714, + 0.6269983649253845 + ], + "credit_map": [ + 0.00420094421133399, + 0.003963320516049862, + 0.006010629702359438, + 0.003955504857003689, + 0.006326320581138134, + 0.006633142940700054, + 0.0034536514431238174, + 0.0037716752849519253, + 0.0020415647886693478, + 0.001592052634805441, + 0.0017229605000466108, + 0.0017751712584868073, + 0.0021378183737397194, + 0.001792422030121088, + 0.0021396237425506115, + 0.003158588893711567, + 0.0030467105098068714, + 0.00328070973046124, + 0.0033081001602113247, + 0.002246672520413995, + 0.004085189662873745, + 0.0035192607901990414, + 0.002814705716446042, + 0.0027840114198625088, + 0.0015061465092003345, + 0.0014040020760148764, + 0.0018248638371005654, + 0.0018822603160515428, + 0.0015653240261599422, + 0.003920772112905979, + 0.0024170398246496916, + 0.004444769583642483, + 0.0058353436179459095, + 0.006231356877833605, + 0.02820734865963459, + 0.004254851024597883, + 0.0063005550764501095, + 0.003848023945465684, + 0.02564789168536663, + 0.0036091834772378206, + 0.00309912022203207, + 0.007501221261918545, + 0.0028991347644478083, + 0.008392211981117725, + 0.006702485494315624, + 0.003566183615475893, + 0.0035452619194984436, + 0.003990011755377054, + 0.0031937984749674797, + 0.0032360453624278307, + 0.004337660036981106, + 0.003603908233344555, + 0.008339082822203636, + 0.0030716899782419205, + 0.004204990342259407, + 0.00309708621352911, + 0.015060899779200554, + 0.005988499615341425, + 0.013706840574741364, + 0.004257282707840204, + 0.005026266444474459, + 0.014497600495815277, + 0.006132669281214476, + 0.005437345709651709, + 0.010788045823574066, + 0.0035433771554380655, + 0.005356448702514172, + 0.0032224676106125116, + 0.003246929496526718, + 0.013499782420694828, + 0.0030844451393932104, + 0.00471233855932951, + 0.2407040297985077, + 0.034417569637298584, + 0.030497578904032707, + 0.00878588855266571, + 0.018887905403971672, + 0.25593051314353943, + 0.023818116635084152, + 0.013956802897155285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.014188600704073906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins in a mid-level shot, focusing on a driverless vehicle constructed from smooth white voxels. It glides smoothly along clear light-gray voxel guides on the floor within a remarkably clean and bright automated piggery. Securely seated within the vehicle are several adorable voxel piglets, composed entirely of pink blocks. They possess simple black voxel eyes, slightly protruding square snouts, and a simple, innocent demeanor. The camera then steadily and deliberately ascends, gradually expanding from this cozy corner. As it rises higher, the imposing structure of the entire facility unfolds: countless uniform compartments, separated by translucent voxel walls, are arranged like a matrix, stretching into the distance to form a multi-layered, bottomless, vast space. Other autonomous vehicles and robotic arms work in an orderly fashion on their respective tracks. The entire scene is illuminated by a uniform, soft white light emitted by luminous voxel panels on the ceiling, creating a near-absence of shadows and a high-tech, sterile atmosphere. Ultimately, the camera hovers in a spectacular bird's-eye view, revealing a stunning voxel world that blends cute aesthetics with futuristic industrial efficiency, creating a tranquil and surreal atmosphere.", + "seed": 3566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_9/index_1264_seed_3566_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": -0.06824552821926773, + "vlm_logprob": -0.06824552821926773, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.06824552821926773, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.28460830450057983 + ], + "opposite_logprob_tokens": [ + -1.4096083641052246 + ], + "logprob_reward": -0.28460830450057983, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28460830450057983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14117823541164398, + 0.039264775812625885, + 0.045392073690891266, + 0.2735549211502075, + 0.06341611593961716, + 0.17899924516677856, + 0.06917057931423187, + 0.07041759043931961, + 0.04020724073052406, + 0.07839927822351456 + ], + "credit_map": [ + 0.018765415996313095, + 0.02024735137820244, + 0.02760126069188118, + 0.012662586756050587, + 0.014828750863671303, + 0.016019050031900406, + 0.014960079453885555, + 0.01609371416270733, + 0.004463599529117346, + 0.006004398223012686, + 0.005735567305237055, + 0.003937159199267626, + 0.006290630903095007, + 0.00497372355312109, + 0.003145634662359953, + 0.004714058246463537, + 0.005259412806481123, + 0.005419272463768721, + 0.006039069499820471, + 0.005123586859554052, + 0.008372648619115353, + 0.00526379095390439, + 0.0034622373059391975, + 0.006452048197388649, + 0.019697492942214012, + 0.06526243686676025, + 0.011694029904901981, + 0.011859959922730923, + 0.1104661300778389, + 0.02928527072072029, + 0.012976109981536865, + 0.012313484214246273, + 0.00680864043533802, + 0.010316272266209126, + 0.007190834730863571, + 0.006831065285950899, + 0.008980780839920044, + 0.008477553725242615, + 0.006228554528206587, + 0.008582403883337975, + 0.016844509169459343, + 0.013928988017141819, + 0.01290684100240469, + 0.011103260330855846, + 0.05281845107674599, + 0.04898205026984215, + 0.012414807453751564, + 0.010000316426157951, + 0.014233622699975967, + 0.007946480065584183, + 0.006178026087582111, + 0.007968645542860031, + 0.006847837939858437, + 0.009833247400820255, + 0.009435364976525307, + 0.006727342028170824, + 0.009120315313339233, + 0.008682996034622192, + 0.007739522028714418, + 0.006276315543800592, + 0.011018273420631886, + 0.014091827906668186, + 0.004706496372818947, + 0.008781840093433857, + 0.0039091287180781364, + 0.007200897671282291, + 0.004659088794142008, + 0.003985244780778885, + 0.005273362156003714, + 0.009712479077279568, + 0.003117194166406989, + 0.0023498402442783117, + 0.005285086575895548, + 0.005171500146389008, + 0.028288109228014946, + 0.02747456170618534, + 0.0026785184163600206, + 0.0025054377038031816, + 0.0035725925117731094, + 0.003423463786020875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.28460830450057983 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.0028269824106246233 + ], + "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": [ + 0.17411817610263824, + 0.08968130499124527, + 0.08646685630083084, + 0.15139514207839966, + 0.23373475670814514, + 0.08777890354394913, + 0.0648508071899414, + 0.019752729684114456, + 0.02374868653714657, + 0.06847260147333145 + ], + "credit_map": [ + 0.02478482574224472, + 0.03173493966460228, + 0.026217585429549217, + 0.014858794398605824, + 0.022590288892388344, + 0.01976454257965088, + 0.01612393744289875, + 0.018043261021375656, + 0.009960021823644638, + 0.01454372238367796, + 0.011250367388129234, + 0.009134703315794468, + 0.018310219049453735, + 0.009099192917346954, + 0.008235350251197815, + 0.00914773065596819, + 0.009893879294395447, + 0.014030931517481804, + 0.00977998971939087, + 0.01038126926869154, + 0.015462706796824932, + 0.009403052739799023, + 0.006039726547896862, + 0.011475305072963238, + 0.009479462169110775, + 0.04112783446907997, + 0.008700706996023655, + 0.008440257981419563, + 0.05004381388425827, + 0.015716826543211937, + 0.008925328031182289, + 0.008960924111306667, + 0.015578988008201122, + 0.09245852380990982, + 0.007469794247299433, + 0.0077461362816393375, + 0.08255642652511597, + 0.013090425170958042, + 0.008930705487728119, + 0.005903774406760931, + 0.008494661189615726, + 0.009316801093518734, + 0.012891450896859169, + 0.007947332225739956, + 0.01178690791130066, + 0.022973813116550446, + 0.0076306103728711605, + 0.0067373355850577354, + 0.01294083334505558, + 0.008568795397877693, + 0.0062449052929878235, + 0.007884649559855461, + 0.005536265671253204, + 0.00811309739947319, + 0.005869983229786158, + 0.009692284278571606, + 0.002546163508668542, + 0.002266116440296173, + 0.0026564598083496094, + 0.0018792329356074333, + 0.0026652843225747347, + 0.0037105893716216087, + 0.0022319061681628227, + 0.0017969772452488542, + 0.0016314233653247356, + 0.0029869587160646915, + 0.002651497721672058, + 0.002418345073238015, + 0.0030625052750110626, + 0.006795576773583889, + 0.002348527777940035, + 0.0018538528820499778, + 0.0045661283656954765, + 0.018545659258961678, + 0.02057904191315174, + 0.005514173302799463, + 0.005195161793380976, + 0.005050253123044968, + 0.005694965831935406, + 0.0033272213768213987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "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 driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.038088809698820114 + ], + "opposite_logprob_tokens": [ + -3.288088798522949 + ], + "logprob_reward": -0.038088809698820114, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038088809698820114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032263435423374176, + 0.048441000282764435, + 0.028246678411960602, + 0.19601431488990784, + 0.050013843923807144, + 0.29356926679611206, + 0.04584430158138275, + 0.015187961049377918, + 0.1048734039068222, + 0.18554580211639404 + ], + "credit_map": [ + 0.005348116625100374, + 0.003149741794914007, + 0.006904503330588341, + 0.002482824958860874, + 0.0033924486488103867, + 0.004126700572669506, + 0.003031049156561494, + 0.0038280498702079058, + 0.006847511976957321, + 0.004181075841188431, + 0.007528793066740036, + 0.004176051821559668, + 0.00488863093778491, + 0.010715769603848457, + 0.003936832305043936, + 0.0061663370579481125, + 0.0026770627591758966, + 0.004620449151843786, + 0.00413052923977375, + 0.0029921187087893486, + 0.0041806395165622234, + 0.00284058996476233, + 0.00258664065040648, + 0.0042186458595097065, + 0.01905578374862671, + 0.03610127791762352, + 0.007419743109494448, + 0.00596679886803031, + 0.0839640349149704, + 0.02222135290503502, + 0.008384812623262405, + 0.012900512665510178, + 0.0050626276060938835, + 0.006728271953761578, + 0.004655621480196714, + 0.004111796151846647, + 0.013185563497245312, + 0.00579612934961915, + 0.003732290118932724, + 0.006741540972143412, + 0.02131449617445469, + 0.013494359329342842, + 0.03379107266664505, + 0.0078118713572621346, + 0.08285031467676163, + 0.11186301708221436, + 0.00836645346134901, + 0.01407766342163086, + 0.008422941900789738, + 0.005153533071279526, + 0.005031942389905453, + 0.008191179484128952, + 0.0037663099355995655, + 0.004180655814707279, + 0.006078219972550869, + 0.005019514821469784, + 0.0021258953493088484, + 0.0018102395115420222, + 0.002201541792601347, + 0.002262247959151864, + 0.0014820951037108898, + 0.0014601266011595726, + 0.0021168706007301807, + 0.0017289441311731935, + 0.009112945757806301, + 0.011554822325706482, + 0.009826907888054848, + 0.008629814721643925, + 0.010655468329787254, + 0.028813768178224564, + 0.011218431405723095, + 0.015061232261359692, + 0.008353890851140022, + 0.013507574796676636, + 0.0744721069931984, + 0.06467293947935104, + 0.007295502349734306, + 0.005904125981032848, + 0.007077396847307682, + 0.004262242931872606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.038088809698820114 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.004632811527699232 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004632811527699232, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632811527699232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059117332100868225, + 0.030267829075455666, + 0.05513901636004448, + 0.15247321128845215, + 0.18548423051834106, + 0.21003003418445587, + 0.041888587176799774, + 0.10296496003866196, + 0.04508896917104721, + 0.11754578351974487 + ], + "credit_map": [ + 0.00842556543648243, + 0.007729208562523127, + 0.008634215220808983, + 0.008642937056720257, + 0.006420728284865618, + 0.006301810499280691, + 0.006137716118246317, + 0.00682515325024724, + 0.0035811609122902155, + 0.003372070612385869, + 0.003675705287605524, + 0.003363499417901039, + 0.005613401066511869, + 0.0042297313921153545, + 0.00286252168007195, + 0.0035697391722351313, + 0.010117565281689167, + 0.007299444172531366, + 0.0048133935779333115, + 0.004989056382328272, + 0.007210968527942896, + 0.005922538693994284, + 0.005274656228721142, + 0.009511398151516914, + 0.019773155450820923, + 0.018669819459319115, + 0.008195415139198303, + 0.008100688457489014, + 0.04509548097848892, + 0.01682312972843647, + 0.008701002225279808, + 0.027114516124129295, + 0.011022775433957577, + 0.0704767107963562, + 0.005733896978199482, + 0.006593120750039816, + 0.06481094658374786, + 0.01296786218881607, + 0.0075453720055520535, + 0.006333538331091404, + 0.01894236169755459, + 0.01680242270231247, + 0.04935453087091446, + 0.015227383002638817, + 0.025821639224886894, + 0.06282374262809753, + 0.011756098829209805, + 0.009301871061325073, + 0.007995161227881908, + 0.005980747751891613, + 0.004127597901970148, + 0.006574627012014389, + 0.003556532319635153, + 0.004605363123118877, + 0.0038786493241786957, + 0.005169908050447702, + 0.0128620695322752, + 0.011643734760582447, + 0.011945659294724464, + 0.008835199289023876, + 0.015134237706661224, + 0.020842820405960083, + 0.009797651320695877, + 0.011903590522706509, + 0.003679363988339901, + 0.003375863889232278, + 0.005072937346994877, + 0.004203680902719498, + 0.006013310980051756, + 0.011264041066169739, + 0.005428294185549021, + 0.006051473319530487, + 0.012210420332849026, + 0.036181312054395676, + 0.008929339237511158, + 0.03092089109122753, + 0.008429733105003834, + 0.009953556582331657, + 0.006053145043551922, + 0.00486739631742239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.004632811527699232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.011070732958614826 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070732958614826, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070732958614826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05815933272242546, + 0.055286359041929245, + 0.023916171863675117, + 0.12060824036598206, + 0.05717712640762329, + 0.12988267838954926, + 0.04244459420442581, + 0.08647997677326202, + 0.21207810938358307, + 0.21396739780902863 + ], + "credit_map": [ + 0.007451277691870928, + 0.0062641906552016735, + 0.011579644866287708, + 0.004976364783942699, + 0.0052037849090993404, + 0.008248354308307171, + 0.004725941456854343, + 0.009709778241813183, + 0.008520525880157948, + 0.005537913180887699, + 0.007164228707551956, + 0.004317121580243111, + 0.00645480677485466, + 0.01247316412627697, + 0.00407006312161684, + 0.006748540792614222, + 0.004513162188231945, + 0.00388101814314723, + 0.0028401047457009554, + 0.002196934074163437, + 0.0022103034425526857, + 0.0028831844683736563, + 0.0015873291995376348, + 0.0038041367661207914, + 0.009328528307378292, + 0.03701013699173927, + 0.006309293210506439, + 0.007109445054084063, + 0.032951682806015015, + 0.009713053703308105, + 0.007324875332415104, + 0.010861236602067947, + 0.00489308126270771, + 0.012254111468791962, + 0.004581140819936991, + 0.005849776323884726, + 0.011279088445007801, + 0.007015366107225418, + 0.005776693113148212, + 0.0055278753861784935, + 0.01478505041450262, + 0.010342861525714397, + 0.036363448947668076, + 0.010238034650683403, + 0.018360355868935585, + 0.02281361073255539, + 0.0075987656600773335, + 0.009380558505654335, + 0.00857221707701683, + 0.005192541982978582, + 0.004243230447173119, + 0.00489208847284317, + 0.003485951106995344, + 0.006050175055861473, + 0.004642253275960684, + 0.005366137716919184, + 0.010744755156338215, + 0.01069522276520729, + 0.010682168416678905, + 0.009289943613111973, + 0.011898775584995747, + 0.016078028827905655, + 0.007892786525189877, + 0.009198300540447235, + 0.020961934700608253, + 0.022179242223501205, + 0.020453691482543945, + 0.021168068051338196, + 0.020966967567801476, + 0.05089811608195305, + 0.02431587316095829, + 0.031134245917201042, + 0.00936904177069664, + 0.02137831225991249, + 0.08675622195005417, + 0.06636162102222443, + 0.00882637407630682, + 0.008720162324607372, + 0.007096041459590197, + 0.005459648557007313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.011070732958614826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins in a mid-level shot, focusing on a driverless vehicle constructed from smooth white voxels. It glides smoothly along clear light-gray voxel guides on the floor within a remarkably clean and bright automated piggery. Securely seated within the vehicle are several adorable voxel piglets, composed entirely of pink blocks. They possess simple black voxel eyes, slightly protruding square snouts, and a simple, innocent demeanor. The camera then steadily and deliberately ascends, gradually expanding from this cozy corner. As it rises higher, the imposing structure of the entire facility unfolds: countless uniform compartments, separated by translucent voxel walls, are arranged like a matrix, stretching into the distance to form a multi-layered, bottomless, vast space. Other autonomous vehicles and robotic arms work in an orderly fashion on their respective tracks. The entire scene is illuminated by a uniform, soft white light emitted by luminous voxel panels on the ceiling, creating a near-absence of shadows and a high-tech, sterile atmosphere. Ultimately, the camera hovers in a spectacular bird's-eye view, revealing a stunning voxel world that blends cute aesthetics with futuristic industrial efficiency, creating a tranquil and surreal atmosphere.", + "seed": 3567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000155/group_9/index_1264_seed_3567_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": -0.012049825675785542, + "vlm_logprob": -0.012049825675785542, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.012049825675785542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.032284487038850784 + ], + "opposite_logprob_tokens": [ + -3.6572844982147217 + ], + "logprob_reward": -0.032284487038850784, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.032284487038850784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15413691103458405, + 0.04306913912296295, + 0.05964307487010956, + 0.065560482442379, + 0.225582093000412, + 0.05129086226224899, + 0.0764312669634819, + 0.05854096636176109, + 0.19757772982120514, + 0.06816751509904861 + ], + "credit_map": [ + 0.02845579758286476, + 0.022060971707105637, + 0.023719828575849533, + 0.01539637055248022, + 0.013952577486634254, + 0.02128382958471775, + 0.013403491117060184, + 0.015864035114645958, + 0.004658871795982122, + 0.006046787369996309, + 0.006050358526408672, + 0.004378407262265682, + 0.0060102250427007675, + 0.005784660577774048, + 0.005074913613498211, + 0.005064916796982288, + 0.006954339798539877, + 0.008004212751984596, + 0.008835063315927982, + 0.005689592566341162, + 0.008097010664641857, + 0.007236639969050884, + 0.0063050054013729095, + 0.008521213196218014, + 0.007739413529634476, + 0.006220137234777212, + 0.009907246567308903, + 0.006853543221950531, + 0.0077281068079173565, + 0.01486166287213564, + 0.005598284304141998, + 0.006652085576206446, + 0.025363720953464508, + 0.028567926958203316, + 0.021192945539951324, + 0.027461791411042213, + 0.044717270880937576, + 0.030545193701982498, + 0.024558767676353455, + 0.02317447029054165, + 0.01279900036752224, + 0.007802033331245184, + 0.006123580038547516, + 0.004541842732578516, + 0.006391214206814766, + 0.004390115849673748, + 0.004995909985154867, + 0.004247165750712156, + 0.00565952667966485, + 0.008573428727686405, + 0.01663924939930439, + 0.007696211338043213, + 0.005382787901908159, + 0.010064925067126751, + 0.013321992941200733, + 0.009093141183257103, + 0.008793927729129791, + 0.007191222161054611, + 0.012785515747964382, + 0.006796725559979677, + 0.004327071830630302, + 0.006551285274326801, + 0.006066528148949146, + 0.006028687581419945, + 0.0186690054833889, + 0.017483685165643692, + 0.03367449715733528, + 0.025279181078076363, + 0.019268397241830826, + 0.02131665125489235, + 0.02573883719742298, + 0.03614746779203415, + 0.008598478510975838, + 0.007125274743884802, + 0.006172447465360165, + 0.007816806435585022, + 0.010755987837910652, + 0.009676172398030758, + 0.007776498794555664, + 0.010245845653116703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.032284487038850784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.0028262692503631115 + ], + "opposite_logprob_tokens": [ + -5.87782621383667 + ], + "logprob_reward": -0.0028262692503631115, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028262692503631115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07161977887153625, + 0.05878246948122978, + 0.20355291664600372, + 0.1534508466720581, + 0.09808105230331421, + 0.07951118797063828, + 0.07071730494499207, + 0.11006221175193787, + 0.08962162584066391, + 0.0646006241440773 + ], + "credit_map": [ + 0.007186379283666611, + 0.012665805406868458, + 0.008670670911669731, + 0.013193061575293541, + 0.008398118428885937, + 0.006593084894120693, + 0.0069173360243439674, + 0.007995326071977615, + 0.006203112658113241, + 0.008926077745854855, + 0.008912337943911552, + 0.005121441092342138, + 0.008629880845546722, + 0.00747075118124485, + 0.0062814634293317795, + 0.007237403187900782, + 0.01378567423671484, + 0.03362240642309189, + 0.031003277748823166, + 0.025711707770824432, + 0.0358927883207798, + 0.012109016068279743, + 0.03069748543202877, + 0.020730558782815933, + 0.013793534599244595, + 0.021174585446715355, + 0.020730121061205864, + 0.017428791150450706, + 0.027994276955723763, + 0.019560925662517548, + 0.014489388093352318, + 0.018279220908880234, + 0.009963990189135075, + 0.009391049854457378, + 0.009989804588258266, + 0.02047514170408249, + 0.014738119207322598, + 0.013312656432390213, + 0.009072134271264076, + 0.011138156987726688, + 0.012586754746735096, + 0.013745534233748913, + 0.008724949322640896, + 0.008963976986706257, + 0.008472507819533348, + 0.007090462371706963, + 0.009879390709102154, + 0.010047614574432373, + 0.00511724129319191, + 0.006253525614738464, + 0.014771591871976852, + 0.008955755271017551, + 0.005917425733059645, + 0.008333885110914707, + 0.013708729296922684, + 0.007659156806766987, + 0.020917445421218872, + 0.009853394702076912, + 0.01880696415901184, + 0.01193135417997837, + 0.008277990855276585, + 0.015478795394301414, + 0.01284598745405674, + 0.011950278654694557, + 0.00817361194640398, + 0.009559962898492813, + 0.014254482463002205, + 0.010796881280839443, + 0.01066047977656126, + 0.009424453601241112, + 0.010576585307717323, + 0.016175169497728348, + 0.007850716821849346, + 0.012332544662058353, + 0.006110349670052528, + 0.007057459559291601, + 0.012610838748514652, + 0.006277319509536028, + 0.006575172301381826, + 0.00578622380271554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.0028262692503631115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.00977727584540844 + ], + "opposite_logprob_tokens": [ + -4.634777069091797 + ], + "logprob_reward": -0.00977727584540844, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977727584540844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10150806605815887, + 0.11874837428331375, + 0.06416518241167068, + 0.06096971035003662, + 0.10672454535961151, + 0.30709466338157654, + 0.07035070657730103, + 0.0854141041636467, + 0.057087481021881104, + 0.027937157079577446 + ], + "credit_map": [ + 0.010054122656583786, + 0.0170829426497221, + 0.012827177532017231, + 0.01946556568145752, + 0.012364774011075497, + 0.007655156310647726, + 0.015283839777112007, + 0.006774492096155882, + 0.012331432662904263, + 0.016638392582535744, + 0.015347074717283249, + 0.014370600692927837, + 0.019520500674843788, + 0.009747996926307678, + 0.020214559510350227, + 0.01057782955467701, + 0.008060426451265812, + 0.009283075109124184, + 0.01074680034071207, + 0.006087118294090033, + 0.005850841291248798, + 0.0062308949418365955, + 0.00558900460600853, + 0.01231702696532011, + 0.014755639247596264, + 0.006279544439166784, + 0.00957022700458765, + 0.0071290466003119946, + 0.005659450776875019, + 0.008386276662349701, + 0.004005298484116793, + 0.0051842350512743, + 0.011210611090064049, + 0.01247388031333685, + 0.012835833244025707, + 0.013004045933485031, + 0.02161053568124771, + 0.013921506702899933, + 0.01175264548510313, + 0.009915493428707123, + 0.04223904013633728, + 0.03751467540860176, + 0.04585561156272888, + 0.046768784523010254, + 0.03149828314781189, + 0.02742619812488556, + 0.03715679422020912, + 0.03863527625799179, + 0.008731349371373653, + 0.007893029600381851, + 0.014334573410451412, + 0.0069760773330926895, + 0.0053418525494635105, + 0.008812381885945797, + 0.009222869761288166, + 0.009038583375513554, + 0.008626810275018215, + 0.009593074209988117, + 0.011788641102612019, + 0.013814542442560196, + 0.009558657184243202, + 0.007403042633086443, + 0.017247900366783142, + 0.007381437346339226, + 0.00632348982617259, + 0.004643576219677925, + 0.012217598035931587, + 0.006570599973201752, + 0.006456338334828615, + 0.005698481108993292, + 0.007848648354411125, + 0.007328754290938377, + 0.0036173781845718622, + 0.003646573983132839, + 0.0029994132928550243, + 0.004366695880889893, + 0.003221032442525029, + 0.004195399582386017, + 0.0022042065393179655, + 0.003686461830511689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.00977727584540844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.006728255655616522 + ], + "opposite_logprob_tokens": [ + -5.006728172302246 + ], + "logprob_reward": -0.006728255655616522, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006728255655616522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05860493704676628, + 0.03510886803269386, + 0.05251726508140564, + 0.041162021458148956, + 0.09781815856695175, + 0.1286318004131317, + 0.06130265071988106, + 0.3261944353580475, + 0.1446831375360489, + 0.05397675558924675 + ], + "credit_map": [ + 0.005989712662994862, + 0.00955752283334732, + 0.007604312617331743, + 0.009371954947710037, + 0.007973750121891499, + 0.006070839241147041, + 0.0073630004189908504, + 0.00467384047806263, + 0.0030950261279940605, + 0.004625496454536915, + 0.004569568671286106, + 0.004937958437949419, + 0.005319369025528431, + 0.003742180997505784, + 0.006092648953199387, + 0.0027266154065728188, + 0.005275900475680828, + 0.007417247164994478, + 0.008384494110941887, + 0.0047248778864741325, + 0.008996807038784027, + 0.005494855809956789, + 0.004201748874038458, + 0.008021336980164051, + 0.004376978147774935, + 0.003682914189994335, + 0.006543810945004225, + 0.0050882925279438496, + 0.004669449292123318, + 0.007706820964813232, + 0.003372992156073451, + 0.0057207634672522545, + 0.01128208264708519, + 0.010982614941895008, + 0.012111280113458633, + 0.011482618749141693, + 0.02067357301712036, + 0.013478415086865425, + 0.009346380829811096, + 0.008461185730993748, + 0.022496690973639488, + 0.016491448506712914, + 0.017925076186656952, + 0.01825454831123352, + 0.013279673643410206, + 0.010441356338560581, + 0.01478748582303524, + 0.014955522492527962, + 0.008176444098353386, + 0.00824461504817009, + 0.015048996545374393, + 0.005045656114816666, + 0.004524331074208021, + 0.0067985085770487785, + 0.007141816429793835, + 0.006322281900793314, + 0.05501989647746086, + 0.02858986333012581, + 0.05501991882920265, + 0.03631926700472832, + 0.025433078408241272, + 0.0390888936817646, + 0.050452422350645065, + 0.036271099001169205, + 0.01476766262203455, + 0.01314533967524767, + 0.020954418927431107, + 0.020176561549305916, + 0.013389726169407368, + 0.01684393361210823, + 0.020673546940088272, + 0.02473195642232895, + 0.006092256400734186, + 0.0074215419590473175, + 0.0052797384560108185, + 0.007393346168100834, + 0.0074602230452001095, + 0.006074120756238699, + 0.007913433015346527, + 0.006342094391584396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.006728255655616522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a driverless vehicle constructed from smooth white voxels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several adorable voxel piglets composed of pink blocks securely seated inside the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin in a mid-level shot and then steadily ascend to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a clean, bright automated piggery with uniform compartments separated by translucent voxel walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by uniform soft white light from ceiling panels creating a near-absence of 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": [ + -0.00863284058868885 + ], + "opposite_logprob_tokens": [ + -4.758632659912109 + ], + "logprob_reward": -0.00863284058868885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863284058868885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11478153616189957, + 0.07188382744789124, + 0.05604632571339607, + 0.08651179075241089, + 0.13558384776115417, + 0.13295263051986694, + 0.05970218777656555, + 0.06119174510240555, + 0.1897171288728714, + 0.09162899851799011 + ], + "credit_map": [ + 0.01040724292397499, + 0.016456350684165955, + 0.015752848237752914, + 0.020855005830526352, + 0.012017794884741306, + 0.007881700061261654, + 0.018584704026579857, + 0.012825882993638515, + 0.0064203329384326935, + 0.010406509973108768, + 0.009526010602712631, + 0.00843412708491087, + 0.011538440361618996, + 0.006349072325974703, + 0.012885695323348045, + 0.006323634646832943, + 0.005368608981370926, + 0.010636516846716404, + 0.008031755685806274, + 0.006167733576148748, + 0.008876066654920578, + 0.0044654724188148975, + 0.006388900335878134, + 0.006111267488449812, + 0.013591203838586807, + 0.007548913825303316, + 0.01240230817347765, + 0.008136694319546223, + 0.009760655462741852, + 0.020603211596608162, + 0.0054127830080688, + 0.00905601866543293, + 0.01708844304084778, + 0.011642114259302616, + 0.018711484968662262, + 0.011952622793614864, + 0.021451212465763092, + 0.019383057951927185, + 0.01448479201644659, + 0.020870117470622063, + 0.02114374190568924, + 0.015252945944666862, + 0.020174233242869377, + 0.016066715121269226, + 0.02001276984810829, + 0.010961887426674366, + 0.013941630721092224, + 0.015398722141981125, + 0.009150314144790173, + 0.008686329238116741, + 0.013020139187574387, + 0.004711516201496124, + 0.005043238867074251, + 0.0072293831035494804, + 0.005353689193725586, + 0.006507574580609798, + 0.008417675271630287, + 0.006358750630170107, + 0.010833323001861572, + 0.008915024809539318, + 0.004172711167484522, + 0.006053583696484566, + 0.009076242335140705, + 0.0073644379153847694, + 0.02322583831846714, + 0.021109944209456444, + 0.03995789214968681, + 0.022985385730862617, + 0.01362551935017109, + 0.02507537230849266, + 0.02016221359372139, + 0.02357495203614235, + 0.010129340924322605, + 0.010253241285681725, + 0.010134621523320675, + 0.01091085560619831, + 0.01367309968918562, + 0.012896476313471794, + 0.010530945844948292, + 0.013100420124828815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.00863284058868885 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file