{"artifacts":{"multimodal":{"bytes":149750,"content_store":{"file_count":400,"path":"multimodal/content-store","total_bytes":408553125,"total_pixels":1206495736},"path":"multimodal/mmmu-pro-vision.jsonl","schedule":{"bytes":243675,"path":"multimodal/mmmu-pro-vision.schedule.json","sha256":"f4fa52e6a24d1a0636b346a60cd4d390cb88ecad9f012297ce6c952423aaca9c"},"schedule_sha256":"62a4d7e96a7d83979d6888f89b3040741c83c66e371408b4b3dd88af3c660b1d","sha256":"fb8397e747ae5f49056ae607b058df4f54832f0e16597538926e4f06eff09d44"},"multimodal_semantic_bank":{"bytes":239109,"path":"multimodal/semantic-bank.jsonl","sha256":"942433c1f933c0bfcaf5b40900c463a397483aada198e63be27b19474689864c"},"text_mistral_instruct":{"bytes":379746,"path":"text/mistral-instruct.jsonl","schedule":{"bytes":467270,"path":"text/mistral-instruct.schedule.json","sha256":"86e6e45c8a2608ac917f5517c4c1145f3723f6e5ec9cd460e4c7cc318ef902f7"},"schedule_sha256":"c780a66ca180bbdba20884775032d3a1a1a3ff5712c54d18d7def58660f7b77b","sha256":"e7dac59f42816cd68ee6a670b0dbbb55b6e6bd06844951940a9bc0436fd72859"},"text_qwen_instruct":{"bytes":396146,"path":"text/qwen-instruct.jsonl","schedule":{"bytes":483667,"path":"text/qwen-instruct.schedule.json","sha256":"178f6be9386fcfbb54a080421ac84e17d47ed2578afda77f6b77bb0d9220ad4c"},"schedule_sha256":"1a2653d04b618057292676a6d795dda4de3d36d9b5cb5fb9a8e4b40a009db572","sha256":"52b568fcbead27884b1c8e375c4e05111bcae25e40000e23c770675869e4a5b8"},"text_raw_causal":{"bytes":305746,"path":"text/raw-causal.jsonl","schedule":{"bytes":393264,"path":"text/raw-causal.schedule.json","sha256":"ab209cd080884279e3b33529b5ea1e25ec96375010d3760c906b50e178a318a7"},"schedule_sha256":"3552a65f461677b7f6a27d24c157f043a6ff5b016f2ddc5834e5f304c7d8be00","sha256":"e4e839b3df7993a228765aa5305c5b7bb01e06a7b79554e3b2abe678b4deb63a"},"text_semantic_bank":{"bytes":339013,"path":"text/semantic-bank.jsonl","sha256":"18a88db999d8157ef051fee0eac4ad48b291853c970a5dc709d40b35e2da4430"}},"distributions":{"multimodal_answers":{"A":40,"B":40,"C":40,"D":40,"E":40,"F":40,"G":40,"H":40,"I":40,"J":40},"multimodal_groups":{"Accounting":14,"Agriculture":14,"Architecture_and_Engineering":14,"Art":14,"Art_Theory":14,"Basic_Medical_Science":14,"Biology":14,"Chemistry":14,"Clinical_Medicine":14,"Computer_Science":14,"Design":13,"Diagnostics_and_Laboratory_Medicine":13,"Economics":13,"Electronics":13,"Energy_and_Power":13,"Finance":13,"Geography":13,"History":13,"Literature":13,"Manage":13,"Marketing":13,"Materials":13,"Math":13,"Mechanical_Engineering":13,"Music":13,"Pharmacy":13,"Physics":13,"Psychology":13,"Public_Health":13,"Sociology":13},"text_answers":{"A":40,"B":40,"C":40,"D":40,"E":40,"F":40,"G":40,"H":40,"I":40,"J":40},"text_groups":{"biology":29,"business":29,"chemistry":29,"computer science":29,"economics":29,"engineering":29,"health":29,"history":29,"law":28,"math":28,"other":28,"philosophy":28,"physics":28,"psychology":28}},"format_version":"invarlock/qualification-suites-v1","normalization_algorithm":"outer-whitespace-only-v1","record_count":400,"selected_ids":{"multimodal":["test_Architecture_and_Engineering_493","test_Literature_36","test_Design_159","validation_Electronics_14","test_Chemistry_447","test_Mechanical_Engineering_104","validation_Physics_8","test_Geography_412","test_Clinical_Medicine_127","test_Physics_307","validation_Literature_11","test_Literature_38","test_Accounting_312","test_Design_4","test_Economics_90","test_Art_35","validation_Art_Theory_6","test_Architecture_and_Engineering_113","validation_Architecture_and_Engineering_17","test_Art_139","test_Architecture_and_Engineering_473","test_Agriculture_262","test_Electronics_235","validation_Finance_27","test_Sociology_68","test_Architecture_and_Engineering_416","validation_Biology_24","test_Clinical_Medicine_148","test_Materials_392","test_Chemistry_40","validation_Literature_10","test_Architecture_and_Engineering_111","validation_Geography_1","test_Computer_Science_77","test_Agriculture_195","validation_Clinical_Medicine_8","test_Public_Health_248","test_Psychology_78","test_Music_310","test_Electronics_68","validation_Physics_20","validation_Sociology_11","test_Computer_Science_151","validation_Materials_21","test_Marketing_150","validation_Literature_8","test_Computer_Science_353","validation_Public_Health_15","test_Finance_5","test_Math_165","test_Public_Health_266","test_Literature_56","test_Mechanical_Engineering_279","validation_Accounting_24","validation_Marketing_4","test_Marketing_135","test_Basic_Medical_Science_209","test_Biology_76","validation_Computer_Science_22","test_Biology_316","test_Economics_137","test_Computer_Science_307","test_Psychology_106","test_Diagnostics_and_Laboratory_Medicine_133","validation_Chemistry_26","validation_Architecture_and_Engineering_26","test_Manage_69","test_History_72","test_Psychology_223","validation_History_15","test_Art_107","test_Chemistry_102","test_Economics_260","test_Geography_301","test_Marketing_61","test_History_79","validation_Literature_7","test_Materials_253","test_Design_146","validation_Energy_and_Power_9","test_Electronics_213","test_Pharmacy_231","test_Accounting_252","validation_Literature_22","validation_Agriculture_11","test_Chemistry_5","validation_Energy_and_Power_6","test_Biology_196","test_Computer_Science_351","test_Literature_60","validation_Mechanical_Engineering_27","test_Design_5","test_Pharmacy_287","test_Electronics_110","validation_Pharmacy_27","test_History_8","validation_Physics_30","validation_Materials_29","test_Sociology_122","test_Pharmacy_140","test_Art_147","test_Geography_8","test_Manage_126","test_Diagnostics_and_Laboratory_Medicine_131","validation_Basic_Medical_Science_3","validation_Sociology_28","test_Art_Theory_246","test_Pharmacy_318","validation_Music_18","test_Economics_162","validation_Materials_18","test_Basic_Medical_Science_253","validation_Accounting_19","test_History_131","test_Finance_238","validation_Public_Health_2","test_Pharmacy_34","test_Electronics_93","test_Biology_302","test_Materials_139","test_Basic_Medical_Science_177","validation_Mechanical_Engineering_25","test_Basic_Medical_Science_270","test_Basic_Medical_Science_313","test_Finance_38","test_Agriculture_139","test_Energy_and_Power_423","test_Energy_and_Power_150","validation_Diagnostics_and_Laboratory_Medicine_5","validation_Agriculture_3","test_Math_41","validation_Chemistry_8","validation_Math_22","test_Electronics_123","validation_Basic_Medical_Science_28","test_History_87","validation_History_8","validation_Marketing_15","test_History_74","test_Computer_Science_70","test_Materials_91","test_Electronics_245","test_Basic_Medical_Science_261","test_Accounting_279","test_Geography_518","test_Chemistry_147","validation_Music_7","test_Architecture_and_Engineering_191","validation_Diagnostics_and_Laboratory_Medicine_13","validation_Art_Theory_28","validation_Math_1","test_History_12","validation_Sociology_20","test_Design_96","test_History_44","validation_Sociology_29","validation_Sociology_22","test_Psychology_285","test_Clinical_Medicine_128","validation_Clinical_Medicine_23","test_Chemistry_570","validation_Pharmacy_22","test_Mechanical_Engineering_186","test_Art_Theory_171","validation_Public_Health_24","test_Finance_250","test_Clinical_Medicine_298","test_Art_121","test_Psychology_270","test_History_27","test_Psychology_56","validation_Manage_26","test_Economics_94","validation_Energy_and_Power_24","test_Pharmacy_174","test_Economics_47","test_Physics_293","test_Music_127","test_Mechanical_Engineering_123","test_Art_159","test_Physics_243","validation_Clinical_Medicine_13","validation_Marketing_9","test_Biology_341","validation_Energy_and_Power_4","validation_Manage_30","test_Computer_Science_250","validation_Art_Theory_30","validation_Chemistry_3","test_Diagnostics_and_Laboratory_Medicine_68","test_Mechanical_Engineering_338","validation_Art_25","test_Design_53","test_Computer_Science_140","test_Accounting_131","validation_Agriculture_13","test_Electronics_188","validation_Art_14","test_Chemistry_233","test_Geography_55","test_Electronics_172","test_Clinical_Medicine_115","test_Finance_163","test_Physics_174","test_Psychology_33","validation_Accounting_3","test_Art_41","validation_Finance_19","validation_Marketing_30","test_Energy_and_Power_180","validation_Literature_2","test_Public_Health_199","validation_Diagnostics_and_Laboratory_Medicine_14","test_Geography_241","test_Computer_Science_226","test_Art_Theory_208","test_Manage_173","test_Economics_16","validation_Diagnostics_and_Laboratory_Medicine_30","validation_Energy_and_Power_22","test_Electronics_254","validation_Music_2","test_Clinical_Medicine_172","validation_Public_Health_16","validation_Chemistry_28","test_Finance_138","test_Clinical_Medicine_312","test_Marketing_62","validation_Music_22","test_Basic_Medical_Science_41","validation_Materials_19","test_Materials_345","test_Diagnostics_and_Laboratory_Medicine_9","test_Design_125","validation_Electronics_19","validation_Design_29","test_Biology_79","test_Agriculture_146","test_Art_Theory_199","test_Architecture_and_Engineering_250","test_Marketing_134","test_Chemistry_189","validation_Biology_30","test_Pharmacy_414","test_Marketing_112","test_Marketing_172","validation_Art_Theory_25","validation_Materials_7","validation_Biology_22","test_Physics_292","test_Finance_290","test_Mechanical_Engineering_55","validation_Math_30","test_Agriculture_15","test_Design_72","validation_Math_2","test_Architecture_and_Engineering_533","test_Marketing_69","validation_Finance_29","test_Psychology_155","validation_Agriculture_9","test_Energy_and_Power_414","validation_Pharmacy_17","validation_Clinical_Medicine_1","validation_Public_Health_28","test_Math_22","validation_Chemistry_22","test_Finance_10","test_Manage_56","validation_Materials_6","test_Accounting_34","test_Physics_210","validation_Mechanical_Engineering_24","test_Design_39","test_Art_Theory_24","test_Biology_156","test_Sociology_229","test_Geography_462","validation_Electronics_17","test_Clinical_Medicine_45","test_Physics_369","validation_Public_Health_20","test_Architecture_and_Engineering_361","test_Math_205","test_Clinical_Medicine_73","test_Art_94","test_Math_115","test_Psychology_258","test_Accounting_160","test_Geography_298","test_Finance_204","test_Music_134","test_Economics_188","test_Accounting_328","test_Music_293","test_Art_24","test_Design_135","test_Agriculture_165","validation_Marketing_29","test_Economics_22","validation_Art_4","test_Design_84","test_Mechanical_Engineering_275","test_Finance_68","test_Diagnostics_and_Laboratory_Medicine_101","test_Economics_195","test_Agriculture_269","test_Public_Health_463","test_Biology_146","test_Computer_Science_51","test_Biology_295","test_Mechanical_Engineering_81","test_Computer_Science_227","test_Materials_314","test_Psychology_105","test_Public_Health_28","test_Computer_Science_369","test_Biology_138","test_Agriculture_234","test_Sociology_63","test_History_125","validation_Pharmacy_7","test_Economics_97","test_Basic_Medical_Science_230","test_Accounting_201","test_Math_257","test_Architecture_and_Engineering_443","validation_Math_28","test_Literature_4","test_Manage_115","test_Mechanical_Engineering_136","validation_Accounting_26","test_Math_11","test_Clinical_Medicine_247","test_Geography_317","test_Music_317","validation_Diagnostics_and_Laboratory_Medicine_16","validation_Art_Theory_23","test_Basic_Medical_Science_233","validation_Art_15","validation_Accounting_8","test_Psychology_128","test_Energy_and_Power_371","test_Geography_191","test_Art_Theory_156","validation_Art_26","validation_Literature_4","test_Chemistry_314","test_Music_324","test_Energy_and_Power_191","test_Economics_74","validation_Manage_12","validation_Public_Health_17","validation_Literature_30","test_Architecture_and_Engineering_312","test_Basic_Medical_Science_76","validation_Pharmacy_11","test_Physics_61","test_Physics_85","test_Energy_and_Power_68","validation_Art_Theory_10","validation_Music_8","test_Accounting_63","validation_Psychology_8","validation_Pharmacy_18","test_History_76","validation_Music_9","validation_Music_6","test_Agriculture_270","test_Computer_Science_276","test_Agriculture_178","test_Math_201","test_Sociology_47","test_Sociology_181","validation_Basic_Medical_Science_23","test_Manage_73","validation_Sociology_15","test_Architecture_and_Engineering_470","test_Basic_Medical_Science_240","test_Art_Theory_273","validation_Art_Theory_16","test_Energy_and_Power_158","test_Psychology_163","test_Materials_415","test_Manage_36","test_Manage_64","test_Economics_243","validation_Diagnostics_and_Laboratory_Medicine_3","test_Mechanical_Engineering_234","validation_Public_Health_9","validation_Sociology_5","validation_Biology_14","validation_Physics_2","test_Manage_35","validation_Diagnostics_and_Laboratory_Medicine_8","test_Geography_267","validation_Manage_24","validation_Geography_7","validation_Design_15","test_Diagnostics_and_Laboratory_Medicine_52"],"text":["mmlu_pro_00075","mmlu_pro_00087","mmlu_pro_00111","mmlu_pro_00154","mmlu_pro_00166","mmlu_pro_00202","mmlu_pro_00250","mmlu_pro_00257","mmlu_pro_00284","mmlu_pro_00292","mmlu_pro_00294","mmlu_pro_00306","mmlu_pro_00323","mmlu_pro_00346","mmlu_pro_00372","mmlu_pro_00375","mmlu_pro_00436","mmlu_pro_00536","mmlu_pro_00564","mmlu_pro_00569","mmlu_pro_00575","mmlu_pro_00614","mmlu_pro_00650","mmlu_pro_00663","mmlu_pro_00685","mmlu_pro_00695","mmlu_pro_00700","mmlu_pro_00807","mmlu_pro_00831","mmlu_pro_00934","mmlu_pro_00941","mmlu_pro_00947","mmlu_pro_01004","mmlu_pro_01020","mmlu_pro_01092","mmlu_pro_01134","mmlu_pro_01179","mmlu_pro_01244","mmlu_pro_01284","mmlu_pro_01337","mmlu_pro_01368","mmlu_pro_01387","mmlu_pro_01413","mmlu_pro_01419","mmlu_pro_01423","mmlu_pro_01449","mmlu_pro_01529","mmlu_pro_01578","mmlu_pro_01660","mmlu_pro_01717","mmlu_pro_01739","mmlu_pro_01761","mmlu_pro_01851","mmlu_pro_01866","mmlu_pro_01894","mmlu_pro_01940","mmlu_pro_01974","mmlu_pro_02018","mmlu_pro_02033","mmlu_pro_02086","mmlu_pro_02090","mmlu_pro_02091","mmlu_pro_02110","mmlu_pro_02147","mmlu_pro_02188","mmlu_pro_02203","mmlu_pro_02225","mmlu_pro_02228","mmlu_pro_02236","mmlu_pro_02238","mmlu_pro_02256","mmlu_pro_02321","mmlu_pro_02356","mmlu_pro_02459","mmlu_pro_02483","mmlu_pro_02519","mmlu_pro_02528","mmlu_pro_02601","mmlu_pro_02626","mmlu_pro_02635","mmlu_pro_02645","mmlu_pro_02667","mmlu_pro_02670","mmlu_pro_02706","mmlu_pro_02728","mmlu_pro_02820","mmlu_pro_02838","mmlu_pro_02851","mmlu_pro_02854","mmlu_pro_02872","mmlu_pro_02905","mmlu_pro_02961","mmlu_pro_02965","mmlu_pro_02986","mmlu_pro_02993","mmlu_pro_03009","mmlu_pro_03030","mmlu_pro_03057","mmlu_pro_03073","mmlu_pro_03074","mmlu_pro_03106","mmlu_pro_03124","mmlu_pro_03140","mmlu_pro_03214","mmlu_pro_03221","mmlu_pro_03279","mmlu_pro_03310","mmlu_pro_03373","mmlu_pro_03415","mmlu_pro_03434","mmlu_pro_03472","mmlu_pro_03493","mmlu_pro_03509","mmlu_pro_03513","mmlu_pro_03582","mmlu_pro_03597","mmlu_pro_03602","mmlu_pro_03663","mmlu_pro_03690","mmlu_pro_03721","mmlu_pro_03752","mmlu_pro_03766","mmlu_pro_03780","mmlu_pro_03797","mmlu_pro_03819","mmlu_pro_03861","mmlu_pro_03871","mmlu_pro_03875","mmlu_pro_03900","mmlu_pro_03908","mmlu_pro_03931","mmlu_pro_03992","mmlu_pro_03998","mmlu_pro_04011","mmlu_pro_04090","mmlu_pro_04129","mmlu_pro_04168","mmlu_pro_04206","mmlu_pro_04234","mmlu_pro_04267","mmlu_pro_04379","mmlu_pro_04628","mmlu_pro_04653","mmlu_pro_04684","mmlu_pro_04708","mmlu_pro_04715","mmlu_pro_04739","mmlu_pro_04748","mmlu_pro_04760","mmlu_pro_04778","mmlu_pro_04828","mmlu_pro_04831","mmlu_pro_04836","mmlu_pro_04837","mmlu_pro_04861","mmlu_pro_04864","mmlu_pro_04881","mmlu_pro_04891","mmlu_pro_04898","mmlu_pro_04914","mmlu_pro_04929","mmlu_pro_04941","mmlu_pro_04942","mmlu_pro_04943","mmlu_pro_04951","mmlu_pro_04962","mmlu_pro_04965","mmlu_pro_04988","mmlu_pro_04993","mmlu_pro_04994","mmlu_pro_05017","mmlu_pro_05044","mmlu_pro_05141","mmlu_pro_05238","mmlu_pro_05266","mmlu_pro_05268","mmlu_pro_05276","mmlu_pro_05307","mmlu_pro_05383","mmlu_pro_05449","mmlu_pro_05471","mmlu_pro_05474","mmlu_pro_05507","mmlu_pro_05527","mmlu_pro_05537","mmlu_pro_05543","mmlu_pro_05628","mmlu_pro_05675","mmlu_pro_05694","mmlu_pro_05706","mmlu_pro_05752","mmlu_pro_05776","mmlu_pro_05782","mmlu_pro_05785","mmlu_pro_05879","mmlu_pro_05910","mmlu_pro_05927","mmlu_pro_05929","mmlu_pro_05974","mmlu_pro_05992","mmlu_pro_06001","mmlu_pro_06022","mmlu_pro_06064","mmlu_pro_06179","mmlu_pro_06198","mmlu_pro_06229","mmlu_pro_06266","mmlu_pro_06271","mmlu_pro_06297","mmlu_pro_06325","mmlu_pro_06340","mmlu_pro_06455","mmlu_pro_06457","mmlu_pro_06481","mmlu_pro_06485","mmlu_pro_06533","mmlu_pro_06536","mmlu_pro_06598","mmlu_pro_06615","mmlu_pro_06619","mmlu_pro_06625","mmlu_pro_06633","mmlu_pro_06652","mmlu_pro_06678","mmlu_pro_06702","mmlu_pro_06710","mmlu_pro_06738","mmlu_pro_06767","mmlu_pro_06781","mmlu_pro_06890","mmlu_pro_06891","mmlu_pro_06893","mmlu_pro_06948","mmlu_pro_06962","mmlu_pro_06990","mmlu_pro_06993","mmlu_pro_07081","mmlu_pro_07145","mmlu_pro_07147","mmlu_pro_07168","mmlu_pro_07272","mmlu_pro_07276","mmlu_pro_07296","mmlu_pro_07336","mmlu_pro_07369","mmlu_pro_07427","mmlu_pro_07465","mmlu_pro_07527","mmlu_pro_07552","mmlu_pro_07565","mmlu_pro_07568","mmlu_pro_07575","mmlu_pro_07595","mmlu_pro_07597","mmlu_pro_07608","mmlu_pro_07632","mmlu_pro_07639","mmlu_pro_07657","mmlu_pro_07742","mmlu_pro_07756","mmlu_pro_07779","mmlu_pro_07903","mmlu_pro_07939","mmlu_pro_08037","mmlu_pro_08073","mmlu_pro_08078","mmlu_pro_08114","mmlu_pro_08118","mmlu_pro_08163","mmlu_pro_08299","mmlu_pro_08332","mmlu_pro_08378","mmlu_pro_08446","mmlu_pro_08449","mmlu_pro_08459","mmlu_pro_08489","mmlu_pro_08554","mmlu_pro_08601","mmlu_pro_08679","mmlu_pro_08798","mmlu_pro_08836","mmlu_pro_08847","mmlu_pro_08868","mmlu_pro_08938","mmlu_pro_08981","mmlu_pro_08988","mmlu_pro_09138","mmlu_pro_09149","mmlu_pro_09251","mmlu_pro_09319","mmlu_pro_09332","mmlu_pro_09338","mmlu_pro_09433","mmlu_pro_09440","mmlu_pro_09461","mmlu_pro_09563","mmlu_pro_09592","mmlu_pro_09624","mmlu_pro_09664","mmlu_pro_09709","mmlu_pro_09727","mmlu_pro_09738","mmlu_pro_09810","mmlu_pro_09862","mmlu_pro_09908","mmlu_pro_09998","mmlu_pro_10035","mmlu_pro_10037","mmlu_pro_10125","mmlu_pro_10217","mmlu_pro_10223","mmlu_pro_10231","mmlu_pro_10243","mmlu_pro_10264","mmlu_pro_10360","mmlu_pro_10380","mmlu_pro_10411","mmlu_pro_10413","mmlu_pro_10414","mmlu_pro_10442","mmlu_pro_10460","mmlu_pro_10475","mmlu_pro_10500","mmlu_pro_10501","mmlu_pro_10508","mmlu_pro_10517","mmlu_pro_10521","mmlu_pro_10530","mmlu_pro_10531","mmlu_pro_10550","mmlu_pro_10586","mmlu_pro_10595","mmlu_pro_10602","mmlu_pro_10609","mmlu_pro_10620","mmlu_pro_10627","mmlu_pro_10637","mmlu_pro_10652","mmlu_pro_10700","mmlu_pro_10737","mmlu_pro_10745","mmlu_pro_10761","mmlu_pro_10771","mmlu_pro_10776","mmlu_pro_10791","mmlu_pro_10809","mmlu_pro_10812","mmlu_pro_10822","mmlu_pro_10856","mmlu_pro_10861","mmlu_pro_10875","mmlu_pro_10876","mmlu_pro_10881","mmlu_pro_10897","mmlu_pro_10901","mmlu_pro_10904","mmlu_pro_10912","mmlu_pro_10922","mmlu_pro_10923","mmlu_pro_10927","mmlu_pro_10949","mmlu_pro_11041","mmlu_pro_11095","mmlu_pro_11128","mmlu_pro_11135","mmlu_pro_11178","mmlu_pro_11195","mmlu_pro_11229","mmlu_pro_11230","mmlu_pro_11234","mmlu_pro_11268","mmlu_pro_11344","mmlu_pro_11363","mmlu_pro_11410","mmlu_pro_11472","mmlu_pro_11514","mmlu_pro_11515","mmlu_pro_11533","mmlu_pro_11539","mmlu_pro_11547","mmlu_pro_11549","mmlu_pro_11642","mmlu_pro_11671","mmlu_pro_11694","mmlu_pro_11716","mmlu_pro_11731","mmlu_pro_11770","mmlu_pro_11841","mmlu_pro_11886","mmlu_pro_11894","mmlu_pro_11916","mmlu_pro_11922","mmlu_pro_12055","mmlu_pro_12056","mmlu_pro_12084","mmlu_pro_12149","mmlu_pro_12197","mmlu_pro_12199","mmlu_pro_12225","mmlu_pro_12236"]},"selection_algorithm":"balanced-bipartite-sha256-v1","sources":{"multimodal":{"config":"vision","dataset":"MMMU/MMMU_Pro","eligible_record_count":1728,"exclusions":{"option_count_limit":1,"unsupported_media_type":1},"license":"Apache-2.0","revision":"563f3e84bb3b90893083a1f039cfa13077f2302b","source_record_count":1730,"split":"test"},"text":{"dataset":"TIGER-Lab/MMLU-Pro","eligible_record_count":12031,"exclusions":{"prompt_exceeds_maximum_input_tokens":1},"license":"MIT","revision":"b189ec765aa7ed75c8acfea42df31fdae71f97be","runtime_profile":{"maximum_input_tokens":1024,"name":"minimum-supported-text-runtime-v1","tokenizer_contract_sha256":"9afc79cdd1fc5f9d0ab0f2bf684f9d2754a3ab865c3a74e4705a684d66c4691f"},"source_record_count":12032,"split":"test"}}}