diff --git a/src/data/flip_labels/flip_labels_cross-router.json b/src/data/flip_labels/flip_labels_cross-router.json new file mode 100644 index 0000000..ad559c6 --- /dev/null +++ b/src/data/flip_labels/flip_labels_cross-router.json @@ -0,0 +1,1682 @@ +[ + { + "global index": "AIME_112", + "flip": 0 + }, + { + "global index": "AIME_58", + "flip": 0 + }, + { + "global index": "ArcMMLU_12", + "flip": 1 + }, + { + "global index": "ArcMMLU_123", + "flip": 0 + }, + { + "global index": "ArcMMLU_16", + "flip": 0 + }, + { + "global index": "ArcMMLU_182", + "flip": 1 + }, + { + "global index": "ArcMMLU_230", + "flip": 1 + }, + { + "global index": "ArcMMLU_293", + "flip": 0 + }, + { + "global index": "ArcMMLU_349", + "flip": 0 + }, + { + "global index": "ArcMMLU_378", + "flip": 0 + }, + { + "global index": "ArcMMLU_443", + "flip": 0 + }, + { + "global index": "ArcMMLU_496", + "flip": 0 + }, + { + "global index": "ArcMMLU_631", + "flip": 0 + }, + { + "global index": "ArcMMLU_646", + "flip": 0 + }, + { + "global index": "ArcMMLU_659", + "flip": 0 + }, + { + "global index": "ArcMMLU_676", + "flip": 0 + }, + { + "global index": "ArcMMLU_685", + "flip": 0 + }, + { + "global index": "ArcMMLU_689", + "flip": 1 + }, + { + "global index": "ArcMMLU_702", + "flip": 0 + }, + { + "global index": "ArcMMLU_713", + "flip": 0 + }, + { + "global index": "ArcMMLU_98", + "flip": 0 + }, + { + "global index": "AsDiv_1165", + "flip": 0 + }, + { + "global index": "AsDiv_1347", + "flip": 0 + }, + { + "global index": "AsDiv_472", + "flip": 0 + }, + { + "global index": "AsDiv_733", + "flip": 0 + }, + { + "global index": "ChessInstruct_0", + "flip": 0 + }, + { + "global index": "ChessInstruct_107", + "flip": 0 + }, + { + "global index": "ChessInstruct_144", + "flip": 0 + }, + { + "global index": "ChessInstruct_42", + "flip": 0 + }, + { + "global index": "ChessInstruct_58", + "flip": 0 + }, + { + "global index": "ChessInstruct_71", + "flip": 0 + }, + { + "global index": "ChessInstruct_84", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_28", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_51", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_6", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_62", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_70", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_85", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_90", + "flip": 1 + }, + { + "global index": "Ethics_deontology_0", + "flip": 0 + }, + { + "global index": "Ethics_deontology_2", + "flip": 0 + }, + { + "global index": "Ethics_deontology_31", + "flip": 1 + }, + { + "global index": "Ethics_deontology_32", + "flip": 0 + }, + { + "global index": "Ethics_deontology_56", + "flip": 0 + }, + { + "global index": "Ethics_justice_1", + "flip": 0 + }, + { + "global index": "Ethics_justice_45", + "flip": 1 + }, + { + "global index": "Ethics_justice_76", + "flip": 0 + }, + { + "global index": "Ethics_justice_84", + "flip": 0 + }, + { + "global index": "Ethics_virtue_14", + "flip": 0 + }, + { + "global index": "Ethics_virtue_26", + "flip": 0 + }, + { + "global index": "Ethics_virtue_30", + "flip": 1 + }, + { + "global index": "Ethics_virtue_48", + "flip": 0 + }, + { + "global index": "Ethics_virtue_51", + "flip": 1 + }, + { + "global index": "FinQA_149", + "flip": 1 + }, + { + "global index": "FinQA_208", + "flip": 1 + }, + { + "global index": "FinQA_56", + "flip": 1 + }, + { + "global index": "FinQA_60", + "flip": 0 + }, + { + "global index": "GSM8K_43", + "flip": 1 + }, + { + "global index": "GeoBench_1002", + "flip": 1 + }, + { + "global index": "GeoBench_1094", + "flip": 0 + }, + { + "global index": "GeoBench_1102", + "flip": 0 + }, + { + "global index": "GeoBench_1113", + "flip": 1 + }, + { + "global index": "GeoBench_124", + "flip": 1 + }, + { + "global index": "GeoBench_1243", + "flip": 0 + }, + { + "global index": "GeoBench_30", + "flip": 1 + }, + { + "global index": "GeoBench_502", + "flip": 1 + }, + { + "global index": "GeoBench_526", + "flip": 0 + }, + { + "global index": "GeoBench_591", + "flip": 1 + }, + { + "global index": "GeoBench_766", + "flip": 1 + }, + { + "global index": "GeoBench_87", + "flip": 0 + }, + { + "global index": "GeoBench_915", + "flip": 1 + }, + { + "global index": "GeoBench_944", + "flip": 0 + }, + { + "global index": "GeoBench_968", + "flip": 0 + }, + { + "global index": "GeoGraphyData_100k_27", + "flip": 1 + }, + { + "global index": "GeoGraphyData_100k_42", + "flip": 1 + }, + { + "global index": "LiveCodeBench_105", + "flip": 0 + }, + { + "global index": "LiveCodeBench_114", + "flip": 0 + }, + { + "global index": "LiveCodeBench_118", + "flip": 0 + }, + { + "global index": "LiveCodeBench_131", + "flip": 0 + }, + { + "global index": "LiveCodeBench_136", + "flip": 1 + }, + { + "global index": "LiveCodeBench_181", + "flip": 0 + }, + { + "global index": "LiveCodeBench_237", + "flip": 0 + }, + { + "global index": "LiveCodeBench_271", + "flip": 1 + }, + { + "global index": "LiveCodeBench_350", + "flip": 0 + }, + { + "global index": "LiveCodeBench_386", + "flip": 0 + }, + { + "global index": "LiveCodeBench_405", + "flip": 1 + }, + { + "global index": "LiveCodeBench_43", + "flip": 1 + }, + { + "global index": "LiveCodeBench_431", + "flip": 1 + }, + { + "global index": "LiveCodeBench_437", + "flip": 0 + }, + { + "global index": "LiveCodeBench_476", + "flip": 1 + }, + { + "global index": "LiveCodeBench_485", + "flip": 1 + }, + { + "global index": "LiveCodeBench_49", + "flip": 0 + }, + { + "global index": "LiveCodeBench_491", + "flip": 0 + }, + { + "global index": "LiveCodeBench_499", + "flip": 1 + }, + { + "global index": "MATH_108", + "flip": 0 + }, + { + "global index": "MATH_442", + "flip": 0 + }, + { + "global index": "MATH_53", + "flip": 0 + }, + { + "global index": "MMLUPro_biology_2808", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2912", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2980", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2985", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_3188", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_3215", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_3225", + "flip": 1 + }, + { + "global index": "MMLUPro_business_226", + "flip": 1 + }, + { + "global index": "MMLUPro_business_294", + "flip": 1 + }, + { + "global index": "MMLUPro_business_378", + "flip": 1 + }, + { + "global index": "MMLUPro_business_430", + "flip": 1 + }, + { + "global index": "MMLUPro_business_503", + "flip": 1 + }, + { + "global index": "MMLUPro_business_507", + "flip": 1 + }, + { + "global index": "MMLUPro_business_6", + "flip": 0 + }, + { + "global index": "MMLUPro_business_784", + "flip": 0 + }, + { + "global index": "MMLUPro_chemistry_3796", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_3837", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_3974", + "flip": 0 + }, + { + "global index": "MMLUPro_chemistry_4067", + "flip": 0 + }, + { + "global index": "MMLUPro_chemistry_4407", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9086", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9110", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9136", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9138", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9149", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9200", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9212", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9239", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9264", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9285", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9289", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9414", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9415", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9430", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9452", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9471", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9475", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_5769", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_5907", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_5931", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_5965", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_6114", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_6122", + "flip": 0 + }, + { + "global index": "MMLUPro_economics_6135", + "flip": 0 + }, + { + "global index": "MMLUPro_economics_6325", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_6353", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10076", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10125", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10179", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10195", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10199", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10298", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10342", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10395", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10428", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10432", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10473", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10537", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10701", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10823", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10864", + "flip": 0 + }, + { + "global index": "MMLUPro_health_4885", + "flip": 1 + }, + { + "global index": "MMLUPro_health_4973", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5093", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5144", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5214", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5215", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5261", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5473", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5514", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4486", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4490", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4497", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4509", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4517", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4523", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4605", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4629", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4638", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4717", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4749", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4752", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4774", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4810", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4833", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4836", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4841", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1007", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1031", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1386", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1462", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1484", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1518", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1818", + "flip": 1 + }, + { + "global index": "MMLUPro_law_806", + "flip": 1 + }, + { + "global index": "MMLUPro_law_899", + "flip": 0 + }, + { + "global index": "MMLUPro_math_6429", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6526", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6623", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6848", + "flip": 0 + }, + { + "global index": "MMLUPro_math_7101", + "flip": 0 + }, + { + "global index": "MMLUPro_math_7249", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7284", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7451", + "flip": 0 + }, + { + "global index": "MMLUPro_math_7577", + "flip": 0 + }, + { + "global index": "MMLUPro_philosophy_9510", + "flip": 0 + }, + { + "global index": "MMLUPro_philosophy_9536", + "flip": 1 + }, + { + "global index": "MMLUPro_philosophy_9663", + "flip": 1 + }, + { + "global index": "MMLUPro_philosophy_9672", + "flip": 1 + }, + { + "global index": "MMLUPro_philosophy_9943", + "flip": 0 + }, + { + "global index": "MMLUPro_physics_7773", + "flip": 0 + }, + { + "global index": "MMLUPro_physics_7887", + "flip": 1 + }, + { + "global index": "MMLUPro_physics_7893", + "flip": 0 + }, + { + "global index": "MMLUPro_physics_8888", + "flip": 0 + }, + { + "global index": "MMLUPro_physics_9017", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2005", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2186", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2329", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2406", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2420", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2450", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2457", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2524", + "flip": 1 + }, + { + "global index": "MMLU_formal_logic_121", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_16", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_32", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_63", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_7", + "flip": 1 + }, + { + "global index": "MMLU_formal_logic_70", + "flip": 1 + }, + { + "global index": "MMLU_formal_logic_85", + "flip": 0 + }, + { + "global index": "MMLU_management_4", + "flip": 1 + }, + { + "global index": "MMLU_management_41", + "flip": 0 + }, + { + "global index": "MMLU_management_77", + "flip": 1 + }, + { + "global index": "MMLU_management_91", + "flip": 1 + }, + { + "global index": "MMLU_management_93", + "flip": 1 + }, + { + "global index": "MathQA_158", + "flip": 1 + }, + { + "global index": "MathQA_1742", + "flip": 1 + }, + { + "global index": "MathQA_202", + "flip": 0 + }, + { + "global index": "MathQA_2092", + "flip": 0 + }, + { + "global index": "MathQA_2102", + "flip": 1 + }, + { + "global index": "MathQA_2851", + "flip": 1 + }, + { + "global index": "MathQA_827", + "flip": 0 + }, + { + "global index": "MathQA_84", + "flip": 0 + }, + { + "global index": "MedMCQA_1005", + "flip": 1 + }, + { + "global index": "MedMCQA_1054", + "flip": 1 + }, + { + "global index": "MedMCQA_1298", + "flip": 1 + }, + { + "global index": "MedMCQA_1309", + "flip": 0 + }, + { + "global index": "MedMCQA_1362", + "flip": 1 + }, + { + "global index": "MedMCQA_145", + "flip": 1 + }, + { + "global index": "MedMCQA_2010", + "flip": 0 + }, + { + "global index": "MedMCQA_2323", + "flip": 1 + }, + { + "global index": "MedMCQA_2366", + "flip": 1 + }, + { + "global index": "MedMCQA_2581", + "flip": 0 + }, + { + "global index": "MedMCQA_511", + "flip": 0 + }, + { + "global index": "MedMCQA_59", + "flip": 0 + }, + { + "global index": "MedMCQA_643", + "flip": 0 + }, + { + "global index": "MedMCQA_853", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_126", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_14", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_147", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_152", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_188", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_189", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_240", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_33", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_337", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_340", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_70", + "flip": 0 + }, + { + "global index": "NarrativeQA_131", + "flip": 0 + }, + { + "global index": "NarrativeQA_1683", + "flip": 0 + }, + { + "global index": "NarrativeQA_2474", + "flip": 0 + }, + { + "global index": "NarrativeQA_2820", + "flip": 1 + }, + { + "global index": "NarrativeQA_3282", + "flip": 0 + }, + { + "global index": "NarrativeQA_4102", + "flip": 1 + }, + { + "global index": "NarrativeQA_4128", + "flip": 0 + }, + { + "global index": "NarrativeQA_4347", + "flip": 0 + }, + { + "global index": "NarrativeQA_4540", + "flip": 0 + }, + { + "global index": "NarrativeQA_5022", + "flip": 1 + }, + { + "global index": "NarrativeQA_5259", + "flip": 0 + }, + { + "global index": "NarrativeQA_533", + "flip": 0 + }, + { + "global index": "NarrativeQA_5894", + "flip": 1 + }, + { + "global index": "NarrativeQA_6829", + "flip": 0 + }, + { + "global index": "NarrativeQA_7678", + "flip": 1 + }, + { + "global index": "NarrativeQA_7964", + "flip": 0 + }, + { + "global index": "NarrativeQA_8215", + "flip": 0 + }, + { + "global index": "NarrativeQA_8598", + "flip": 0 + }, + { + "global index": "NarrativeQA_927", + "flip": 0 + }, + { + "global index": "OpenTDB_Animals_2545", + "flip": 0 + }, + { + "global index": "OpenTDB_Animals_262", + "flip": 0 + }, + { + "global index": "OpenTDB_Art_1429", + "flip": 0 + }, + { + "global index": "OpenTDB_Celebrities_1078", + "flip": 1 + }, + { + "global index": "OpenTDB_Celebrities_1456", + "flip": 1 + }, + { + "global index": "OpenTDB_Celebrities_3577", + "flip": 1 + }, + { + "global index": "OpenTDB_Celebrities_3968", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_1178", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_1807", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_1957", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_2181", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_3404", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_3727", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_4019", + "flip": 0 + }, + { + "global index": "OpenTDB_Geography_2099", + "flip": 0 + }, + { + "global index": "OpenTDB_Geography_2346", + "flip": 0 + }, + { + "global index": "OpenTDB_Geography_3880", + "flip": 0 + }, + { + "global index": "OpenTDB_Geography_792", + "flip": 0 + }, + { + "global index": "OpenTDB_History_1162", + "flip": 1 + }, + { + "global index": "OpenTDB_History_2026", + "flip": 0 + }, + { + "global index": "OpenTDB_History_3712", + "flip": 1 + }, + { + "global index": "OpenTDB_History_3902", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_1175", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_1560", + "flip": 1 + }, + { + "global index": "OpenTDB_Science & Nature_3716", + "flip": 1 + }, + { + "global index": "OpenTDB_Science & Nature_476", + "flip": 0 + }, + { + "global index": "OpenTDB_Sports_2289", + "flip": 1 + }, + { + "global index": "OpenTDB_Vehicles_1173", + "flip": 1 + }, + { + "global index": "OpenTDB_Vehicles_1419", + "flip": 0 + }, + { + "global index": "OpenTDB_Vehicles_2519", + "flip": 0 + }, + { + "global index": "OpenTDB_Vehicles_3234", + "flip": 0 + }, + { + "global index": "PubMedQA_0", + "flip": 0 + }, + { + "global index": "PubMedQA_154", + "flip": 0 + }, + { + "global index": "PubMedQA_18", + "flip": 0 + }, + { + "global index": "PubMedQA_238", + "flip": 1 + }, + { + "global index": "PubMedQA_250", + "flip": 0 + }, + { + "global index": "PubMedQA_337", + "flip": 0 + }, + { + "global index": "PubMedQA_362", + "flip": 1 + }, + { + "global index": "PubMedQA_437", + "flip": 1 + }, + { + "global index": "PubMedQA_510", + "flip": 0 + }, + { + "global index": "PubMedQA_520", + "flip": 0 + }, + { + "global index": "PubMedQA_575", + "flip": 1 + }, + { + "global index": "PubMedQA_582", + "flip": 1 + }, + { + "global index": "PubMedQA_588", + "flip": 1 + }, + { + "global index": "PubMedQA_610", + "flip": 0 + }, + { + "global index": "PubMedQA_63", + "flip": 0 + }, + { + "global index": "PubMedQA_643", + "flip": 0 + }, + { + "global index": "PubMedQA_687", + "flip": 1 + }, + { + "global index": "PubMedQA_722", + "flip": 0 + }, + { + "global index": "PubMedQA_73", + "flip": 1 + }, + { + "global index": "PubMedQA_755", + "flip": 0 + }, + { + "global index": "PubMedQA_8", + "flip": 0 + }, + { + "global index": "PubMedQA_81", + "flip": 1 + }, + { + "global index": "PubMedQA_854", + "flip": 1 + }, + { + "global index": "PubMedQA_905", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_1212", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_1702", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_828", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_865", + "flip": 1 + }, + { + "global index": "QANTA_Geography_1023", + "flip": 0 + }, + { + "global index": "QANTA_Geography_1555", + "flip": 1 + }, + { + "global index": "QANTA_Geography_304", + "flip": 0 + }, + { + "global index": "QANTA_History_1084", + "flip": 1 + }, + { + "global index": "QANTA_History_1154", + "flip": 1 + }, + { + "global index": "QANTA_History_433", + "flip": 1 + }, + { + "global index": "QANTA_History_473", + "flip": 0 + }, + { + "global index": "QANTA_History_926", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1045", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1073", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1239", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1326", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1727", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1843", + "flip": 0 + }, + { + "global index": "QANTA_Literature_386", + "flip": 0 + }, + { + "global index": "QANTA_Literature_408", + "flip": 0 + }, + { + "global index": "QANTA_Literature_475", + "flip": 0 + }, + { + "global index": "QANTA_Literature_833", + "flip": 0 + }, + { + "global index": "QANTA_Philosophy_1270", + "flip": 0 + }, + { + "global index": "QANTA_Philosophy_499", + "flip": 1 + }, + { + "global index": "QANTA_Philosophy_91", + "flip": 0 + }, + { + "global index": "QANTA_Science_1360", + "flip": 1 + }, + { + "global index": "QANTA_Science_1473", + "flip": 1 + }, + { + "global index": "QANTA_Science_308", + "flip": 0 + }, + { + "global index": "QANTA_Science_619", + "flip": 1 + }, + { + "global index": "QANTA_Social Science_1847", + "flip": 1 + }, + { + "global index": "QANTA_Social Science_2", + "flip": 1 + }, + { + "global index": "QANTA_Social Science_77", + "flip": 0 + }, + { + "global index": "SocialiQA_13810", + "flip": 0 + }, + { + "global index": "SocialiQA_22095", + "flip": 0 + }, + { + "global index": "SocialiQA_26846", + "flip": 0 + }, + { + "global index": "SocialiQA_7839", + "flip": 0 + }, + { + "global index": "SuperGLUE-CausalReasoning_4526", + "flip": 0 + }, + { + "global index": "SuperGLUE-ClozeTest_12894", + "flip": 0 + }, + { + "global index": "SuperGLUE-ClozeTest_17965", + "flip": 1 + }, + { + "global index": "SuperGLUE-ClozeTest_18766", + "flip": 1 + }, + { + "global index": "SuperGLUE-Entailment_19410", + "flip": 1 + }, + { + "global index": "SuperGLUE-Entailment_19567", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_522", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_767", + "flip": 1 + }, + { + "global index": "SuperGLUE-QA_1408", + "flip": 1 + }, + { + "global index": "SuperGLUE-QA_3137", + "flip": 1 + }, + { + "global index": "SuperGLUE-QA_4046", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_4102", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_4160", + "flip": 0 + }, + { + "global index": "SuperGLUE-RC_7725", + "flip": 0 + }, + { + "global index": "SuperGLUE-RC_7738", + "flip": 0 + }, + { + "global index": "SuperGLUE-RC_8531", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_19695", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_19738", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_20079", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_20189", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_20253", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wsc_20368", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wsc_20370", + "flip": 0 + }, + { + "global index": "WMT19-cs-en_156", + "flip": 1 + }, + { + "global index": "WMT19-cs-en_246", + "flip": 0 + }, + { + "global index": "WMT19-cs-en_568", + "flip": 1 + }, + { + "global index": "WMT19-de-en_46", + "flip": 0 + }, + { + "global index": "WMT19-de-en_715", + "flip": 0 + }, + { + "global index": "WMT19-de-en_883", + "flip": 0 + }, + { + "global index": "WMT19-fi-en_222", + "flip": 0 + }, + { + "global index": "WMT19-fi-en_610", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_116", + "flip": 1 + }, + { + "global index": "WMT19-gu-en_123", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_191", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_491", + "flip": 1 + }, + { + "global index": "WMT19-gu-en_968", + "flip": 1 + }, + { + "global index": "WMT19-kk-en_528", + "flip": 0 + }, + { + "global index": "WMT19-kk-en_826", + "flip": 0 + }, + { + "global index": "WMT19-lt-en_135", + "flip": 0 + }, + { + "global index": "WMT19-lt-en_269", + "flip": 0 + }, + { + "global index": "WMT19-lt-en_636", + "flip": 0 + }, + { + "global index": "WMT19-ru-en_222", + "flip": 0 + }, + { + "global index": "WMT19-zh-en_218", + "flip": 0 + }, + { + "global index": "WMT19-zh-en_252", + "flip": 0 + }, + { + "global index": "WMT19-zh-en_59", + "flip": 0 + } +] diff --git a/src/data/flip_labels/flip_labels_llm-router.json b/src/data/flip_labels/flip_labels_llm-router.json new file mode 100644 index 0000000..daf98f7 --- /dev/null +++ b/src/data/flip_labels/flip_labels_llm-router.json @@ -0,0 +1,1682 @@ +[ + { + "global index": "AIME_112", + "flip": 1 + }, + { + "global index": "AIME_58", + "flip": 1 + }, + { + "global index": "ArcMMLU_12", + "flip": 1 + }, + { + "global index": "ArcMMLU_123", + "flip": 0 + }, + { + "global index": "ArcMMLU_16", + "flip": 1 + }, + { + "global index": "ArcMMLU_182", + "flip": 0 + }, + { + "global index": "ArcMMLU_230", + "flip": 0 + }, + { + "global index": "ArcMMLU_293", + "flip": 1 + }, + { + "global index": "ArcMMLU_349", + "flip": 1 + }, + { + "global index": "ArcMMLU_378", + "flip": 1 + }, + { + "global index": "ArcMMLU_443", + "flip": 1 + }, + { + "global index": "ArcMMLU_496", + "flip": 1 + }, + { + "global index": "ArcMMLU_631", + "flip": 1 + }, + { + "global index": "ArcMMLU_646", + "flip": 0 + }, + { + "global index": "ArcMMLU_659", + "flip": 1 + }, + { + "global index": "ArcMMLU_676", + "flip": 1 + }, + { + "global index": "ArcMMLU_685", + "flip": 1 + }, + { + "global index": "ArcMMLU_689", + "flip": 1 + }, + { + "global index": "ArcMMLU_702", + "flip": 1 + }, + { + "global index": "ArcMMLU_713", + "flip": 1 + }, + { + "global index": "ArcMMLU_98", + "flip": 1 + }, + { + "global index": "AsDiv_1165", + "flip": 1 + }, + { + "global index": "AsDiv_1347", + "flip": 1 + }, + { + "global index": "AsDiv_472", + "flip": 1 + }, + { + "global index": "AsDiv_733", + "flip": 1 + }, + { + "global index": "ChessInstruct_0", + "flip": 0 + }, + { + "global index": "ChessInstruct_107", + "flip": 0 + }, + { + "global index": "ChessInstruct_144", + "flip": 1 + }, + { + "global index": "ChessInstruct_42", + "flip": 0 + }, + { + "global index": "ChessInstruct_58", + "flip": 0 + }, + { + "global index": "ChessInstruct_71", + "flip": 0 + }, + { + "global index": "ChessInstruct_84", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_28", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_51", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_6", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_62", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_70", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_85", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_90", + "flip": 1 + }, + { + "global index": "Ethics_deontology_0", + "flip": 0 + }, + { + "global index": "Ethics_deontology_2", + "flip": 0 + }, + { + "global index": "Ethics_deontology_31", + "flip": 0 + }, + { + "global index": "Ethics_deontology_32", + "flip": 0 + }, + { + "global index": "Ethics_deontology_56", + "flip": 0 + }, + { + "global index": "Ethics_justice_1", + "flip": 0 + }, + { + "global index": "Ethics_justice_45", + "flip": 0 + }, + { + "global index": "Ethics_justice_76", + "flip": 0 + }, + { + "global index": "Ethics_justice_84", + "flip": 0 + }, + { + "global index": "Ethics_virtue_14", + "flip": 0 + }, + { + "global index": "Ethics_virtue_26", + "flip": 0 + }, + { + "global index": "Ethics_virtue_30", + "flip": 0 + }, + { + "global index": "Ethics_virtue_48", + "flip": 0 + }, + { + "global index": "Ethics_virtue_51", + "flip": 0 + }, + { + "global index": "FinQA_149", + "flip": 1 + }, + { + "global index": "FinQA_208", + "flip": 0 + }, + { + "global index": "FinQA_56", + "flip": 1 + }, + { + "global index": "FinQA_60", + "flip": 1 + }, + { + "global index": "GSM8K_43", + "flip": 1 + }, + { + "global index": "GeoBench_1002", + "flip": 1 + }, + { + "global index": "GeoBench_1094", + "flip": 1 + }, + { + "global index": "GeoBench_1102", + "flip": 1 + }, + { + "global index": "GeoBench_1113", + "flip": 1 + }, + { + "global index": "GeoBench_124", + "flip": 1 + }, + { + "global index": "GeoBench_1243", + "flip": 1 + }, + { + "global index": "GeoBench_30", + "flip": 1 + }, + { + "global index": "GeoBench_502", + "flip": 1 + }, + { + "global index": "GeoBench_526", + "flip": 1 + }, + { + "global index": "GeoBench_591", + "flip": 1 + }, + { + "global index": "GeoBench_766", + "flip": 1 + }, + { + "global index": "GeoBench_87", + "flip": 1 + }, + { + "global index": "GeoBench_915", + "flip": 1 + }, + { + "global index": "GeoBench_944", + "flip": 1 + }, + { + "global index": "GeoBench_968", + "flip": 1 + }, + { + "global index": "GeoGraphyData_100k_27", + "flip": 1 + }, + { + "global index": "GeoGraphyData_100k_42", + "flip": 1 + }, + { + "global index": "LiveCodeBench_105", + "flip": 1 + }, + { + "global index": "LiveCodeBench_114", + "flip": 0 + }, + { + "global index": "LiveCodeBench_118", + "flip": 0 + }, + { + "global index": "LiveCodeBench_131", + "flip": 1 + }, + { + "global index": "LiveCodeBench_136", + "flip": 1 + }, + { + "global index": "LiveCodeBench_181", + "flip": 1 + }, + { + "global index": "LiveCodeBench_237", + "flip": 1 + }, + { + "global index": "LiveCodeBench_271", + "flip": 0 + }, + { + "global index": "LiveCodeBench_350", + "flip": 1 + }, + { + "global index": "LiveCodeBench_386", + "flip": 1 + }, + { + "global index": "LiveCodeBench_405", + "flip": 0 + }, + { + "global index": "LiveCodeBench_43", + "flip": 0 + }, + { + "global index": "LiveCodeBench_431", + "flip": 1 + }, + { + "global index": "LiveCodeBench_437", + "flip": 1 + }, + { + "global index": "LiveCodeBench_476", + "flip": 1 + }, + { + "global index": "LiveCodeBench_485", + "flip": 1 + }, + { + "global index": "LiveCodeBench_49", + "flip": 1 + }, + { + "global index": "LiveCodeBench_491", + "flip": 1 + }, + { + "global index": "LiveCodeBench_499", + "flip": 1 + }, + { + "global index": "MATH_108", + "flip": 1 + }, + { + "global index": "MATH_442", + "flip": 1 + }, + { + "global index": "MATH_53", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2808", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2912", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2980", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2985", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_3188", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_3215", + "flip": 0 + }, + { + "global index": "MMLUPro_biology_3225", + "flip": 1 + }, + { + "global index": "MMLUPro_business_226", + "flip": 1 + }, + { + "global index": "MMLUPro_business_294", + "flip": 1 + }, + { + "global index": "MMLUPro_business_378", + "flip": 1 + }, + { + "global index": "MMLUPro_business_430", + "flip": 1 + }, + { + "global index": "MMLUPro_business_503", + "flip": 0 + }, + { + "global index": "MMLUPro_business_507", + "flip": 1 + }, + { + "global index": "MMLUPro_business_6", + "flip": 1 + }, + { + "global index": "MMLUPro_business_784", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_3796", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_3837", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_3974", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_4067", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_4407", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9086", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9110", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9136", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9138", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9149", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9200", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9212", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9239", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9264", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9285", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9289", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9414", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9415", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9430", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9452", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9471", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9475", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_5769", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_5907", + "flip": 0 + }, + { + "global index": "MMLUPro_economics_5931", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_5965", + "flip": 0 + }, + { + "global index": "MMLUPro_economics_6114", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_6122", + "flip": 1 + }, + { + "global index": "MMLUPro_economics_6135", + "flip": 0 + }, + { + "global index": "MMLUPro_economics_6325", + "flip": 0 + }, + { + "global index": "MMLUPro_economics_6353", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10076", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10125", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10179", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10195", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10199", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10298", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10342", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10395", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10428", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10432", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10473", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10537", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10701", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10823", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10864", + "flip": 0 + }, + { + "global index": "MMLUPro_health_4885", + "flip": 1 + }, + { + "global index": "MMLUPro_health_4973", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5093", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5144", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5214", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5215", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5261", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5473", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5514", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4486", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4490", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4497", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4509", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4517", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4523", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4605", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4629", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4638", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4717", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4749", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4752", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4774", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4810", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4833", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4836", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4841", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1007", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1031", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1386", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1462", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1484", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1518", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1818", + "flip": 1 + }, + { + "global index": "MMLUPro_law_806", + "flip": 1 + }, + { + "global index": "MMLUPro_law_899", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6429", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6526", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6623", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6848", + "flip": 0 + }, + { + "global index": "MMLUPro_math_7101", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7249", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7284", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7451", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7577", + "flip": 1 + }, + { + "global index": "MMLUPro_philosophy_9510", + "flip": 0 + }, + { + "global index": "MMLUPro_philosophy_9536", + "flip": 0 + }, + { + "global index": "MMLUPro_philosophy_9663", + "flip": 1 + }, + { + "global index": "MMLUPro_philosophy_9672", + "flip": 0 + }, + { + "global index": "MMLUPro_philosophy_9943", + "flip": 0 + }, + { + "global index": "MMLUPro_physics_7773", + "flip": 1 + }, + { + "global index": "MMLUPro_physics_7887", + "flip": 1 + }, + { + "global index": "MMLUPro_physics_7893", + "flip": 1 + }, + { + "global index": "MMLUPro_physics_8888", + "flip": 1 + }, + { + "global index": "MMLUPro_physics_9017", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2005", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2186", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2329", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2406", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2420", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2450", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2457", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2524", + "flip": 1 + }, + { + "global index": "MMLU_formal_logic_121", + "flip": 1 + }, + { + "global index": "MMLU_formal_logic_16", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_32", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_63", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_7", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_70", + "flip": 0 + }, + { + "global index": "MMLU_formal_logic_85", + "flip": 0 + }, + { + "global index": "MMLU_management_4", + "flip": 0 + }, + { + "global index": "MMLU_management_41", + "flip": 1 + }, + { + "global index": "MMLU_management_77", + "flip": 0 + }, + { + "global index": "MMLU_management_91", + "flip": 1 + }, + { + "global index": "MMLU_management_93", + "flip": 1 + }, + { + "global index": "MathQA_158", + "flip": 1 + }, + { + "global index": "MathQA_1742", + "flip": 1 + }, + { + "global index": "MathQA_202", + "flip": 1 + }, + { + "global index": "MathQA_2092", + "flip": 1 + }, + { + "global index": "MathQA_2102", + "flip": 1 + }, + { + "global index": "MathQA_2851", + "flip": 1 + }, + { + "global index": "MathQA_827", + "flip": 1 + }, + { + "global index": "MathQA_84", + "flip": 1 + }, + { + "global index": "MedMCQA_1005", + "flip": 1 + }, + { + "global index": "MedMCQA_1054", + "flip": 1 + }, + { + "global index": "MedMCQA_1298", + "flip": 0 + }, + { + "global index": "MedMCQA_1309", + "flip": 0 + }, + { + "global index": "MedMCQA_1362", + "flip": 0 + }, + { + "global index": "MedMCQA_145", + "flip": 0 + }, + { + "global index": "MedMCQA_2010", + "flip": 1 + }, + { + "global index": "MedMCQA_2323", + "flip": 1 + }, + { + "global index": "MedMCQA_2366", + "flip": 1 + }, + { + "global index": "MedMCQA_2581", + "flip": 0 + }, + { + "global index": "MedMCQA_511", + "flip": 0 + }, + { + "global index": "MedMCQA_59", + "flip": 1 + }, + { + "global index": "MedMCQA_643", + "flip": 0 + }, + { + "global index": "MedMCQA_853", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_126", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_14", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_147", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_152", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_188", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_189", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_240", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_33", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_337", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_340", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_70", + "flip": 1 + }, + { + "global index": "NarrativeQA_131", + "flip": 1 + }, + { + "global index": "NarrativeQA_1683", + "flip": 1 + }, + { + "global index": "NarrativeQA_2474", + "flip": 1 + }, + { + "global index": "NarrativeQA_2820", + "flip": 1 + }, + { + "global index": "NarrativeQA_3282", + "flip": 1 + }, + { + "global index": "NarrativeQA_4102", + "flip": 1 + }, + { + "global index": "NarrativeQA_4128", + "flip": 1 + }, + { + "global index": "NarrativeQA_4347", + "flip": 1 + }, + { + "global index": "NarrativeQA_4540", + "flip": 1 + }, + { + "global index": "NarrativeQA_5022", + "flip": 1 + }, + { + "global index": "NarrativeQA_5259", + "flip": 1 + }, + { + "global index": "NarrativeQA_533", + "flip": 1 + }, + { + "global index": "NarrativeQA_5894", + "flip": 1 + }, + { + "global index": "NarrativeQA_6829", + "flip": 1 + }, + { + "global index": "NarrativeQA_7678", + "flip": 1 + }, + { + "global index": "NarrativeQA_7964", + "flip": 1 + }, + { + "global index": "NarrativeQA_8215", + "flip": 1 + }, + { + "global index": "NarrativeQA_8598", + "flip": 1 + }, + { + "global index": "NarrativeQA_927", + "flip": 1 + }, + { + "global index": "OpenTDB_Animals_2545", + "flip": 1 + }, + { + "global index": "OpenTDB_Animals_262", + "flip": 0 + }, + { + "global index": "OpenTDB_Art_1429", + "flip": 1 + }, + { + "global index": "OpenTDB_Celebrities_1078", + "flip": 0 + }, + { + "global index": "OpenTDB_Celebrities_1456", + "flip": 1 + }, + { + "global index": "OpenTDB_Celebrities_3577", + "flip": 0 + }, + { + "global index": "OpenTDB_Celebrities_3968", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_1178", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_1807", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_1957", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_2181", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_3404", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_3727", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_4019", + "flip": 0 + }, + { + "global index": "OpenTDB_Geography_2099", + "flip": 0 + }, + { + "global index": "OpenTDB_Geography_2346", + "flip": 1 + }, + { + "global index": "OpenTDB_Geography_3880", + "flip": 1 + }, + { + "global index": "OpenTDB_Geography_792", + "flip": 1 + }, + { + "global index": "OpenTDB_History_1162", + "flip": 1 + }, + { + "global index": "OpenTDB_History_2026", + "flip": 1 + }, + { + "global index": "OpenTDB_History_3712", + "flip": 1 + }, + { + "global index": "OpenTDB_History_3902", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_1175", + "flip": 1 + }, + { + "global index": "OpenTDB_Science & Nature_1560", + "flip": 1 + }, + { + "global index": "OpenTDB_Science & Nature_3716", + "flip": 1 + }, + { + "global index": "OpenTDB_Science & Nature_476", + "flip": 1 + }, + { + "global index": "OpenTDB_Sports_2289", + "flip": 0 + }, + { + "global index": "OpenTDB_Vehicles_1173", + "flip": 1 + }, + { + "global index": "OpenTDB_Vehicles_1419", + "flip": 1 + }, + { + "global index": "OpenTDB_Vehicles_2519", + "flip": 1 + }, + { + "global index": "OpenTDB_Vehicles_3234", + "flip": 1 + }, + { + "global index": "PubMedQA_0", + "flip": 0 + }, + { + "global index": "PubMedQA_154", + "flip": 1 + }, + { + "global index": "PubMedQA_18", + "flip": 1 + }, + { + "global index": "PubMedQA_238", + "flip": 1 + }, + { + "global index": "PubMedQA_250", + "flip": 1 + }, + { + "global index": "PubMedQA_337", + "flip": 0 + }, + { + "global index": "PubMedQA_362", + "flip": 1 + }, + { + "global index": "PubMedQA_437", + "flip": 1 + }, + { + "global index": "PubMedQA_510", + "flip": 1 + }, + { + "global index": "PubMedQA_520", + "flip": 1 + }, + { + "global index": "PubMedQA_575", + "flip": 1 + }, + { + "global index": "PubMedQA_582", + "flip": 0 + }, + { + "global index": "PubMedQA_588", + "flip": 1 + }, + { + "global index": "PubMedQA_610", + "flip": 1 + }, + { + "global index": "PubMedQA_63", + "flip": 0 + }, + { + "global index": "PubMedQA_643", + "flip": 1 + }, + { + "global index": "PubMedQA_687", + "flip": 1 + }, + { + "global index": "PubMedQA_722", + "flip": 1 + }, + { + "global index": "PubMedQA_73", + "flip": 1 + }, + { + "global index": "PubMedQA_755", + "flip": 1 + }, + { + "global index": "PubMedQA_8", + "flip": 1 + }, + { + "global index": "PubMedQA_81", + "flip": 1 + }, + { + "global index": "PubMedQA_854", + "flip": 1 + }, + { + "global index": "PubMedQA_905", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_1212", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_1702", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_828", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_865", + "flip": 1 + }, + { + "global index": "QANTA_Geography_1023", + "flip": 1 + }, + { + "global index": "QANTA_Geography_1555", + "flip": 1 + }, + { + "global index": "QANTA_Geography_304", + "flip": 1 + }, + { + "global index": "QANTA_History_1084", + "flip": 1 + }, + { + "global index": "QANTA_History_1154", + "flip": 1 + }, + { + "global index": "QANTA_History_433", + "flip": 1 + }, + { + "global index": "QANTA_History_473", + "flip": 1 + }, + { + "global index": "QANTA_History_926", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1045", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1073", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1239", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1326", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1727", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1843", + "flip": 1 + }, + { + "global index": "QANTA_Literature_386", + "flip": 1 + }, + { + "global index": "QANTA_Literature_408", + "flip": 1 + }, + { + "global index": "QANTA_Literature_475", + "flip": 1 + }, + { + "global index": "QANTA_Literature_833", + "flip": 1 + }, + { + "global index": "QANTA_Philosophy_1270", + "flip": 0 + }, + { + "global index": "QANTA_Philosophy_499", + "flip": 1 + }, + { + "global index": "QANTA_Philosophy_91", + "flip": 1 + }, + { + "global index": "QANTA_Science_1360", + "flip": 1 + }, + { + "global index": "QANTA_Science_1473", + "flip": 1 + }, + { + "global index": "QANTA_Science_308", + "flip": 1 + }, + { + "global index": "QANTA_Science_619", + "flip": 1 + }, + { + "global index": "QANTA_Social Science_1847", + "flip": 1 + }, + { + "global index": "QANTA_Social Science_2", + "flip": 1 + }, + { + "global index": "QANTA_Social Science_77", + "flip": 1 + }, + { + "global index": "SocialiQA_13810", + "flip": 1 + }, + { + "global index": "SocialiQA_22095", + "flip": 0 + }, + { + "global index": "SocialiQA_26846", + "flip": 1 + }, + { + "global index": "SocialiQA_7839", + "flip": 0 + }, + { + "global index": "SuperGLUE-CausalReasoning_4526", + "flip": 0 + }, + { + "global index": "SuperGLUE-ClozeTest_12894", + "flip": 1 + }, + { + "global index": "SuperGLUE-ClozeTest_17965", + "flip": 1 + }, + { + "global index": "SuperGLUE-ClozeTest_18766", + "flip": 1 + }, + { + "global index": "SuperGLUE-Entailment_19410", + "flip": 1 + }, + { + "global index": "SuperGLUE-Entailment_19567", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_522", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_767", + "flip": 1 + }, + { + "global index": "SuperGLUE-QA_1408", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_3137", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_4046", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_4102", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_4160", + "flip": 1 + }, + { + "global index": "SuperGLUE-RC_7725", + "flip": 0 + }, + { + "global index": "SuperGLUE-RC_7738", + "flip": 0 + }, + { + "global index": "SuperGLUE-RC_8531", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_19695", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_19738", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_20079", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_20189", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_20253", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wsc_20368", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wsc_20370", + "flip": 0 + }, + { + "global index": "WMT19-cs-en_156", + "flip": 0 + }, + { + "global index": "WMT19-cs-en_246", + "flip": 0 + }, + { + "global index": "WMT19-cs-en_568", + "flip": 1 + }, + { + "global index": "WMT19-de-en_46", + "flip": 0 + }, + { + "global index": "WMT19-de-en_715", + "flip": 0 + }, + { + "global index": "WMT19-de-en_883", + "flip": 0 + }, + { + "global index": "WMT19-fi-en_222", + "flip": 0 + }, + { + "global index": "WMT19-fi-en_610", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_116", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_123", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_191", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_491", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_968", + "flip": 0 + }, + { + "global index": "WMT19-kk-en_528", + "flip": 0 + }, + { + "global index": "WMT19-kk-en_826", + "flip": 0 + }, + { + "global index": "WMT19-lt-en_135", + "flip": 0 + }, + { + "global index": "WMT19-lt-en_269", + "flip": 0 + }, + { + "global index": "WMT19-lt-en_636", + "flip": 0 + }, + { + "global index": "WMT19-ru-en_222", + "flip": 0 + }, + { + "global index": "WMT19-zh-en_218", + "flip": 0 + }, + { + "global index": "WMT19-zh-en_252", + "flip": 0 + }, + { + "global index": "WMT19-zh-en_59", + "flip": 0 + } +] diff --git a/src/data/routerMetrics/category_scores.json b/src/data/routerMetrics/category_scores.json index a1a826f..62db215 100644 --- a/src/data/routerMetrics/category_scores.json +++ b/src/data/routerMetrics/category_scores.json @@ -18590,5 +18590,1869 @@ } } } + }, + "cross-router": { + "metrics": { + "easy": { + "accuracy": 96.1, + "cost": 0.0001, + "robustness": 59.1 + }, + "medium": { + "accuracy": 73.4, + "cost": 0.0003, + "robustness": 52.2 + }, + "hard": { + "accuracy": 35.0, + "cost": 0.0003, + "robustness": 66.1 + }, + "all": { + "accuracy": 75.2, + "cost": 0.0002, + "robustness": 59.0 + } + }, + "categories": { + "Arts & recreation": { + "metrics": { + "easy": { + "accuracy": 93.0, + "cost": 0.0001, + "robustness": 37.5 + }, + "medium": { + "accuracy": 62.0, + "cost": 0.0004, + "robustness": 63.6 + }, + "hard": { + "accuracy": 17.0, + "cost": 0.0004, + "robustness": 60.0 + }, + "all": { + "accuracy": 65.7, + "cost": 0.0003, + "robustness": 54.2 + } + }, + "subcategories": { + "Arts": { + "metrics": { + "easy": { + "accuracy": 98.3, + "cost": 0.0001, + "robustness": 100.0 + }, + "medium": { + "accuracy": 74.4, + "cost": 0.0002, + "robustness": 0.0 + }, + "hard": { + "accuracy": 19.6, + "cost": 0.0002, + "robustness": 50.0 + }, + "all": { + "accuracy": 63.6, + "cost": 0.0001, + "robustness": 40.0 + } + } + }, + "Music": { + "metrics": { + "easy": { + "accuracy": 96.0, + "cost": 0.0002, + "robustness": 20.0 + }, + "medium": { + "accuracy": 51.7, + "cost": 0.0007, + "robustness": 60.0 + }, + "hard": { + "accuracy": 12.0, + "cost": 0.0012, + "robustness": 0.0 + }, + "all": { + "accuracy": 65.8, + "cost": 0.0006, + "robustness": 36.4 + } + } + }, + "Sports, games and entertainment": { + "metrics": { + "easy": { + "accuracy": 88.6, + "cost": 0.0001, + "robustness": 50.0 + }, + "medium": { + "accuracy": 69.0, + "cost": 0.0002, + "robustness": 100.0 + }, + "hard": { + "accuracy": 16.7, + "cost": 0.0003, + "robustness": 100.0 + }, + "all": { + "accuracy": 66.7, + "cost": 0.0002, + "robustness": 87.5 + } + } + } + } + }, + "Computer science, information, and general works": { + "metrics": { + "easy": { + "accuracy": 96.5, + "cost": 0.0002, + "robustness": 73.1 + }, + "medium": { + "accuracy": 77.7, + "cost": 0.0004, + "robustness": 47.6 + }, + "hard": { + "accuracy": 38.8, + "cost": 0.0007, + "robustness": 80.0 + }, + "all": { + "accuracy": 80.2, + "cost": 0.0003, + "robustness": 66.1 + } + }, + "subcategories": { + "Computer science, knowledge, and systems": { + "metrics": { + "easy": { + "accuracy": 97.1, + "cost": 0.0002, + "robustness": 69.2 + }, + "medium": { + "accuracy": 81.2, + "cost": 0.0005, + "robustness": 41.2 + }, + "hard": { + "accuracy": 44.0, + "cost": 0.0008, + "robustness": 76.9 + }, + "all": { + "accuracy": 80.1, + "cost": 0.0004, + "robustness": 60.5 + } + } + }, + "Library and information sciences": { + "metrics": { + "easy": { + "accuracy": 95.6, + "cost": 0.0001, + "robustness": 76.9 + }, + "medium": { + "accuracy": 62.4, + "cost": 0.0001, + "robustness": 75.0 + }, + "hard": { + "accuracy": 8.3, + "cost": 0.0001, + "robustness": 100.0 + }, + "all": { + "accuracy": 80.6, + "cost": 0.0001, + "robustness": 78.9 + } + } + } + } + }, + "History": { + "metrics": { + "easy": { + "accuracy": 97.6, + "cost": 0.0001, + "robustness": 52.2 + }, + "medium": { + "accuracy": 80.8, + "cost": 0.0002, + "robustness": 16.7 + }, + "hard": { + "accuracy": 14.8, + "cost": 0.0002, + "robustness": 40.0 + }, + "all": { + "accuracy": 75.7, + "cost": 0.0002, + "robustness": 43.6 + } + }, + "subcategories": { + "Biography and genealogy": { + "metrics": { + "easy": { + "accuracy": 97.6, + "cost": 0.0, + "robustness": 0.0 + }, + "medium": { + "accuracy": 77.8, + "cost": 0.0001, + "robustness": 0 + }, + "hard": { + "accuracy": 0.0, + "cost": 0.0, + "robustness": 0 + }, + "all": { + "accuracy": 92.3, + "cost": 0.0001, + "robustness": 0.0 + } + } + }, + "Geography": { + "metrics": { + "easy": { + "accuracy": 97.2, + "cost": 0.0, + "robustness": 62.5 + }, + "medium": { + "accuracy": 69.0, + "cost": 0.0001, + "robustness": 100.0 + }, + "hard": { + "accuracy": 33.3, + "cost": 0.0001, + "robustness": 0 + }, + "all": { + "accuracy": 86.7, + "cost": 0.0001, + "robustness": 66.7 + } + } + }, + "History": { + "metrics": { + "easy": { + "accuracy": 97.8, + "cost": 0.0002, + "robustness": 63.6 + }, + "medium": { + "accuracy": 83.5, + "cost": 0.0003, + "robustness": 0.0 + }, + "hard": { + "accuracy": 13.2, + "cost": 0.0002, + "robustness": 40.0 + }, + "all": { + "accuracy": 70.7, + "cost": 0.0002, + "robustness": 42.3 + } + } + } + } + }, + "Language": { + "metrics": { + "easy": { + "accuracy": 91.9, + "cost": 0.0001, + "robustness": 72.7 + }, + "medium": { + "accuracy": 50.9, + "cost": 0.0002, + "robustness": 77.8 + }, + "hard": { + "accuracy": 65.0, + "cost": 0.0002, + "robustness": 72.0 + }, + "all": { + "accuracy": 69.6, + "cost": 0.0001, + "robustness": 73.3 + } + }, + "subcategories": { + "Language": { + "metrics": { + "easy": { + "accuracy": 91.9, + "cost": 0.0001, + "robustness": 72.7 + }, + "medium": { + "accuracy": 50.9, + "cost": 0.0002, + "robustness": 77.8 + }, + "hard": { + "accuracy": 65.0, + "cost": 0.0002, + "robustness": 72.0 + }, + "all": { + "accuracy": 69.6, + "cost": 0.0001, + "robustness": 73.3 + } + } + } + } + }, + "Literature": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 100.0 + }, + "medium": { + "accuracy": 76.5, + "cost": 0.0002, + "robustness": 100.0 + }, + "hard": { + "accuracy": 35.7, + "cost": 0.0002, + "robustness": 75.0 + }, + "all": { + "accuracy": 50.1, + "cost": 0.0002, + "robustness": 79.3 + } + }, + "subcategories": { + "Literature, rhetoric and criticism": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 100.0 + }, + "medium": { + "accuracy": 76.5, + "cost": 0.0002, + "robustness": 100.0 + }, + "hard": { + "accuracy": 35.7, + "cost": 0.0002, + "robustness": 75.0 + }, + "all": { + "accuracy": 50.1, + "cost": 0.0002, + "robustness": 79.3 + } + } + } + } + }, + "Philosophy and psychology": { + "metrics": { + "easy": { + "accuracy": 95.6, + "cost": 0.0001, + "robustness": 66.7 + }, + "medium": { + "accuracy": 70.1, + "cost": 0.0002, + "robustness": 90.0 + }, + "hard": { + "accuracy": 33.3, + "cost": 0.0002, + "robustness": 57.1 + }, + "all": { + "accuracy": 80.7, + "cost": 0.0002, + "robustness": 70.5 + } + }, + "subcategories": { + "Ethics": { + "metrics": { + "easy": { + "accuracy": 93.5, + "cost": 0.0001, + "robustness": 71.4 + }, + "medium": { + "accuracy": 45.9, + "cost": 0.0001, + "robustness": 100.0 + }, + "hard": { + "accuracy": 30.8, + "cost": 0.0001, + "robustness": 50.0 + }, + "all": { + "accuracy": 76.3, + "cost": 0.0001, + "robustness": 76.2 + } + } + }, + "Philosophical logic": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0002, + "robustness": 50.0 + }, + "medium": { + "accuracy": 86.8, + "cost": 0.0002, + "robustness": 100.0 + }, + "hard": { + "accuracy": 80.0, + "cost": 0.0002, + "robustness": 0 + }, + "all": { + "accuracy": 94.4, + "cost": 0.0002, + "robustness": 71.4 + } + } + }, + "Philosophy": { + "metrics": { + "easy": { + "accuracy": 94.6, + "cost": 0.0001, + "robustness": 33.3 + }, + "medium": { + "accuracy": 80.0, + "cost": 0.0002, + "robustness": 50.0 + }, + "hard": { + "accuracy": 35.6, + "cost": 0.0002, + "robustness": 66.7 + }, + "all": { + "accuracy": 68.0, + "cost": 0.0002, + "robustness": 50.0 + } + } + }, + "Psychology": { + "metrics": { + "easy": { + "accuracy": 96.7, + "cost": 0.0002, + "robustness": 83.3 + }, + "medium": { + "accuracy": 91.7, + "cost": 0.0002, + "robustness": 0 + }, + "hard": { + "accuracy": 16.7, + "cost": 0.0002, + "robustness": 50.0 + }, + "all": { + "accuracy": 87.7, + "cost": 0.0002, + "robustness": 75.0 + } + } + } + } + }, + "Science": { + "metrics": { + "easy": { + "accuracy": 98.6, + "cost": 0.0002, + "robustness": 51.3 + }, + "medium": { + "accuracy": 81.9, + "cost": 0.0003, + "robustness": 42.9 + }, + "hard": { + "accuracy": 28.9, + "cost": 0.0005, + "robustness": 66.7 + }, + "all": { + "accuracy": 83.8, + "cost": 0.0003, + "robustness": 50.7 + } + }, + "subcategories": { + "Animals (Zoology)": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 100.0 + }, + "medium": { + "accuracy": 60.0, + "cost": 0.0001, + "robustness": 0 + }, + "hard": { + "accuracy": 0, + "cost": 0, + "robustness": 0 + }, + "all": { + "accuracy": 92.0, + "cost": 0.0001, + "robustness": 100.0 + } + } + }, + "Biology": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0002, + "robustness": 0.0 + }, + "medium": { + "accuracy": 76.5, + "cost": 0.0003, + "robustness": 0 + }, + "hard": { + "accuracy": 33.3, + "cost": 0.0004, + "robustness": 0 + }, + "all": { + "accuracy": 93.0, + "cost": 0.0002, + "robustness": 0.0 + } + } + }, + "Chemistry": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0002, + "robustness": 33.3 + }, + "medium": { + "accuracy": 94.1, + "cost": 0.0004, + "robustness": 50.0 + }, + "hard": { + "accuracy": 23.1, + "cost": 0.0004, + "robustness": 0 + }, + "all": { + "accuracy": 86.7, + "cost": 0.0003, + "robustness": 40.0 + } + } + }, + "Earth sciences and geology": { + "metrics": { + "easy": { + "accuracy": 97.3, + "cost": 0.0001, + "robustness": 62.5 + }, + "medium": { + "accuracy": 66.1, + "cost": 0.0002, + "robustness": 33.3 + }, + "hard": { + "accuracy": 23.1, + "cost": 0.0002, + "robustness": 0.0 + }, + "all": { + "accuracy": 88.8, + "cost": 0.0001, + "robustness": 46.7 + } + } + }, + "Mathematics": { + "metrics": { + "easy": { + "accuracy": 98.5, + "cost": 0.0002, + "robustness": 63.6 + }, + "medium": { + "accuracy": 84.0, + "cost": 0.0003, + "robustness": 45.5 + }, + "hard": { + "accuracy": 33.7, + "cost": 0.0009, + "robustness": 100.0 + }, + "all": { + "accuracy": 82.9, + "cost": 0.0003, + "robustness": 63.0 + } + } + }, + "Physics": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0002, + "robustness": 100.0 + }, + "medium": { + "accuracy": 90.5, + "cost": 0.0003, + "robustness": 50.0 + }, + "hard": { + "accuracy": 45.5, + "cost": 0.0003, + "robustness": 0 + }, + "all": { + "accuracy": 89.7, + "cost": 0.0003, + "robustness": 80.0 + } + } + }, + "Science": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 40.0 + }, + "medium": { + "accuracy": 76.0, + "cost": 0.0001, + "robustness": 0 + }, + "hard": { + "accuracy": 20.3, + "cost": 0.0002, + "robustness": 33.3 + }, + "all": { + "accuracy": 53.5, + "cost": 0.0001, + "robustness": 37.5 + } + } + } + } + }, + "Social Science": { + "metrics": { + "easy": { + "accuracy": 96.3, + "cost": 0.0002, + "robustness": 55.6 + }, + "medium": { + "accuracy": 74.7, + "cost": 0.0005, + "robustness": 33.3 + }, + "hard": { + "accuracy": 16.1, + "cost": 0.0004, + "robustness": 42.9 + }, + "all": { + "accuracy": 71.4, + "cost": 0.0003, + "robustness": 45.9 + } + }, + "subcategories": { + "Economics": { + "metrics": { + "easy": { + "accuracy": 97.8, + "cost": 0.0002, + "robustness": 33.3 + }, + "medium": { + "accuracy": 79.8, + "cost": 0.0004, + "robustness": 33.3 + }, + "hard": { + "accuracy": 16.2, + "cost": 0.0004, + "robustness": 0.0 + }, + "all": { + "accuracy": 75.6, + "cost": 0.0003, + "robustness": 28.6 + } + } + }, + "Law": { + "metrics": { + "easy": { + "accuracy": 90.8, + "cost": 0.0004, + "robustness": 100.0 + }, + "medium": { + "accuracy": 68.3, + "cost": 0.0006, + "robustness": 25.0 + }, + "hard": { + "accuracy": 20.0, + "cost": 0.0006, + "robustness": 100.0 + }, + "all": { + "accuracy": 68.4, + "cost": 0.0005, + "robustness": 66.7 + } + } + }, + "Social problems": { + "metrics": { + "easy": { + "accuracy": 97.9, + "cost": 0.0001, + "robustness": 100.0 + }, + "medium": { + "accuracy": 22.2, + "cost": 0.0001, + "robustness": 0 + }, + "hard": { + "accuracy": 0.0, + "cost": 0.0001, + "robustness": 100.0 + }, + "all": { + "accuracy": 78.7, + "cost": 0.0001, + "robustness": 100.0 + } + } + }, + "Social sciences, sociology, and anthropology": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 0 + }, + "medium": { + "accuracy": 77.8, + "cost": 0.0002, + "robustness": 50.0 + }, + "hard": { + "accuracy": 14.7, + "cost": 0.0002, + "robustness": 0.0 + }, + "all": { + "accuracy": 40.0, + "cost": 0.0002, + "robustness": 33.3 + } + } + } + } + }, + "Technology": { + "metrics": { + "easy": { + "accuracy": 94.5, + "cost": 0.0002, + "robustness": 55.8 + }, + "medium": { + "accuracy": 72.9, + "cost": 0.0003, + "robustness": 47.6 + }, + "hard": { + "accuracy": 27.4, + "cost": 0.0004, + "robustness": 57.1 + }, + "all": { + "accuracy": 80.5, + "cost": 0.0002, + "robustness": 53.5 + } + }, + "subcategories": { + "Engineering": { + "metrics": { + "easy": { + "accuracy": 97.6, + "cost": 0.0002, + "robustness": 80.0 + }, + "medium": { + "accuracy": 86.8, + "cost": 0.0005, + "robustness": 61.5 + }, + "hard": { + "accuracy": 59.5, + "cost": 0.0007, + "robustness": 100.0 + }, + "all": { + "accuracy": 87.5, + "cost": 0.0004, + "robustness": 68.4 + } + } + }, + "Management and public relations": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 25.0 + }, + "medium": { + "accuracy": 63.6, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 0.0, + "cost": 0.0001, + "robustness": 0 + }, + "all": { + "accuracy": 91.8, + "cost": 0.0001, + "robustness": 20.0 + } + } + }, + "Medicine and health": { + "metrics": { + "easy": { + "accuracy": 93.4, + "cost": 0.0001, + "robustness": 55.9 + }, + "medium": { + "accuracy": 63.6, + "cost": 0.0002, + "robustness": 28.6 + }, + "hard": { + "accuracy": 15.9, + "cost": 0.0003, + "robustness": 50.0 + }, + "all": { + "accuracy": 77.3, + "cost": 0.0002, + "robustness": 51.1 + } + } + } + } + } + } + }, + "llm-router": { + "metrics": { + "easy": { + "accuracy": 95.6, + "cost": 0.0001, + "robustness": 30.8 + }, + "medium": { + "accuracy": 69.1, + "cost": 0.0002, + "robustness": 23.9 + }, + "hard": { + "accuracy": 27.0, + "cost": 0.0002, + "robustness": 34.9 + }, + "all": { + "accuracy": 71.8, + "cost": 0.0001, + "robustness": 30.0 + } + }, + "categories": { + "Arts & recreation": { + "metrics": { + "easy": { + "accuracy": 92.6, + "cost": 0.0001, + "robustness": 37.5 + }, + "medium": { + "accuracy": 52.2, + "cost": 0.0003, + "robustness": 27.3 + }, + "hard": { + "accuracy": 10.2, + "cost": 0.0004, + "robustness": 20.0 + }, + "all": { + "accuracy": 60.6, + "cost": 0.0002, + "robustness": 29.2 + } + }, + "subcategories": { + "Arts": { + "metrics": { + "easy": { + "accuracy": 93.2, + "cost": 0.0, + "robustness": 0.0 + }, + "medium": { + "accuracy": 64.1, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 8.9, + "cost": 0.0004, + "robustness": 0.0 + }, + "all": { + "accuracy": 55.2, + "cost": 0.0002, + "robustness": 0.0 + } + } + }, + "Music": { + "metrics": { + "easy": { + "accuracy": 97.0, + "cost": 0.0001, + "robustness": 20.0 + }, + "medium": { + "accuracy": 47.4, + "cost": 0.0005, + "robustness": 0.0 + }, + "hard": { + "accuracy": 12.0, + "cost": 0.0008, + "robustness": 0.0 + }, + "all": { + "accuracy": 64.2, + "cost": 0.0003, + "robustness": 9.1 + } + } + }, + "Sports, games and entertainment": { + "metrics": { + "easy": { + "accuracy": 89.3, + "cost": 0.0, + "robustness": 100.0 + }, + "medium": { + "accuracy": 53.0, + "cost": 0.0001, + "robustness": 75.0 + }, + "hard": { + "accuracy": 10.6, + "cost": 0.0003, + "robustness": 50.0 + }, + "all": { + "accuracy": 60.5, + "cost": 0.0001, + "robustness": 75.0 + } + } + } + } + }, + "Computer science, information, and general works": { + "metrics": { + "easy": { + "accuracy": 96.2, + "cost": 0.0001, + "robustness": 23.1 + }, + "medium": { + "accuracy": 76.6, + "cost": 0.0001, + "robustness": 14.3 + }, + "hard": { + "accuracy": 28.6, + "cost": 0.0001, + "robustness": 20.0 + }, + "all": { + "accuracy": 77.9, + "cost": 0.0001, + "robustness": 19.4 + } + }, + "subcategories": { + "Computer science, knowledge, and systems": { + "metrics": { + "easy": { + "accuracy": 96.9, + "cost": 0.0001, + "robustness": 23.1 + }, + "medium": { + "accuracy": 81.4, + "cost": 0.0001, + "robustness": 11.8 + }, + "hard": { + "accuracy": 31.1, + "cost": 0.0002, + "robustness": 23.1 + }, + "all": { + "accuracy": 77.4, + "cost": 0.0001, + "robustness": 18.6 + } + } + }, + "Library and information sciences": { + "metrics": { + "easy": { + "accuracy": 95.3, + "cost": 0.0001, + "robustness": 23.1 + }, + "medium": { + "accuracy": 55.3, + "cost": 0.0001, + "robustness": 25.0 + }, + "hard": { + "accuracy": 13.9, + "cost": 0.0001, + "robustness": 0.0 + }, + "all": { + "accuracy": 79.3, + "cost": 0.0001, + "robustness": 21.1 + } + } + } + } + }, + "History": { + "metrics": { + "easy": { + "accuracy": 96.8, + "cost": 0.0001, + "robustness": 34.8 + }, + "medium": { + "accuracy": 75.1, + "cost": 0.0002, + "robustness": 16.7 + }, + "hard": { + "accuracy": 17.4, + "cost": 0.0002, + "robustness": 10.0 + }, + "all": { + "accuracy": 74.4, + "cost": 0.0001, + "robustness": 25.6 + } + }, + "subcategories": { + "Biography and genealogy": { + "metrics": { + "easy": { + "accuracy": 97.6, + "cost": 0.0, + "robustness": 75.0 + }, + "medium": { + "accuracy": 77.8, + "cost": 0.0, + "robustness": 0 + }, + "hard": { + "accuracy": 100.0, + "cost": 0.0, + "robustness": 0 + }, + "all": { + "accuracy": 94.2, + "cost": 0.0, + "robustness": 75.0 + } + } + }, + "Geography": { + "metrics": { + "easy": { + "accuracy": 96.3, + "cost": 0.0001, + "robustness": 12.5 + }, + "medium": { + "accuracy": 55.2, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 25.0, + "cost": 0.0002, + "robustness": 0 + }, + "all": { + "accuracy": 82.7, + "cost": 0.0001, + "robustness": 11.1 + } + } + }, + "History": { + "metrics": { + "easy": { + "accuracy": 96.9, + "cost": 0.0002, + "robustness": 36.4 + }, + "medium": { + "accuracy": 79.1, + "cost": 0.0002, + "robustness": 20.0 + }, + "hard": { + "accuracy": 16.2, + "cost": 0.0002, + "robustness": 10.0 + }, + "all": { + "accuracy": 69.9, + "cost": 0.0002, + "robustness": 23.1 + } + } + } + } + }, + "Language": { + "metrics": { + "easy": { + "accuracy": 88.5, + "cost": 0.0, + "robustness": 45.5 + }, + "medium": { + "accuracy": 48.5, + "cost": 0.0001, + "robustness": 66.7 + }, + "hard": { + "accuracy": 48.3, + "cost": 0.0, + "robustness": 84.0 + }, + "all": { + "accuracy": 60.3, + "cost": 0.0, + "robustness": 71.1 + } + }, + "subcategories": { + "Language": { + "metrics": { + "easy": { + "accuracy": 88.5, + "cost": 0.0, + "robustness": 45.5 + }, + "medium": { + "accuracy": 48.5, + "cost": 0.0001, + "robustness": 66.7 + }, + "hard": { + "accuracy": 48.3, + "cost": 0.0, + "robustness": 84.0 + }, + "all": { + "accuracy": 60.3, + "cost": 0.0, + "robustness": 71.1 + } + } + } + } + }, + "Literature": { + "metrics": { + "easy": { + "accuracy": 99.1, + "cost": 0.0, + "robustness": 0.0 + }, + "medium": { + "accuracy": 72.5, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 31.9, + "cost": 0.0002, + "robustness": 0.0 + }, + "all": { + "accuracy": 46.7, + "cost": 0.0002, + "robustness": 0.0 + } + }, + "subcategories": { + "Literature, rhetoric and criticism": { + "metrics": { + "easy": { + "accuracy": 99.1, + "cost": 0.0, + "robustness": 0.0 + }, + "medium": { + "accuracy": 72.5, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 31.9, + "cost": 0.0002, + "robustness": 0.0 + }, + "all": { + "accuracy": 46.7, + "cost": 0.0002, + "robustness": 0.0 + } + } + } + } + }, + "Philosophy and psychology": { + "metrics": { + "easy": { + "accuracy": 93.6, + "cost": 0.0, + "robustness": 85.2 + }, + "medium": { + "accuracy": 66.8, + "cost": 0.0, + "robustness": 70.0 + }, + "hard": { + "accuracy": 16.0, + "cost": 0.0002, + "robustness": 57.1 + }, + "all": { + "accuracy": 76.6, + "cost": 0.0001, + "robustness": 77.3 + } + }, + "subcategories": { + "Ethics": { + "metrics": { + "easy": { + "accuracy": 90.8, + "cost": 0.0, + "robustness": 92.9 + }, + "medium": { + "accuracy": 43.5, + "cost": 0.0, + "robustness": 80.0 + }, + "hard": { + "accuracy": 0.0, + "cost": 0.0, + "robustness": 100.0 + }, + "all": { + "accuracy": 72.4, + "cost": 0.0, + "robustness": 90.5 + } + } + }, + "Philosophical logic": { + "metrics": { + "easy": { + "accuracy": 95.4, + "cost": 0.0, + "robustness": 100.0 + }, + "medium": { + "accuracy": 92.1, + "cost": 0.0, + "robustness": 66.7 + }, + "hard": { + "accuracy": 60.0, + "cost": 0.0, + "robustness": 0 + }, + "all": { + "accuracy": 92.6, + "cost": 0.0, + "robustness": 85.7 + } + } + }, + "Philosophy": { + "metrics": { + "easy": { + "accuracy": 97.3, + "cost": 0.0, + "robustness": 100.0 + }, + "medium": { + "accuracy": 72.5, + "cost": 0.0001, + "robustness": 50.0 + }, + "hard": { + "accuracy": 17.8, + "cost": 0.0003, + "robustness": 33.3 + }, + "all": { + "accuracy": 59.8, + "cost": 0.0001, + "robustness": 62.5 + } + } + }, + "Psychology": { + "metrics": { + "easy": { + "accuracy": 95.9, + "cost": 0.0001, + "robustness": 50.0 + }, + "medium": { + "accuracy": 83.3, + "cost": 0.0001, + "robustness": 0 + }, + "hard": { + "accuracy": 11.1, + "cost": 0.0001, + "robustness": 50.0 + }, + "all": { + "accuracy": 84.5, + "cost": 0.0001, + "robustness": 50.0 + } + } + } + } + }, + "Science": { + "metrics": { + "easy": { + "accuracy": 98.6, + "cost": 0.0001, + "robustness": 7.7 + }, + "medium": { + "accuracy": 79.4, + "cost": 0.0002, + "robustness": 0.0 + }, + "hard": { + "accuracy": 18.0, + "cost": 0.0005, + "robustness": 0.0 + }, + "all": { + "accuracy": 81.5, + "cost": 0.0002, + "robustness": 4.3 + } + }, + "subcategories": { + "Animals (Zoology)": { + "metrics": { + "easy": { + "accuracy": 95.0, + "cost": 0.0001, + "robustness": 50.0 + }, + "medium": { + "accuracy": 60.0, + "cost": 0.0, + "robustness": 0 + }, + "hard": { + "accuracy": 0, + "cost": 0, + "robustness": 0 + }, + "all": { + "accuracy": 88.0, + "cost": 0.0001, + "robustness": 50.0 + } + } + }, + "Biology": { + "metrics": { + "easy": { + "accuracy": 99.1, + "cost": 0.0001, + "robustness": 14.3 + }, + "medium": { + "accuracy": 76.5, + "cost": 0.0002, + "robustness": 0 + }, + "hard": { + "accuracy": 22.2, + "cost": 0.0002, + "robustness": 0 + }, + "all": { + "accuracy": 91.6, + "cost": 0.0001, + "robustness": 14.3 + } + } + }, + "Chemistry": { + "metrics": { + "easy": { + "accuracy": 95.3, + "cost": 0.0001, + "robustness": 0.0 + }, + "medium": { + "accuracy": 85.3, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 23.1, + "cost": 0.0004, + "robustness": 0 + }, + "all": { + "accuracy": 81.1, + "cost": 0.0001, + "robustness": 0.0 + } + } + }, + "Earth sciences and geology": { + "metrics": { + "easy": { + "accuracy": 99.6, + "cost": 0.0001, + "robustness": 0.0 + }, + "medium": { + "accuracy": 59.3, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 23.1, + "cost": 0.0001, + "robustness": 0.0 + }, + "all": { + "accuracy": 89.4, + "cost": 0.0001, + "robustness": 0.0 + } + } + }, + "Mathematics": { + "metrics": { + "easy": { + "accuracy": 98.1, + "cost": 0.0001, + "robustness": 9.1 + }, + "medium": { + "accuracy": 83.6, + "cost": 0.0002, + "robustness": 0.0 + }, + "hard": { + "accuracy": 22.5, + "cost": 0.0008, + "robustness": 0.0 + }, + "all": { + "accuracy": 80.9, + "cost": 0.0002, + "robustness": 3.7 + } + } + }, + "Physics": { + "metrics": { + "easy": { + "accuracy": 97.7, + "cost": 0.0002, + "robustness": 0.0 + }, + "medium": { + "accuracy": 92.9, + "cost": 0.0002, + "robustness": 0.0 + }, + "hard": { + "accuracy": 45.5, + "cost": 0.0003, + "robustness": 0 + }, + "all": { + "accuracy": 89.7, + "cost": 0.0002, + "robustness": 0.0 + } + } + }, + "Science": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 0.0 + }, + "medium": { + "accuracy": 60.0, + "cost": 0.0001, + "robustness": 0 + }, + "hard": { + "accuracy": 3.4, + "cost": 0.0002, + "robustness": 0.0 + }, + "all": { + "accuracy": 41.2, + "cost": 0.0001, + "robustness": 0.0 + } + } + } + } + }, + "Social Science": { + "metrics": { + "easy": { + "accuracy": 97.0, + "cost": 0.0001, + "robustness": 16.7 + }, + "medium": { + "accuracy": 67.6, + "cost": 0.0002, + "robustness": 41.7 + }, + "hard": { + "accuracy": 11.4, + "cost": 0.0002, + "robustness": 42.9 + }, + "all": { + "accuracy": 68.1, + "cost": 0.0001, + "robustness": 29.7 + } + }, + "subcategories": { + "Economics": { + "metrics": { + "easy": { + "accuracy": 97.3, + "cost": 0.0001, + "robustness": 8.3 + }, + "medium": { + "accuracy": 71.2, + "cost": 0.0002, + "robustness": 66.7 + }, + "hard": { + "accuracy": 8.8, + "cost": 0.0002, + "robustness": 33.3 + }, + "all": { + "accuracy": 70.7, + "cost": 0.0001, + "robustness": 28.6 + } + } + }, + "Law": { + "metrics": { + "easy": { + "accuracy": 95.4, + "cost": 0.0002, + "robustness": 33.3 + }, + "medium": { + "accuracy": 73.0, + "cost": 0.0002, + "robustness": 25.0 + }, + "hard": { + "accuracy": 23.3, + "cost": 0.0002, + "robustness": 50.0 + }, + "all": { + "accuracy": 72.8, + "cost": 0.0002, + "robustness": 33.3 + } + } + }, + "Social problems": { + "metrics": { + "easy": { + "accuracy": 97.9, + "cost": 0.0, + "robustness": 33.3 + }, + "medium": { + "accuracy": 0.0, + "cost": 0.0, + "robustness": 0 + }, + "hard": { + "accuracy": 20.0, + "cost": 0.0, + "robustness": 100.0 + }, + "all": { + "accuracy": 77.0, + "cost": 0.0, + "robustness": 50.0 + } + } + }, + "Social sciences, sociology, and anthropology": { + "metrics": { + "easy": { + "accuracy": 100.0, + "cost": 0.0001, + "robustness": 0 + }, + "medium": { + "accuracy": 50.0, + "cost": 0.0001, + "robustness": 0.0 + }, + "hard": { + "accuracy": 5.9, + "cost": 0.0001, + "robustness": 0.0 + }, + "all": { + "accuracy": 25.5, + "cost": 0.0001, + "robustness": 0.0 + } + } + } + } + }, + "Technology": { + "metrics": { + "easy": { + "accuracy": 94.8, + "cost": 0.0001, + "robustness": 23.3 + }, + "medium": { + "accuracy": 67.4, + "cost": 0.0002, + "robustness": 9.5 + }, + "hard": { + "accuracy": 21.0, + "cost": 0.0002, + "robustness": 71.4 + }, + "all": { + "accuracy": 78.3, + "cost": 0.0002, + "robustness": 23.9 + } + }, + "subcategories": { + "Engineering": { + "metrics": { + "easy": { + "accuracy": 95.2, + "cost": 0.0001, + "robustness": 0.0 + }, + "medium": { + "accuracy": 84.4, + "cost": 0.0003, + "robustness": 7.7 + }, + "hard": { + "accuracy": 47.6, + "cost": 0.0005, + "robustness": 0.0 + }, + "all": { + "accuracy": 83.9, + "cost": 0.0003, + "robustness": 5.3 + } + } + }, + "Management and public relations": { + "metrics": { + "easy": { + "accuracy": 95.0, + "cost": 0.0, + "robustness": 50.0 + }, + "medium": { + "accuracy": 63.6, + "cost": 0.0, + "robustness": 0.0 + }, + "hard": { + "accuracy": 0.0, + "cost": 0.0, + "robustness": 0 + }, + "all": { + "accuracy": 87.7, + "cost": 0.0, + "robustness": 40.0 + } + } + }, + "Medicine and health": { + "metrics": { + "easy": { + "accuracy": 94.7, + "cost": 0.0001, + "robustness": 23.5 + }, + "medium": { + "accuracy": 55.8, + "cost": 0.0001, + "robustness": 14.3 + }, + "hard": { + "accuracy": 11.5, + "cost": 0.0001, + "robustness": 83.3 + }, + "all": { + "accuracy": 75.7, + "cost": 0.0001, + "robustness": 29.8 + } + } + } + } + } + } } } diff --git a/src/data/routerMetrics/leaderboard.json b/src/data/routerMetrics/leaderboard.json index 22f7284..a62dd77 100644 --- a/src/data/routerMetrics/leaderboard.json +++ b/src/data/routerMetrics/leaderboard.json @@ -32,6 +32,17 @@ "Accuracy": 74.72, "Cost per 1k": 0.13 }, + { + "Router Name": "Cross-Router", + "Arena Score": 73.71, + "Optimal Selection Score": 40.18, + "Optimal Cost Score": 44.03, + "Optimal Acc. Score": 89.58, + "Robustness Score": 59.05, + "Latency Score": null, + "Accuracy": 75.13, + "Cost per 1k": 0.26 + }, { "Router Name": "Weave Router", "Arena Score": 72.82, @@ -87,6 +98,17 @@ "Accuracy": 71.23, "Cost per 1k": 0.06 }, + { + "Router Name": "LLM Router", + "Arena Score": 71.26, + "Optimal Selection Score": 18.01, + "Optimal Cost Score": 20.46, + "Optimal Acc. Score": 89.13, + "Robustness Score": 30.0, + "Latency Score": null, + "Accuracy": 72.05, + "Cost per 1k": 0.2 + }, { "Router Name": "azure", "Arena Score": 70.42, diff --git a/src/data/routers.json b/src/data/routers.json index 04e45db..00e382a 100644 --- a/src/data/routers.json +++ b/src/data/routers.json @@ -414,5 +414,31 @@ "qwen/qwen3-30b-a3b-instruct-2507", "mistralai/ministral-3b-2512" ] + }, + "Cross-Router": { + "name": "Cross-Router", + "affiliation": "@JiaHg", + "githubUrl": "https://github.com/JiaHg", + "type": "open-source", + "description": "Submitted by @JiaHg.", + "modelPool": [ + "google/gemini-3.1-flash-lite", + "deepseek/deepseek-v4-flash", + "deepseek/deepseek-v4-pro" + ] + }, + "LLM Router": { + "name": "LLM Router", + "affiliation": "@ypollak2", + "githubUrl": "https://github.com/ypollak2", + "type": "open-source", + "description": "Submitted by @ypollak2.", + "modelPool": [ + "qwen/qwen3-235b-a22b-2507", + "qwen/qwen3-next-80b-a3b-instruct", + "Qwen/Qwen3-Coder-Next", + "google/gemini-3.1-flash-lite", + "deepseek/deepseek-v4-flash" + ] } }