From f96d639cfd56bd5546b1d1c880a001fbe128c6d5 Mon Sep 17 00:00:00 2001 From: Shen Zhang Date: Mon, 15 Dec 2025 18:14:41 -0600 Subject: [PATCH 1/2] made columns closer --- src/pages/LeaderboardPage.css | 18 ++++++------------ 1 file changed, 6 insertions(+), 12 deletions(-) diff --git a/src/pages/LeaderboardPage.css b/src/pages/LeaderboardPage.css index bac938a..26965a8 100644 --- a/src/pages/LeaderboardPage.css +++ b/src/pages/LeaderboardPage.css @@ -22,7 +22,7 @@ } .leaderboard-full-bleed .leaderboard-container { - width:90vw; + width:85vw; margin: 0 auto; } @@ -339,8 +339,8 @@ --lb-grid: minmax(0, 0.5fr) /* Select */ minmax(0, 0.7fr) /* Rank */ - minmax(0, 2.0fr) /* Name */ - minmax(0, 1.1fr) /* Affiliation */ + minmax(0, 1.8fr) /* Name */ + minmax(0, 0.7fr) /* Affiliation */ minmax(0, 0.9fr) /* Arena */ minmax(0, 0.9fr) /* Accuracy */ minmax(0, 0.9fr) /* Cost/1k */ @@ -1407,7 +1407,7 @@ .metrics-col { display: flex; justify-content: center; - align-items: center; + align-items: left; } .metric-value { @@ -1781,13 +1781,6 @@ } -/* Preserve left alignment for name and affiliation columns */ -.leaderboard-header .name-col, -.leaderboard-header .affiliation-col, -.leaderboard-row .name-col, -.leaderboard-row .affiliation-col { - text-align: left; -} /* Fix type badge and prevent flex growth */ .type-col { @@ -1820,7 +1813,8 @@ /* put padding on cells (same for header and rows) */ .leaderboard-header > div, .leaderboard-row > div { - padding: 0.75rem 0.75rem !important; + padding: 0.75rem 0rem !important; + text-align: left; overflow: hidden; /* text-overflow: ellipsis; */ white-space: nowrap; From 89e827d3f10115a61d7a637ae595cdfd5d8bdb5f Mon Sep 17 00:00:00 2001 From: Shen Zhang Date: Mon, 22 Dec 2025 15:39:36 -0500 Subject: [PATCH 2/2] Added flip label files, updated router robustness data for router comparison --- src/data/flip_labels/flip_labels_azure.json | 1622 +++++++++++ .../flip_labels/flip_labels_graphrouter.json | 1682 ++++++++++++ .../flip_labels/flip_labels_notdiamond.json | 1626 +++++++++++ src/data/routerMetrics/category_scores.json | 2442 ++++++++--------- src/data/routerMetrics/leaderboard.json | 22 +- src/pages/LeaderboardPage.tsx | 2 +- 6 files changed, 6163 insertions(+), 1233 deletions(-) create mode 100644 src/data/flip_labels/flip_labels_azure.json create mode 100644 src/data/flip_labels/flip_labels_graphrouter.json create mode 100644 src/data/flip_labels/flip_labels_notdiamond.json diff --git a/src/data/flip_labels/flip_labels_azure.json b/src/data/flip_labels/flip_labels_azure.json new file mode 100644 index 0000000..abd8a07 --- /dev/null +++ b/src/data/flip_labels/flip_labels_azure.json @@ -0,0 +1,1622 @@ +[ + { + "global index": "AIME_112", + "flip": 0 + }, + { + "global index": "AIME_58", + "flip": 0 + }, + { + "global index": "ArcMMLU_12", + "flip": 1 + }, + { + "global index": "ArcMMLU_123", + "flip": 1 + }, + { + "global index": "ArcMMLU_16", + "flip": 1 + }, + { + "global index": "ArcMMLU_182", + "flip": 0 + }, + { + "global index": "ArcMMLU_230", + "flip": 0 + }, + { + "global index": "ArcMMLU_293", + "flip": 0 + }, + { + "global index": "ArcMMLU_349", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "ArcMMLU_702", + "flip": 1 + }, + { + "global index": "ArcMMLU_713", + "flip": 1 + }, + { + "global index": "ArcMMLU_98", + "flip": 0 + }, + { + "global index": "AsDiv_1165", + "flip": 1 + }, + { + "global index": "AsDiv_1347", + "flip": 0 + }, + { + "global index": "AsDiv_472", + "flip": 1 + }, + { + "global index": "AsDiv_733", + "flip": 1 + }, + { + "global index": "ChessInstruct_0", + "flip": 1 + }, + { + "global index": "ChessInstruct_107", + "flip": 1 + }, + { + "global index": "ChessInstruct_144", + "flip": 1 + }, + { + "global index": "ChessInstruct_42", + "flip": 1 + }, + { + "global index": "ChessInstruct_58", + "flip": 1 + }, + { + "global index": "ChessInstruct_71", + "flip": 1 + }, + { + "global index": "ChessInstruct_84", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_28", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_51", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_6", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_62", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_70", + "flip": 1 + }, + { + "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": 1 + }, + { + "global index": "Ethics_deontology_31", + "flip": 0 + }, + { + "global index": "Ethics_deontology_32", + "flip": 1 + }, + { + "global index": "Ethics_deontology_56", + "flip": 0 + }, + { + "global index": "Ethics_justice_1", + "flip": 1 + }, + { + "global index": "Ethics_justice_45", + "flip": 1 + }, + { + "global index": "Ethics_justice_76", + "flip": 1 + }, + { + "global index": "Ethics_justice_84", + "flip": 1 + }, + { + "global index": "Ethics_virtue_14", + "flip": 1 + }, + { + "global index": "Ethics_virtue_26", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "FinQA_208", + "flip": 1 + }, + { + "global index": "FinQA_56", + "flip": 0 + }, + { + "global index": "FinQA_60", + "flip": 0 + }, + { + "global index": "GSM8K_43", + "flip": 0 + }, + { + "global index": "GeoBench_1002", + "flip": 1 + }, + { + "global index": "GeoBench_1094", + "flip": 0 + }, + { + "global index": "GeoBench_1102", + "flip": 0 + }, + { + "global index": "GeoBench_1113", + "flip": 0 + }, + { + "global index": "GeoBench_124", + "flip": 0 + }, + { + "global index": "GeoBench_1243", + "flip": 1 + }, + { + "global index": "GeoBench_30", + "flip": 0 + }, + { + "global index": "GeoBench_502", + "flip": 0 + }, + { + "global index": "GeoBench_526", + "flip": 1 + }, + { + "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": 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": 0 + }, + { + "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": 0 + }, + { + "global index": "LiveCodeBench_43", + "flip": 0 + }, + { + "global index": "LiveCodeBench_431", + "flip": 0 + }, + { + "global index": "LiveCodeBench_437", + "flip": 0 + }, + { + "global index": "LiveCodeBench_476", + "flip": 1 + }, + { + "global index": "LiveCodeBench_485", + "flip": 0 + }, + { + "global index": "LiveCodeBench_49", + "flip": 1 + }, + { + "global index": "LiveCodeBench_491", + "flip": 0 + }, + { + "global index": "LiveCodeBench_499", + "flip": 1 + }, + { + "global index": "MATH_108", + "flip": 0 + }, + { + "global index": "MATH_442", + "flip": 1 + }, + { + "global index": "MATH_53", + "flip": 0 + }, + { + "global index": "MMLUPro_biology_2808", + "flip": 0 + }, + { + "global index": "MMLUPro_biology_2912", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2980", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2985", + "flip": 0 + }, + { + "global index": "MMLUPro_biology_3188", + "flip": 0 + }, + { + "global index": "MMLUPro_biology_3215", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_3225", + "flip": 0 + }, + { + "global index": "MMLUPro_business_226", + "flip": 0 + }, + { + "global index": "MMLUPro_business_294", + "flip": 1 + }, + { + "global index": "MMLUPro_business_378", + "flip": 1 + }, + { + "global index": "MMLUPro_business_430", + "flip": 0 + }, + { + "global index": "MMLUPro_business_503", + "flip": 0 + }, + { + "global index": "MMLUPro_business_507", + "flip": 0 + }, + { + "global index": "MMLUPro_business_6", + "flip": 1 + }, + { + "global index": "MMLUPro_business_784", + "flip": 1 + }, + { + "global index": "MMLUPro_chemistry_3796", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_computer science_9086", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9110", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9136", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9138", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9149", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9200", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9212", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9239", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9264", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9285", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9289", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9414", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_computer science_9475", + "flip": 0 + }, + { + "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": 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": 0 + }, + { + "global index": "MMLUPro_engineering_10125", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10179", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10195", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10199", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10298", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10342", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10395", + "flip": 0 + }, + { + "global index": "MMLUPro_engineering_10428", + "flip": 1 + }, + { + "global index": "MMLUPro_engineering_10432", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_engineering_10864", + "flip": 0 + }, + { + "global index": "MMLUPro_health_4885", + "flip": 0 + }, + { + "global index": "MMLUPro_health_4973", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5093", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5144", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5214", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5215", + "flip": 1 + }, + { + "global index": "MMLUPro_health_5261", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5473", + "flip": 0 + }, + { + "global index": "MMLUPro_health_5514", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4486", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_history_4523", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4605", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4629", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4638", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4717", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4749", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_history_4841", + "flip": 0 + }, + { + "global index": "MMLUPro_law_1007", + "flip": 0 + }, + { + "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": 1 + }, + { + "global index": "MMLUPro_law_1518", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1818", + "flip": 0 + }, + { + "global index": "MMLUPro_law_806", + "flip": 1 + }, + { + "global index": "MMLUPro_law_899", + "flip": 0 + }, + { + "global index": "MMLUPro_math_6429", + "flip": 0 + }, + { + "global index": "MMLUPro_math_6526", + "flip": 0 + }, + { + "global index": "MMLUPro_math_6623", + "flip": 1 + }, + { + "global index": "MMLUPro_math_6848", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7101", + "flip": 1 + }, + { + "global index": "MMLUPro_math_7249", + "flip": 0 + }, + { + "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": 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": 0 + }, + { + "global index": "MMLUPro_physics_7893", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_psychology_2450", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2457", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2524", + "flip": 0 + }, + { + "global index": "MathQA_158", + "flip": 1 + }, + { + "global index": "MathQA_1742", + "flip": 0 + }, + { + "global index": "MathQA_202", + "flip": 0 + }, + { + "global index": "MathQA_2092", + "flip": 1 + }, + { + "global index": "MathQA_2102", + "flip": 0 + }, + { + "global index": "MathQA_2851", + "flip": 0 + }, + { + "global index": "MathQA_827", + "flip": 0 + }, + { + "global index": "MathQA_84", + "flip": 0 + }, + { + "global index": "MedMCQA_1005", + "flip": 0 + }, + { + "global index": "MedMCQA_1054", + "flip": 1 + }, + { + "global index": "MedMCQA_1298", + "flip": 0 + }, + { + "global index": "MedMCQA_1309", + "flip": 1 + }, + { + "global index": "MedMCQA_1362", + "flip": 1 + }, + { + "global index": "MedMCQA_145", + "flip": 0 + }, + { + "global index": "MedMCQA_2010", + "flip": 0 + }, + { + "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": 1 + }, + { + "global index": "MedMCQA_853", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_126", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_14", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_147", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_152", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_188", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_189", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_240", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_33", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_337", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_340", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_70", + "flip": 1 + }, + { + "global index": "NarrativeQA_131", + "flip": 0 + }, + { + "global index": "NarrativeQA_1683", + "flip": 0 + }, + { + "global index": "NarrativeQA_2474", + "flip": 0 + }, + { + "global index": "NarrativeQA_2820", + "flip": 0 + }, + { + "global index": "NarrativeQA_3282", + "flip": 1 + }, + { + "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": 0 + }, + { + "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": 1 + }, + { + "global index": "NarrativeQA_927", + "flip": 0 + }, + { + "global index": "OpenTDB_Animals_2545", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "OpenTDB_Celebrities_3577", + "flip": 0 + }, + { + "global index": "OpenTDB_Celebrities_3968", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_1178", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_1807", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_1957", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "OpenTDB_Science & Nature_3716", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_476", + "flip": 0 + }, + { + "global index": "OpenTDB_Sports_2289", + "flip": 0 + }, + { + "global index": "OpenTDB_Vehicles_1173", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "PubMedQA_238", + "flip": 0 + }, + { + "global index": "PubMedQA_250", + "flip": 0 + }, + { + "global index": "PubMedQA_337", + "flip": 0 + }, + { + "global index": "PubMedQA_362", + "flip": 0 + }, + { + "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": 1 + }, + { + "global index": "PubMedQA_687", + "flip": 0 + }, + { + "global index": "PubMedQA_722", + "flip": 1 + }, + { + "global index": "PubMedQA_73", + "flip": 1 + }, + { + "global index": "PubMedQA_755", + "flip": 1 + }, + { + "global index": "PubMedQA_8", + "flip": 0 + }, + { + "global index": "PubMedQA_81", + "flip": 0 + }, + { + "global index": "PubMedQA_854", + "flip": 0 + }, + { + "global index": "PubMedQA_905", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_1212", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_1702", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_828", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_865", + "flip": 0 + }, + { + "global index": "QANTA_Geography_1023", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "QANTA_History_433", + "flip": 1 + }, + { + "global index": "QANTA_History_473", + "flip": 1 + }, + { + "global index": "QANTA_History_926", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1045", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1073", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1239", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1326", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1727", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "QANTA_Literature_833", + "flip": 1 + }, + { + "global index": "QANTA_Philosophy_1270", + "flip": 1 + }, + { + "global index": "QANTA_Philosophy_499", + "flip": 0 + }, + { + "global index": "QANTA_Philosophy_91", + "flip": 1 + }, + { + "global index": "QANTA_Science_1360", + "flip": 1 + }, + { + "global index": "QANTA_Science_1473", + "flip": 0 + }, + { + "global index": "QANTA_Science_308", + "flip": 1 + }, + { + "global index": "QANTA_Science_619", + "flip": 0 + }, + { + "global index": "QANTA_Social Science_1847", + "flip": 1 + }, + { + "global index": "QANTA_Social Science_2", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "SocialiQA_7839", + "flip": 1 + }, + { + "global index": "SuperGLUE-CausalReasoning_4526", + "flip": 1 + }, + { + "global index": "SuperGLUE-ClozeTest_12894", + "flip": 1 + }, + { + "global index": "SuperGLUE-ClozeTest_17965", + "flip": 0 + }, + { + "global index": "SuperGLUE-ClozeTest_18766", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_19410", + "flip": 0 + }, + { + "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": 1 + }, + { + "global index": "SuperGLUE-QA_4102", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_4160", + "flip": 0 + }, + { + "global index": "SuperGLUE-RC_7725", + "flip": 1 + }, + { + "global index": "SuperGLUE-RC_7738", + "flip": 0 + }, + { + "global index": "SuperGLUE-RC_8531", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_19695", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_19738", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_20079", + "flip": 0 + }, + { + "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": 1 + }, + { + "global index": "WMT19-cs-en_156", + "flip": 1 + }, + { + "global index": "WMT19-cs-en_246", + "flip": 1 + }, + { + "global index": "WMT19-cs-en_568", + "flip": 1 + }, + { + "global index": "WMT19-de-en_46", + "flip": 1 + }, + { + "global index": "WMT19-de-en_715", + "flip": 1 + }, + { + "global index": "WMT19-de-en_883", + "flip": 1 + }, + { + "global index": "WMT19-fi-en_222", + "flip": 1 + }, + { + "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": 1 + }, + { + "global index": "WMT19-gu-en_491", + "flip": 0 + }, + { + "global index": "WMT19-gu-en_968", + "flip": 1 + }, + { + "global index": "WMT19-kk-en_528", + "flip": 1 + }, + { + "global index": "WMT19-kk-en_826", + "flip": 1 + }, + { + "global index": "WMT19-lt-en_135", + "flip": 1 + }, + { + "global index": "WMT19-lt-en_269", + "flip": 1 + }, + { + "global index": "WMT19-lt-en_636", + "flip": 1 + }, + { + "global index": "WMT19-ru-en_222", + "flip": 1 + }, + { + "global index": "WMT19-zh-en_218", + "flip": 1 + }, + { + "global index": "WMT19-zh-en_252", + "flip": 0 + }, + { + "global index": "WMT19-zh-en_59", + "flip": 1 + } +] diff --git a/src/data/flip_labels/flip_labels_graphrouter.json b/src/data/flip_labels/flip_labels_graphrouter.json new file mode 100644 index 0000000..8da2159 --- /dev/null +++ b/src/data/flip_labels/flip_labels_graphrouter.json @@ -0,0 +1,1682 @@ +[ + { + "global_index": "AIME_112", + "flip": 0 + }, + { + "global_index": "AIME_58", + "flip": 0 + }, + { + "global_index": "ArcMMLU_12", + "flip": 0 + }, + { + "global_index": "ArcMMLU_123", + "flip": 0 + }, + { + "global_index": "ArcMMLU_16", + "flip": 0 + }, + { + "global_index": "ArcMMLU_182", + "flip": 0 + }, + { + "global_index": "ArcMMLU_230", + "flip": 0 + }, + { + "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": 0 + }, + { + "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": 1 + }, + { + "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": 1 + }, + { + "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": 0 + }, + { + "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": 0 + }, + { + "global_index": "FinQA_208", + "flip": 0 + }, + { + "global_index": "FinQA_56", + "flip": 0 + }, + { + "global_index": "FinQA_60", + "flip": 0 + }, + { + "global_index": "GSM8K_43", + "flip": 0 + }, + { + "global_index": "GeoBench_1002", + "flip": 0 + }, + { + "global_index": "GeoBench_1094", + "flip": 0 + }, + { + "global_index": "GeoBench_1102", + "flip": 0 + }, + { + "global_index": "GeoBench_1113", + "flip": 0 + }, + { + "global_index": "GeoBench_124", + "flip": 0 + }, + { + "global_index": "GeoBench_1243", + "flip": 0 + }, + { + "global_index": "GeoBench_30", + "flip": 0 + }, + { + "global_index": "GeoBench_502", + "flip": 0 + }, + { + "global_index": "GeoBench_526", + "flip": 0 + }, + { + "global_index": "GeoBench_591", + "flip": 0 + }, + { + "global_index": "GeoBench_766", + "flip": 0 + }, + { + "global_index": "GeoBench_87", + "flip": 0 + }, + { + "global_index": "GeoBench_915", + "flip": 0 + }, + { + "global_index": "GeoBench_944", + "flip": 0 + }, + { + "global_index": "GeoBench_968", + "flip": 0 + }, + { + "global_index": "GeoGraphyData_100k_27", + "flip": 0 + }, + { + "global_index": "GeoGraphyData_100k_42", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_105", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_114", + "flip": 1 + }, + { + "global_index": "LiveCodeBench_118", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_131", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_136", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_181", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_237", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_271", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_350", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_386", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_405", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_43", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_431", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_437", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_476", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_485", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_49", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_491", + "flip": 0 + }, + { + "global_index": "LiveCodeBench_499", + "flip": 1 + }, + { + "global_index": "MATH_108", + "flip": 1 + }, + { + "global_index": "MATH_442", + "flip": 0 + }, + { + "global_index": "MATH_53", + "flip": 0 + }, + { + "global_index": "MMLUPro_biology_2808", + "flip": 0 + }, + { + "global_index": "MMLUPro_biology_2912", + "flip": 0 + }, + { + "global_index": "MMLUPro_biology_2980", + "flip": 0 + }, + { + "global_index": "MMLUPro_biology_2985", + "flip": 0 + }, + { + "global_index": "MMLUPro_biology_3188", + "flip": 0 + }, + { + "global_index": "MMLUPro_biology_3215", + "flip": 0 + }, + { + "global_index": "MMLUPro_biology_3225", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_226", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_294", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_378", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_430", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_503", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_507", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_6", + "flip": 0 + }, + { + "global_index": "MMLUPro_business_784", + "flip": 0 + }, + { + "global_index": "MMLUPro_chemistry_3796", + "flip": 0 + }, + { + "global_index": "MMLUPro_chemistry_3837", + "flip": 0 + }, + { + "global_index": "MMLUPro_chemistry_3974", + "flip": 0 + }, + { + "global_index": "MMLUPro_chemistry_4067", + "flip": 1 + }, + { + "global_index": "MMLUPro_chemistry_4407", + "flip": 0 + }, + { + "global_index": "MMLUPro_computer science_9086", + "flip": 0 + }, + { + "global_index": "MMLUPro_computer science_9110", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "MMLUPro_computer science_9264", + "flip": 0 + }, + { + "global_index": "MMLUPro_computer science_9285", + "flip": 0 + }, + { + "global_index": "MMLUPro_computer science_9289", + "flip": 1 + }, + { + "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": 0 + }, + { + "global_index": "MMLUPro_computer science_9475", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_5769", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_5907", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_5931", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_5965", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_6114", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_6122", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_6135", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_6325", + "flip": 0 + }, + { + "global_index": "MMLUPro_economics_6353", + "flip": 0 + }, + { + "global_index": "MMLUPro_engineering_10076", + "flip": 0 + }, + { + "global_index": "MMLUPro_engineering_10125", + "flip": 0 + }, + { + "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": 0 + }, + { + "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": 0 + }, + { + "global_index": "MMLUPro_engineering_10864", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_4885", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_4973", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_5093", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_5144", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_5214", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_5215", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_5261", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_5473", + "flip": 0 + }, + { + "global_index": "MMLUPro_health_5514", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4486", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4490", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4497", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4509", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4517", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4523", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "MMLUPro_history_4749", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4752", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4774", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4810", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4833", + "flip": 0 + }, + { + "global_index": "MMLUPro_history_4836", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "MMLUPro_law_806", + "flip": 1 + }, + { + "global_index": "MMLUPro_law_899", + "flip": 0 + }, + { + "global_index": "MMLUPro_math_6429", + "flip": 0 + }, + { + "global_index": "MMLUPro_math_6526", + "flip": 0 + }, + { + "global_index": "MMLUPro_math_6623", + "flip": 0 + }, + { + "global_index": "MMLUPro_math_6848", + "flip": 0 + }, + { + "global_index": "MMLUPro_math_7101", + "flip": 0 + }, + { + "global_index": "MMLUPro_math_7249", + "flip": 0 + }, + { + "global_index": "MMLUPro_math_7284", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "MMLUPro_philosophy_9663", + "flip": 0 + }, + { + "global_index": "MMLUPro_philosophy_9672", + "flip": 0 + }, + { + "global_index": "MMLUPro_philosophy_9943", + "flip": 0 + }, + { + "global_index": "MMLUPro_physics_7773", + "flip": 0 + }, + { + "global_index": "MMLUPro_physics_7887", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "MMLUPro_psychology_2450", + "flip": 0 + }, + { + "global_index": "MMLUPro_psychology_2457", + "flip": 0 + }, + { + "global_index": "MMLUPro_psychology_2524", + "flip": 0 + }, + { + "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": 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": 0 + }, + { + "global_index": "MMLU_management_77", + "flip": 0 + }, + { + "global_index": "MMLU_management_91", + "flip": 0 + }, + { + "global_index": "MMLU_management_93", + "flip": 0 + }, + { + "global_index": "MathQA_158", + "flip": 0 + }, + { + "global_index": "MathQA_1742", + "flip": 0 + }, + { + "global_index": "MathQA_202", + "flip": 0 + }, + { + "global_index": "MathQA_2092", + "flip": 0 + }, + { + "global_index": "MathQA_2102", + "flip": 0 + }, + { + "global_index": "MathQA_2851", + "flip": 0 + }, + { + "global_index": "MathQA_827", + "flip": 1 + }, + { + "global_index": "MathQA_84", + "flip": 1 + }, + { + "global_index": "MedMCQA_1005", + "flip": 0 + }, + { + "global_index": "MedMCQA_1054", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "MedMCQA_2323", + "flip": 0 + }, + { + "global_index": "MedMCQA_2366", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "MusicTheoryBench_126", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_14", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_147", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_152", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_188", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_189", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_240", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_33", + "flip": 0 + }, + { + "global_index": "MusicTheoryBench_337", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "NarrativeQA_3282", + "flip": 0 + }, + { + "global_index": "NarrativeQA_4102", + "flip": 0 + }, + { + "global_index": "NarrativeQA_4128", + "flip": 0 + }, + { + "global_index": "NarrativeQA_4347", + "flip": 0 + }, + { + "global_index": "NarrativeQA_4540", + "flip": 0 + }, + { + "global_index": "NarrativeQA_5022", + "flip": 0 + }, + { + "global_index": "NarrativeQA_5259", + "flip": 0 + }, + { + "global_index": "NarrativeQA_533", + "flip": 0 + }, + { + "global_index": "NarrativeQA_5894", + "flip": 0 + }, + { + "global_index": "NarrativeQA_6829", + "flip": 0 + }, + { + "global_index": "NarrativeQA_7678", + "flip": 0 + }, + { + "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": 1 + }, + { + "global_index": "OpenTDB_Celebrities_1078", + "flip": 0 + }, + { + "global_index": "OpenTDB_Celebrities_1456", + "flip": 0 + }, + { + "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": 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": 0 + }, + { + "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": 0 + }, + { + "global_index": "OpenTDB_History_2026", + "flip": 0 + }, + { + "global_index": "OpenTDB_History_3712", + "flip": 0 + }, + { + "global_index": "OpenTDB_History_3902", + "flip": 0 + }, + { + "global_index": "OpenTDB_Science & Nature_1175", + "flip": 0 + }, + { + "global_index": "OpenTDB_Science & Nature_1560", + "flip": 0 + }, + { + "global_index": "OpenTDB_Science & Nature_3716", + "flip": 0 + }, + { + "global_index": "OpenTDB_Science & Nature_476", + "flip": 0 + }, + { + "global_index": "OpenTDB_Sports_2289", + "flip": 0 + }, + { + "global_index": "OpenTDB_Vehicles_1173", + "flip": 0 + }, + { + "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": 1 + }, + { + "global_index": "PubMedQA_238", + "flip": 0 + }, + { + "global_index": "PubMedQA_250", + "flip": 0 + }, + { + "global_index": "PubMedQA_337", + "flip": 0 + }, + { + "global_index": "PubMedQA_362", + "flip": 0 + }, + { + "global_index": "PubMedQA_437", + "flip": 0 + }, + { + "global_index": "PubMedQA_510", + "flip": 0 + }, + { + "global_index": "PubMedQA_520", + "flip": 0 + }, + { + "global_index": "PubMedQA_575", + "flip": 0 + }, + { + "global_index": "PubMedQA_582", + "flip": 0 + }, + { + "global_index": "PubMedQA_588", + "flip": 0 + }, + { + "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": 0 + }, + { + "global_index": "PubMedQA_755", + "flip": 0 + }, + { + "global_index": "PubMedQA_8", + "flip": 0 + }, + { + "global_index": "PubMedQA_81", + "flip": 0 + }, + { + "global_index": "PubMedQA_854", + "flip": 0 + }, + { + "global_index": "PubMedQA_905", + "flip": 0 + }, + { + "global_index": "QANTA_Fine Arts_1212", + "flip": 0 + }, + { + "global_index": "QANTA_Fine Arts_1702", + "flip": 0 + }, + { + "global_index": "QANTA_Fine Arts_828", + "flip": 0 + }, + { + "global_index": "QANTA_Fine Arts_865", + "flip": 0 + }, + { + "global_index": "QANTA_Geography_1023", + "flip": 0 + }, + { + "global_index": "QANTA_Geography_1555", + "flip": 0 + }, + { + "global_index": "QANTA_Geography_304", + "flip": 0 + }, + { + "global_index": "QANTA_History_1084", + "flip": 0 + }, + { + "global_index": "QANTA_History_1154", + "flip": 0 + }, + { + "global_index": "QANTA_History_433", + "flip": 0 + }, + { + "global_index": "QANTA_History_473", + "flip": 0 + }, + { + "global_index": "QANTA_History_926", + "flip": 0 + }, + { + "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": 0 + }, + { + "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": 0 + }, + { + "global_index": "QANTA_Philosophy_91", + "flip": 0 + }, + { + "global_index": "QANTA_Science_1360", + "flip": 0 + }, + { + "global_index": "QANTA_Science_1473", + "flip": 0 + }, + { + "global_index": "QANTA_Science_308", + "flip": 0 + }, + { + "global_index": "QANTA_Science_619", + "flip": 0 + }, + { + "global_index": "QANTA_Social Science_1847", + "flip": 0 + }, + { + "global_index": "QANTA_Social Science_2", + "flip": 0 + }, + { + "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": 1 + }, + { + "global_index": "SuperGLUE-CausalReasoning_4526", + "flip": 0 + }, + { + "global_index": "SuperGLUE-ClozeTest_12894", + "flip": 0 + }, + { + "global_index": "SuperGLUE-ClozeTest_17965", + "flip": 0 + }, + { + "global_index": "SuperGLUE-ClozeTest_18766", + "flip": 0 + }, + { + "global_index": "SuperGLUE-Entailment_19410", + "flip": 0 + }, + { + "global_index": "SuperGLUE-Entailment_19567", + "flip": 0 + }, + { + "global_index": "SuperGLUE-Entailment_522", + "flip": 0 + }, + { + "global_index": "SuperGLUE-Entailment_767", + "flip": 0 + }, + { + "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": 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": 0 + }, + { + "global_index": "SuperGLUE-Wic_20189", + "flip": 0 + }, + { + "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": 1 + }, + { + "global_index": "WMT19-cs-en_246", + "flip": 1 + }, + { + "global_index": "WMT19-cs-en_568", + "flip": 0 + }, + { + "global_index": "WMT19-de-en_46", + "flip": 1 + }, + { + "global_index": "WMT19-de-en_715", + "flip": 0 + }, + { + "global_index": "WMT19-de-en_883", + "flip": 0 + }, + { + "global_index": "WMT19-fi-en_222", + "flip": 1 + }, + { + "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": 1 + }, + { + "global_index": "WMT19-zh-en_218", + "flip": 0 + }, + { + "global_index": "WMT19-zh-en_252", + "flip": 0 + }, + { + "global_index": "WMT19-zh-en_59", + "flip": 1 + } +] diff --git a/src/data/flip_labels/flip_labels_notdiamond.json b/src/data/flip_labels/flip_labels_notdiamond.json new file mode 100644 index 0000000..3403e0d --- /dev/null +++ b/src/data/flip_labels/flip_labels_notdiamond.json @@ -0,0 +1,1626 @@ +[ + { + "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": 1 + }, + { + "global index": "ArcMMLU_182", + "flip": 0 + }, + { + "global index": "ArcMMLU_230", + "flip": 0 + }, + { + "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": 0 + }, + { + "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": 1 + }, + { + "global index": "ChessInstruct_107", + "flip": 0 + }, + { + "global index": "ChessInstruct_144", + "flip": 0 + }, + { + "global index": "ChessInstruct_42", + "flip": 1 + }, + { + "global index": "ChessInstruct_58", + "flip": 1 + }, + { + "global index": "ChessInstruct_71", + "flip": 1 + }, + { + "global index": "ChessInstruct_84", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_28", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_51", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_6", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_62", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_70", + "flip": 1 + }, + { + "global index": "Ethics_commonsense_85", + "flip": 0 + }, + { + "global index": "Ethics_commonsense_90", + "flip": 0 + }, + { + "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": 1 + }, + { + "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": 1 + }, + { + "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": 1 + }, + { + "global index": "FinQA_149", + "flip": 0 + }, + { + "global index": "FinQA_208", + "flip": 0 + }, + { + "global index": "FinQA_56", + "flip": 0 + }, + { + "global index": "FinQA_60", + "flip": 0 + }, + { + "global index": "GSM8K_43", + "flip": 0 + }, + { + "global index": "GeoBench_1002", + "flip": 1 + }, + { + "global index": "GeoBench_1094", + "flip": 1 + }, + { + "global index": "GeoBench_1102", + "flip": 0 + }, + { + "global index": "GeoBench_1113", + "flip": 0 + }, + { + "global index": "GeoBench_124", + "flip": 0 + }, + { + "global index": "GeoBench_1243", + "flip": 0 + }, + { + "global index": "GeoBench_30", + "flip": 1 + }, + { + "global index": "GeoBench_502", + "flip": 0 + }, + { + "global index": "GeoBench_526", + "flip": 0 + }, + { + "global index": "GeoBench_591", + "flip": 0 + }, + { + "global index": "GeoBench_766", + "flip": 0 + }, + { + "global index": "GeoBench_87", + "flip": 0 + }, + { + "global index": "GeoBench_915", + "flip": 0 + }, + { + "global index": "GeoBench_944", + "flip": 0 + }, + { + "global index": "GeoBench_968", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "LiveCodeBench_181", + "flip": 0 + }, + { + "global index": "LiveCodeBench_237", + "flip": 0 + }, + { + "global index": "LiveCodeBench_271", + "flip": 0 + }, + { + "global index": "LiveCodeBench_350", + "flip": 0 + }, + { + "global index": "LiveCodeBench_386", + "flip": 0 + }, + { + "global index": "LiveCodeBench_405", + "flip": 0 + }, + { + "global index": "LiveCodeBench_43", + "flip": 0 + }, + { + "global index": "LiveCodeBench_431", + "flip": 0 + }, + { + "global index": "LiveCodeBench_437", + "flip": 0 + }, + { + "global index": "LiveCodeBench_476", + "flip": 0 + }, + { + "global index": "LiveCodeBench_485", + "flip": 0 + }, + { + "global index": "LiveCodeBench_49", + "flip": 0 + }, + { + "global index": "LiveCodeBench_491", + "flip": 0 + }, + { + "global index": "LiveCodeBench_499", + "flip": 0 + }, + { + "global index": "MATH_108", + "flip": 0 + }, + { + "global index": "MATH_442", + "flip": 0 + }, + { + "global index": "MATH_53", + "flip": 1 + }, + { + "global index": "MMLUPro_biology_2808", + "flip": 0 + }, + { + "global index": "MMLUPro_biology_2912", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_business_226", + "flip": 1 + }, + { + "global index": "MMLUPro_business_294", + "flip": 0 + }, + { + "global index": "MMLUPro_business_378", + "flip": 1 + }, + { + "global index": "MMLUPro_business_430", + "flip": 0 + }, + { + "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": 1 + }, + { + "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": 1 + }, + { + "global index": "MMLUPro_chemistry_4407", + "flip": 1 + }, + { + "global index": "MMLUPro_computer science_9086", + "flip": 0 + }, + { + "global index": "MMLUPro_computer science_9110", + "flip": 0 + }, + { + "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": 0 + }, + { + "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": 0 + }, + { + "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": 1 + }, + { + "global index": "MMLUPro_economics_6353", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_engineering_10864", + "flip": 1 + }, + { + "global index": "MMLUPro_health_4885", + "flip": 1 + }, + { + "global index": "MMLUPro_health_4973", + "flip": 0 + }, + { + "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": 0 + }, + { + "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": 1 + }, + { + "global index": "MMLUPro_history_4509", + "flip": 0 + }, + { + "global index": "MMLUPro_history_4517", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4523", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4605", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4629", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4638", + "flip": 1 + }, + { + "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": 1 + }, + { + "global index": "MMLUPro_history_4810", + "flip": 1 + }, + { + "global index": "MMLUPro_history_4833", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MMLUPro_law_1386", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1462", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1484", + "flip": 1 + }, + { + "global index": "MMLUPro_law_1518", + "flip": 1 + }, + { + "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": 1 + }, + { + "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": 1 + }, + { + "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": 1 + }, + { + "global index": "MMLUPro_psychology_2406", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2420", + "flip": 0 + }, + { + "global index": "MMLUPro_psychology_2450", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2457", + "flip": 1 + }, + { + "global index": "MMLUPro_psychology_2524", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MathQA_2851", + "flip": 0 + }, + { + "global index": "MathQA_827", + "flip": 0 + }, + { + "global index": "MathQA_84", + "flip": 1 + }, + { + "global index": "MedMCQA_1005", + "flip": 0 + }, + { + "global index": "MedMCQA_1054", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MedMCQA_2323", + "flip": 0 + }, + { + "global index": "MedMCQA_2366", + "flip": 0 + }, + { + "global index": "MedMCQA_2581", + "flip": 1 + }, + { + "global index": "MedMCQA_511", + "flip": 1 + }, + { + "global index": "MedMCQA_59", + "flip": 0 + }, + { + "global index": "MedMCQA_643", + "flip": 0 + }, + { + "global index": "MedMCQA_853", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_126", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_14", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "MusicTheoryBench_33", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_337", + "flip": 1 + }, + { + "global index": "MusicTheoryBench_340", + "flip": 0 + }, + { + "global index": "MusicTheoryBench_70", + "flip": 1 + }, + { + "global index": "NarrativeQA_131", + "flip": 0 + }, + { + "global index": "NarrativeQA_1683", + "flip": 0 + }, + { + "global index": "NarrativeQA_2474", + "flip": 0 + }, + { + "global index": "NarrativeQA_2820", + "flip": 0 + }, + { + "global index": "NarrativeQA_3282", + "flip": 0 + }, + { + "global index": "NarrativeQA_4102", + "flip": 0 + }, + { + "global index": "NarrativeQA_4128", + "flip": 1 + }, + { + "global index": "NarrativeQA_4347", + "flip": 0 + }, + { + "global index": "NarrativeQA_4540", + "flip": 0 + }, + { + "global index": "NarrativeQA_5022", + "flip": 0 + }, + { + "global index": "NarrativeQA_5259", + "flip": 0 + }, + { + "global index": "NarrativeQA_533", + "flip": 0 + }, + { + "global index": "NarrativeQA_5894", + "flip": 0 + }, + { + "global index": "NarrativeQA_6829", + "flip": 0 + }, + { + "global index": "NarrativeQA_7678", + "flip": 1 + }, + { + "global index": "NarrativeQA_7964", + "flip": 1 + }, + { + "global index": "NarrativeQA_8215", + "flip": 0 + }, + { + "global index": "NarrativeQA_8598", + "flip": 1 + }, + { + "global index": "NarrativeQA_927", + "flip": 1 + }, + { + "global index": "OpenTDB_Animals_2545", + "flip": 0 + }, + { + "global index": "OpenTDB_Animals_262", + "flip": 1 + }, + { + "global index": "OpenTDB_Art_1429", + "flip": 0 + }, + { + "global index": "OpenTDB_Celebrities_1078", + "flip": 0 + }, + { + "global index": "OpenTDB_Celebrities_1456", + "flip": 0 + }, + { + "global index": "OpenTDB_Celebrities_3577", + "flip": 1 + }, + { + "global index": "OpenTDB_Celebrities_3968", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_1178", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_1807", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_1957", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_2181", + "flip": 1 + }, + { + "global index": "OpenTDB_General Knowledge_3404", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_3727", + "flip": 0 + }, + { + "global index": "OpenTDB_General Knowledge_4019", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "OpenTDB_History_2026", + "flip": 1 + }, + { + "global index": "OpenTDB_History_3712", + "flip": 0 + }, + { + "global index": "OpenTDB_History_3902", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_1175", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_1560", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_3716", + "flip": 0 + }, + { + "global index": "OpenTDB_Science & Nature_476", + "flip": 1 + }, + { + "global index": "OpenTDB_Sports_2289", + "flip": 1 + }, + { + "global index": "OpenTDB_Vehicles_1173", + "flip": 0 + }, + { + "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": 1 + }, + { + "global index": "PubMedQA_154", + "flip": 1 + }, + { + "global index": "PubMedQA_18", + "flip": 1 + }, + { + "global index": "PubMedQA_238", + "flip": 0 + }, + { + "global index": "PubMedQA_250", + "flip": 1 + }, + { + "global index": "PubMedQA_337", + "flip": 1 + }, + { + "global index": "PubMedQA_362", + "flip": 1 + }, + { + "global index": "PubMedQA_437", + "flip": 0 + }, + { + "global index": "PubMedQA_510", + "flip": 1 + }, + { + "global index": "PubMedQA_520", + "flip": 0 + }, + { + "global index": "PubMedQA_575", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "PubMedQA_722", + "flip": 0 + }, + { + "global index": "PubMedQA_73", + "flip": 1 + }, + { + "global index": "PubMedQA_755", + "flip": 1 + }, + { + "global index": "PubMedQA_8", + "flip": 1 + }, + { + "global index": "PubMedQA_81", + "flip": 0 + }, + { + "global index": "PubMedQA_854", + "flip": 1 + }, + { + "global index": "PubMedQA_905", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_1212", + "flip": 1 + }, + { + "global index": "QANTA_Fine Arts_1702", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_828", + "flip": 0 + }, + { + "global index": "QANTA_Fine Arts_865", + "flip": 0 + }, + { + "global index": "QANTA_Geography_1023", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "QANTA_History_473", + "flip": 0 + }, + { + "global index": "QANTA_History_926", + "flip": 0 + }, + { + "global index": "QANTA_Literature_1045", + "flip": 1 + }, + { + "global index": "QANTA_Literature_1073", + "flip": 0 + }, + { + "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": 0 + }, + { + "global index": "QANTA_Literature_408", + "flip": 0 + }, + { + "global index": "QANTA_Literature_475", + "flip": 1 + }, + { + "global index": "QANTA_Literature_833", + "flip": 0 + }, + { + "global index": "QANTA_Philosophy_1270", + "flip": 1 + }, + { + "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": 0 + }, + { + "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": 1 + }, + { + "global index": "SocialiQA_13810", + "flip": 1 + }, + { + "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": 0 + }, + { + "global index": "SuperGLUE-ClozeTest_18766", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_19410", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_19567", + "flip": 1 + }, + { + "global index": "SuperGLUE-Entailment_522", + "flip": 0 + }, + { + "global index": "SuperGLUE-Entailment_767", + "flip": 0 + }, + { + "global index": "SuperGLUE-QA_1408", + "flip": 0 + }, + { + "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": 1 + }, + { + "global index": "SuperGLUE-RC_8531", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wic_19695", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_19738", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_20079", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_20189", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wic_20253", + "flip": 1 + }, + { + "global index": "SuperGLUE-Wsc_20368", + "flip": 0 + }, + { + "global index": "SuperGLUE-Wsc_20370", + "flip": 1 + }, + { + "global index": "WMT19-cs-en_156", + "flip": 0 + }, + { + "global index": "WMT19-cs-en_246", + "flip": 0 + }, + { + "global index": "WMT19-cs-en_568", + "flip": 0 + }, + { + "global index": "WMT19-de-en_46", + "flip": 0 + }, + { + "global index": "WMT19-de-en_715", + "flip": 1 + }, + { + "global index": "WMT19-de-en_883", + "flip": 1 + }, + { + "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": 1 + }, + { + "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 d73d6b7..a0e1d63 100644 --- a/src/data/routerMetrics/category_scores.json +++ b/src/data/routerMetrics/category_scores.json @@ -4,22 +4,22 @@ "easy": { "accuracy": 95.2, "cost": 0.0018, - "robustness": 0 + "robustness": 92.9 }, "medium": { "accuracy": 58.6, "cost": 0.0021, - "robustness": 0 + "robustness": 90.3 }, "hard": { "accuracy": 21.2, "cost": 0.0026, - "robustness": 0 + "robustness": 80.7 }, "all": { "accuracy": 67.2, "cost": 0.0021, - "robustness": 0 + "robustness": 89.0 } }, "categories": { @@ -28,22 +28,22 @@ "easy": { "accuracy": 97.0, "cost": 0.0016, - "robustness": 0 + "robustness": 80.8 }, "medium": { "accuracy": 62.8, "cost": 0.0024, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 7.3, "cost": 0.003, - "robustness": 0 + "robustness": 93.3 }, "all": { "accuracy": 70.0, "cost": 0.0021, - "robustness": 0 + "robustness": 87.1 } }, "subcategories": { @@ -52,22 +52,22 @@ "easy": { "accuracy": 96.0, "cost": 0.0013, - "robustness": 0 + "robustness": 76.9 }, "medium": { "accuracy": 67.1, "cost": 0.0015, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 5.6, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 81.6, "cost": 0.0014, - "robustness": 0 + "robustness": 84.2 } } }, @@ -76,22 +76,22 @@ "easy": { "accuracy": 97.6, "cost": 0.0017, - "robustness": 0 + "robustness": 84.6 }, "medium": { "accuracy": 61.8, "cost": 0.0026, - "robustness": 0 + "robustness": 88.2 }, "hard": { "accuracy": 7.7, "cost": 0.0032, - "robustness": 0 + "robustness": 92.3 }, "all": { "accuracy": 65.4, "cost": 0.0024, - "robustness": 0 + "robustness": 88.4 } } } @@ -102,22 +102,22 @@ "easy": { "accuracy": 92.9, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 56.9, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 7.4, "cost": 0.0018, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 72.1, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 } }, "subcategories": { @@ -126,22 +126,22 @@ "easy": { "accuracy": 90.3, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 38.8, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 70.7, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 } } }, @@ -150,22 +150,22 @@ "easy": { "accuracy": 100.0, "cost": 0.0018, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 70.0, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.9, "cost": 0.002, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 56.6, "cost": 0.002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -174,7 +174,7 @@ "easy": { "accuracy": 95.9, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 75.0, @@ -184,12 +184,12 @@ "hard": { "accuracy": 5.6, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 81.8, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 } } }, @@ -198,12 +198,12 @@ "easy": { "accuracy": 90.8, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 60.5, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 20.0, @@ -213,7 +213,7 @@ "all": { "accuracy": 76.9, "cost": 0.0018, - "robustness": 0 + "robustness": 100.0 } } } @@ -224,22 +224,22 @@ "easy": { "accuracy": 94.6, "cost": 0.0019, - "robustness": 0 + "robustness": 83.3 }, "medium": { "accuracy": 60.9, "cost": 0.0024, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.1, "cost": 0.0027, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 64.0, "cost": 0.0022, - "robustness": 0 + "robustness": 91.9 } }, "subcategories": { @@ -248,22 +248,22 @@ "easy": { "accuracy": 96.2, "cost": 0.0018, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 57.7, "cost": 0.0023, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 6.2, "cost": 0.0033, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 64.6, "cost": 0.0023, - "robustness": 0 + "robustness": 85.7 } } }, @@ -272,22 +272,22 @@ "easy": { "accuracy": 89.2, "cost": 0.0024, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 69.8, "cost": 0.0028, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 13.3, "cost": 0.0025, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 67.1, "cost": 0.0026, - "robustness": 0 + "robustness": 100.0 } } }, @@ -301,17 +301,17 @@ "medium": { "accuracy": 72.2, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 5.9, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 32.7, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 } } }, @@ -320,7 +320,7 @@ "easy": { "accuracy": 95.7, "cost": 0.0015, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 33.3, @@ -330,12 +330,12 @@ "hard": { "accuracy": 20.0, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 80.3, "cost": 0.0015, - "robustness": 0 + "robustness": 100.0 } } } @@ -346,22 +346,22 @@ "easy": { "accuracy": 80.4, "cost": 0.001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 36.5, "cost": 0.0019, - "robustness": 0 + "robustness": 77.8 }, "hard": { "accuracy": 52.9, "cost": 0.0018, - "robustness": 0 + "robustness": 56.0 }, "all": { "accuracy": 57.2, "cost": 0.0016, - "robustness": 0 + "robustness": 71.1 } }, "subcategories": { @@ -370,22 +370,22 @@ "easy": { "accuracy": 80.4, "cost": 0.001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 36.5, "cost": 0.0019, - "robustness": 0 + "robustness": 77.8 }, "hard": { "accuracy": 52.9, "cost": 0.0018, - "robustness": 0 + "robustness": 56.0 }, "all": { "accuracy": 57.2, "cost": 0.0016, - "robustness": 0 + "robustness": 71.1 } } } @@ -396,22 +396,22 @@ "easy": { "accuracy": 97.4, "cost": 0.0016, - "robustness": 0 + "robustness": 84.6 }, "medium": { "accuracy": 56.2, "cost": 0.0018, - "robustness": 0 + "robustness": 85.7 }, "hard": { "accuracy": 3.6, "cost": 0.0017, - "robustness": 0 + "robustness": 55.6 }, "all": { "accuracy": 71.7, "cost": 0.0017, - "robustness": 0 + "robustness": 81.2 } }, "subcategories": { @@ -420,22 +420,22 @@ "easy": { "accuracy": 96.2, "cost": 0.0014, - "robustness": 0 + "robustness": 72.7 }, "medium": { "accuracy": 51.6, "cost": 0.0018, - "robustness": 0 + "robustness": 81.8 }, "hard": { "accuracy": 2.2, "cost": 0.0018, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 63.7, "cost": 0.0016, - "robustness": 0 + "robustness": 74.1 } } }, @@ -444,22 +444,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0015, - "robustness": 0 + "robustness": 87.5 }, "medium": { "accuracy": 54.2, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 7.7, "cost": 0.0014, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 86.1, "cost": 0.0015, - "robustness": 0 + "robustness": 86.7 } } }, @@ -468,7 +468,7 @@ "easy": { "accuracy": 99.1, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 58.8, @@ -483,7 +483,7 @@ "all": { "accuracy": 88.1, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 } } }, @@ -492,12 +492,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0017, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 70.6, "cost": 0.0019, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 0.0, @@ -507,7 +507,7 @@ "all": { "accuracy": 73.3, "cost": 0.0018, - "robustness": 0 + "robustness": 60.0 } } }, @@ -516,12 +516,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 64.3, "cost": 0.0021, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 18.2, @@ -531,7 +531,7 @@ "all": { "accuracy": 74.2, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 } } }, @@ -540,7 +540,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0015, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 100.0, @@ -555,7 +555,7 @@ "all": { "accuracy": 100.0, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 } } }, @@ -564,7 +564,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0015, - "robustness": 0 + "robustness": 80.0 }, "medium": { "accuracy": 64.0, @@ -574,12 +574,12 @@ "hard": { "accuracy": 3.4, "cost": 0.0016, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 42.1, "cost": 0.0016, - "robustness": 0 + "robustness": 75.0 } } } @@ -590,22 +590,22 @@ "easy": { "accuracy": 95.7, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 58.1, "cost": 0.0022, - "robustness": 0 + "robustness": 85.7 }, "hard": { "accuracy": 6.4, "cost": 0.0023, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 74.4, "cost": 0.0022, - "robustness": 0 + "robustness": 95.8 } }, "subcategories": { @@ -614,22 +614,22 @@ "easy": { "accuracy": 94.4, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 62.3, "cost": 0.0026, - "robustness": 0 + "robustness": 84.6 }, "hard": { "accuracy": 4.8, "cost": 0.0026, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 67.2, "cost": 0.0023, - "robustness": 0 + "robustness": 89.5 } } }, @@ -638,22 +638,22 @@ "easy": { "accuracy": 95.8, "cost": 0.0024, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 55.4, "cost": 0.002, - "robustness": 0 + "robustness": 85.7 }, "hard": { "accuracy": 7.1, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 75.8, "cost": 0.0023, - "robustness": 0 + "robustness": 97.9 } } }, @@ -662,12 +662,12 @@ "easy": { "accuracy": 98.3, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 54.5, "cost": 0.0011, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -677,7 +677,7 @@ "all": { "accuracy": 89.0, "cost": 0.0012, - "robustness": 0 + "robustness": 100.0 } } } @@ -688,22 +688,22 @@ "easy": { "accuracy": 93.3, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 52.9, "cost": 0.002, - "robustness": 0 + "robustness": 90.9 }, "hard": { "accuracy": 10.9, "cost": 0.0026, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 61.3, "cost": 0.002, - "robustness": 0 + "robustness": 91.7 } }, "subcategories": { @@ -712,22 +712,22 @@ "easy": { "accuracy": 90.7, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 70.0, "cost": 0.0024, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 13.6, "cost": 0.0037, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 67.3, "cost": 0.0025, - "robustness": 0 + "robustness": 75.0 } } }, @@ -736,22 +736,22 @@ "easy": { "accuracy": 92.9, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 28.4, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.0, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 52.9, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 } } }, @@ -760,22 +760,22 @@ "easy": { "accuracy": 100.0, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 82.1, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.9, "cost": 0.0018, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 62.3, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 } } } @@ -786,22 +786,22 @@ "easy": { "accuracy": 99.1, "cost": 0.002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 77.9, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 30.8, "cost": 0.0034, - "robustness": 0 + "robustness": 83.3 }, "all": { "accuracy": 46.4, "cost": 0.0031, - "robustness": 0 + "robustness": 86.2 } }, "subcategories": { @@ -810,22 +810,22 @@ "easy": { "accuracy": 99.1, "cost": 0.002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 77.9, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 30.8, "cost": 0.0034, - "robustness": 0 + "robustness": 83.3 }, "all": { "accuracy": 46.4, "cost": 0.0031, - "robustness": 0 + "robustness": 86.2 } } } @@ -836,22 +836,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0021, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 74.6, "cost": 0.0023, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 11.4, "cost": 0.0023, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 73.3, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 } }, "subcategories": { @@ -860,12 +860,12 @@ "easy": { "accuracy": 97.2, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 72.4, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 41.7, @@ -875,7 +875,7 @@ "all": { "accuracy": 88.0, "cost": 0.0015, - "robustness": 0 + "robustness": 100.0 } } }, @@ -884,22 +884,22 @@ "easy": { "accuracy": 97.8, "cost": 0.0024, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 74.8, "cost": 0.0025, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.8, "cost": 0.0024, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 67.1, "cost": 0.0024, - "robustness": 0 + "robustness": 100.0 } } }, @@ -908,7 +908,7 @@ "easy": { "accuracy": 95.2, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 77.8, @@ -923,7 +923,7 @@ "all": { "accuracy": 90.4, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 } } } @@ -936,22 +936,22 @@ "easy": { "accuracy": 89.8, "cost": 0.0003, - "robustness": 0 + "robustness": 94.9 }, "medium": { "accuracy": 38.2, "cost": 0.0004, - "robustness": 0 + "robustness": 94.7 }, "hard": { "accuracy": 13.9, "cost": 0.0005, - "robustness": 0 + "robustness": 92.7 }, "all": { "accuracy": 57.0, "cost": 0.0003, - "robustness": 0 + "robustness": 94.3 } }, "categories": { @@ -960,22 +960,22 @@ "easy": { "accuracy": 93.7, "cost": 0.0002, - "robustness": 0 + "robustness": 92.3 }, "medium": { "accuracy": 41.6, "cost": 0.0005, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 2.4, "cost": 0.0007, - "robustness": 0 + "robustness": 93.3 }, "all": { "accuracy": 60.5, "cost": 0.0004, - "robustness": 0 + "robustness": 95.2 } }, "subcategories": { @@ -984,22 +984,22 @@ "easy": { "accuracy": 96.7, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 32.9, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 74.2, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1008,22 +1008,22 @@ "easy": { "accuracy": 91.6, "cost": 0.0003, - "robustness": 0 + "robustness": 84.6 }, "medium": { "accuracy": 43.5, "cost": 0.0005, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 2.9, "cost": 0.0008, - "robustness": 0 + "robustness": 92.3 }, "all": { "accuracy": 55.1, "cost": 0.0005, - "robustness": 0 + "robustness": 93.0 } } } @@ -1034,22 +1034,22 @@ "easy": { "accuracy": 90.0, "cost": 0.0002, - "robustness": 0 + "robustness": 96.3 }, "medium": { "accuracy": 35.5, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 2.5, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 63.4, "cost": 0.0002, - "robustness": 0 + "robustness": 97.7 } }, "subcategories": { @@ -1058,22 +1058,22 @@ "easy": { "accuracy": 91.4, "cost": 0.0002, - "robustness": 0 + "robustness": 92.9 }, "medium": { "accuracy": 29.4, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 68.6, "cost": 0.0002, - "robustness": 0 + "robustness": 95.2 } } }, @@ -1082,22 +1082,22 @@ "easy": { "accuracy": 89.2, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 40.0, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 2.2, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 41.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1106,7 +1106,7 @@ "easy": { "accuracy": 87.6, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 41.7, @@ -1116,12 +1116,12 @@ "hard": { "accuracy": 5.6, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 67.9, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1130,12 +1130,12 @@ "easy": { "accuracy": 90.8, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 36.8, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -1145,7 +1145,7 @@ "all": { "accuracy": 67.6, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } } @@ -1156,22 +1156,22 @@ "easy": { "accuracy": 85.9, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 32.8, "cost": 0.0004, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 2.7, "cost": 0.0005, - "robustness": 0 + "robustness": 85.7 }, "all": { "accuracy": 49.0, "cost": 0.0004, - "robustness": 0 + "robustness": 91.9 } }, "subcategories": { @@ -1180,22 +1180,22 @@ "easy": { "accuracy": 86.9, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 34.4, "cost": 0.0005, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 2.5, "cost": 0.0007, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 50.9, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1204,22 +1204,22 @@ "easy": { "accuracy": 84.6, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 30.2, "cost": 0.0004, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 3.3, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 47.5, "cost": 0.0003, - "robustness": 0 + "robustness": 77.8 } } }, @@ -1233,17 +1233,17 @@ "medium": { "accuracy": 38.9, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 2.9, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 20.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1252,7 +1252,7 @@ "easy": { "accuracy": 83.0, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 11.1, @@ -1262,12 +1262,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 65.6, "cost": 0.0001, - "robustness": 0 + "robustness": 75.0 } } } @@ -1278,22 +1278,22 @@ "easy": { "accuracy": 67.0, "cost": 0.0002, - "robustness": 0 + "robustness": 90.9 }, "medium": { "accuracy": 25.1, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 28.8, "cost": 0.0001, - "robustness": 0 + "robustness": 76.0 }, "all": { "accuracy": 39.3, "cost": 0.0002, - "robustness": 0 + "robustness": 84.4 } }, "subcategories": { @@ -1302,22 +1302,22 @@ "easy": { "accuracy": 67.0, "cost": 0.0002, - "robustness": 0 + "robustness": 90.9 }, "medium": { "accuracy": 25.1, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 28.8, "cost": 0.0001, - "robustness": 0 + "robustness": 76.0 }, "all": { "accuracy": 39.3, "cost": 0.0002, - "robustness": 0 + "robustness": 84.4 } } } @@ -1328,22 +1328,22 @@ "easy": { "accuracy": 86.4, "cost": 0.0002, - "robustness": 0 + "robustness": 92.3 }, "medium": { "accuracy": 36.6, "cost": 0.0003, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 3.1, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 59.5, "cost": 0.0003, - "robustness": 0 + "robustness": 92.8 } }, "subcategories": { @@ -1352,22 +1352,22 @@ "easy": { "accuracy": 67.9, "cost": 0.0002, - "robustness": 0 + "robustness": 72.7 }, "medium": { "accuracy": 33.2, "cost": 0.0004, - "robustness": 0 + "robustness": 90.9 }, "hard": { "accuracy": 2.2, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 43.8, "cost": 0.0003, - "robustness": 0 + "robustness": 85.2 } } }, @@ -1376,22 +1376,22 @@ "easy": { "accuracy": 99.2, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 64.4, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 89.1, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1400,7 +1400,7 @@ "easy": { "accuracy": 94.9, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 17.6, @@ -1415,7 +1415,7 @@ "all": { "accuracy": 80.4, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1424,12 +1424,12 @@ "easy": { "accuracy": 76.7, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 26.5, "cost": 0.0003, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 7.7, @@ -1439,7 +1439,7 @@ "all": { "accuracy": 47.8, "cost": 0.0003, - "robustness": 0 + "robustness": 80.0 } } }, @@ -1448,12 +1448,12 @@ "easy": { "accuracy": 95.5, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 33.3, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 9.1, @@ -1463,7 +1463,7 @@ "all": { "accuracy": 58.8, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1472,7 +1472,7 @@ "easy": { "accuracy": 95.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 40.0, @@ -1487,7 +1487,7 @@ "all": { "accuracy": 84.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1496,7 +1496,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 36.0, @@ -1506,12 +1506,12 @@ "hard": { "accuracy": 1.7, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 35.1, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } } @@ -1522,22 +1522,22 @@ "easy": { "accuracy": 92.8, "cost": 0.0003, - "robustness": 0 + "robustness": 95.3 }, "medium": { "accuracy": 40.5, "cost": 0.0003, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 1.9, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 66.9, "cost": 0.0003, - "robustness": 0 + "robustness": 94.4 } }, "subcategories": { @@ -1546,22 +1546,22 @@ "easy": { "accuracy": 86.5, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 37.7, "cost": 0.0003, - "robustness": 0 + "robustness": 84.6 }, "hard": { "accuracy": 2.4, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 51.6, "cost": 0.0003, - "robustness": 0 + "robustness": 89.5 } } }, @@ -1570,22 +1570,22 @@ "easy": { "accuracy": 94.2, "cost": 0.0003, - "robustness": 0 + "robustness": 94.1 }, "medium": { "accuracy": 41.7, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 1.8, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 70.9, "cost": 0.0003, - "robustness": 0 + "robustness": 95.7 } } }, @@ -1594,12 +1594,12 @@ "easy": { "accuracy": 91.7, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 54.5, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -1609,7 +1609,7 @@ "all": { "accuracy": 83.6, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } } @@ -1620,22 +1620,22 @@ "easy": { "accuracy": 94.0, "cost": 0.0003, - "robustness": 0 + "robustness": 87.5 }, "medium": { "accuracy": 35.7, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.7, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 53.1, "cost": 0.0004, - "robustness": 0 + "robustness": 95.8 } }, "subcategories": { @@ -1644,22 +1644,22 @@ "easy": { "accuracy": 92.1, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 39.0, "cost": 0.0007, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 1.5, "cost": 0.0008, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 55.2, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1668,22 +1668,22 @@ "easy": { "accuracy": 94.9, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 29.3, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 53.3, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1692,22 +1692,22 @@ "easy": { "accuracy": 96.6, "cost": 0.0002, - "robustness": 0 + "robustness": 0.0 }, "medium": { "accuracy": 46.2, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 48.7, "cost": 0.0003, - "robustness": 0 + "robustness": 80.0 } } } @@ -1718,22 +1718,22 @@ "easy": { "accuracy": 98.2, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 51.0, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 29.3, "cost": 0.0007, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 42.5, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 } }, "subcategories": { @@ -1742,22 +1742,22 @@ "easy": { "accuracy": 98.2, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 51.0, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 29.3, "cost": 0.0007, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 42.5, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 } } } @@ -1768,22 +1768,22 @@ "easy": { "accuracy": 92.5, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 49.7, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 3.4, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 62.7, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 } }, "subcategories": { @@ -1792,12 +1792,12 @@ "easy": { "accuracy": 85.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 58.6, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -1807,7 +1807,7 @@ "all": { "accuracy": 73.3, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1816,22 +1816,22 @@ "easy": { "accuracy": 95.5, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 48.9, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 3.7, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 57.4, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 } } }, @@ -1840,7 +1840,7 @@ "easy": { "accuracy": 95.2, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 33.3, @@ -1855,7 +1855,7 @@ "all": { "accuracy": 82.7, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 } } } @@ -1868,22 +1868,22 @@ "easy": { "accuracy": 92.9, "cost": 0.0054, - "robustness": 0 + "robustness": 56.9 }, "medium": { "accuracy": 64.0, "cost": 0.008, - "robustness": 0 + "robustness": 45.9 }, "hard": { "accuracy": 24.4, "cost": 0.0186, - "robustness": 0 + "robustness": 64.2 }, "all": { "accuracy": 68.0, "cost": 0.0093, - "robustness": 0 + "robustness": 55.9 } }, "categories": { @@ -1892,22 +1892,22 @@ "easy": { "accuracy": 92.4, "cost": 0.005, - "robustness": 0 + "robustness": 51.3 }, "medium": { "accuracy": 61.6, "cost": 0.0062, - "robustness": 0 + "robustness": 30.0 }, "hard": { "accuracy": 17.4, "cost": 0.0061, - "robustness": 0 + "robustness": 57.1 }, "all": { "accuracy": 74.2, "cost": 0.0055, - "robustness": 0 + "robustness": 45.5 } }, "subcategories": { @@ -1916,22 +1916,22 @@ "easy": { "accuracy": 91.7, "cost": 0.005, - "robustness": 0 + "robustness": 47.1 }, "medium": { "accuracy": 52.5, "cost": 0.0053, - "robustness": 0 + "robustness": 71.4 }, "hard": { "accuracy": 12.4, "cost": 0.0051, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 73.1, "cost": 0.0051, - "robustness": 0 + "robustness": 53.2 } } }, @@ -1940,22 +1940,22 @@ "easy": { "accuracy": 96.0, "cost": 0.0048, - "robustness": 0 + "robustness": 80.0 }, "medium": { "accuracy": 74.9, "cost": 0.0075, - "robustness": 0 + "robustness": 7.7 }, "hard": { "accuracy": 31.0, "cost": 0.0089, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 77.3, "cost": 0.0066, - "robustness": 0 + "robustness": 26.3 } } } @@ -1966,22 +1966,22 @@ "easy": { "accuracy": 89.8, "cost": 0.0044, - "robustness": 0 + "robustness": 60.9 }, "medium": { "accuracy": 54.3, "cost": 0.005, - "robustness": 0 + "robustness": 71.4 }, "hard": { "accuracy": 14.5, "cost": 0.0049, - "robustness": 0 + "robustness": 28.6 }, "all": { "accuracy": 69.8, "cost": 0.0047, - "robustness": 0 + "robustness": 56.8 } }, "subcategories": { @@ -1990,22 +1990,22 @@ "easy": { "accuracy": 89.2, "cost": 0.0047, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 65.0, "cost": 0.0047, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 20.0, "cost": 0.0046, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 55.7, "cost": 0.0047, - "robustness": 0 + "robustness": 37.5 } } }, @@ -2014,7 +2014,7 @@ "easy": { "accuracy": 95.9, "cost": 0.0047, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 81.2, @@ -2024,12 +2024,12 @@ "hard": { "accuracy": 5.6, "cost": 0.0054, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 83.4, "cost": 0.0048, - "robustness": 0 + "robustness": 62.5 } } }, @@ -2038,22 +2038,22 @@ "easy": { "accuracy": 85.9, "cost": 0.0042, - "robustness": 0 + "robustness": 64.3 }, "medium": { "accuracy": 34.1, "cost": 0.0052, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 7.7, "cost": 0.005, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 66.8, "cost": 0.0046, - "robustness": 0 + "robustness": 61.9 } } } @@ -2064,22 +2064,22 @@ "easy": { "accuracy": 89.9, "cost": 0.005, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 60.9, "cost": 0.0063, - "robustness": 0 + "robustness": 25.0 }, "hard": { "accuracy": 7.4, "cost": 0.0074, - "robustness": 0 + "robustness": 57.1 }, "all": { "accuracy": 61.9, "cost": 0.006, - "robustness": 0 + "robustness": 43.2 } }, "subcategories": { @@ -2093,17 +2093,17 @@ "medium": { "accuracy": 66.7, "cost": 0.0045, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 5.9, "cost": 0.0034, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 30.9, "cost": 0.0037, - "robustness": 0 + "robustness": 0.0 } } }, @@ -2112,22 +2112,22 @@ "easy": { "accuracy": 94.0, "cost": 0.0053, - "robustness": 0 + "robustness": 58.3 }, "medium": { "accuracy": 68.7, "cost": 0.0068, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 7.5, "cost": 0.01, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 68.1, "cost": 0.0068, - "robustness": 0 + "robustness": 57.1 } } }, @@ -2136,7 +2136,7 @@ "easy": { "accuracy": 91.5, "cost": 0.0041, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 22.2, @@ -2146,12 +2146,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0037, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 73.8, "cost": 0.0042, - "robustness": 0 + "robustness": 75.0 } } }, @@ -2160,22 +2160,22 @@ "easy": { "accuracy": 76.9, "cost": 0.0049, - "robustness": 0 + "robustness": 0.0 }, "medium": { "accuracy": 44.4, "cost": 0.0058, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 10.0, "cost": 0.0054, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 51.3, "cost": 0.0054, - "robustness": 0 + "robustness": 11.1 } } } @@ -2186,22 +2186,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0049, - "robustness": 0 + "robustness": 51.3 }, "medium": { "accuracy": 74.1, "cost": 0.0074, - "robustness": 0 + "robustness": 47.6 }, "hard": { "accuracy": 12.9, "cost": 0.011, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 78.4, "cost": 0.0065, - "robustness": 0 + "robustness": 52.2 } }, "subcategories": { @@ -2210,7 +2210,7 @@ "easy": { "accuracy": 93.3, "cost": 0.0034, - "robustness": 0 + "robustness": 80.0 }, "medium": { "accuracy": 60.0, @@ -2220,12 +2220,12 @@ "hard": { "accuracy": 6.8, "cost": 0.005, - "robustness": 0 + "robustness": 33.3 }, "all": { "accuracy": 41.2, "cost": 0.0046, - "robustness": 0 + "robustness": 62.5 } } }, @@ -2234,22 +2234,22 @@ "easy": { "accuracy": 98.5, "cost": 0.0052, - "robustness": 0 + "robustness": 36.4 }, "medium": { "accuracy": 78.4, "cost": 0.009, - "robustness": 0 + "robustness": 54.5 }, "hard": { "accuracy": 13.5, "cost": 0.018, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 77.5, "cost": 0.0087, - "robustness": 0 + "robustness": 51.9 } } }, @@ -2258,7 +2258,7 @@ "easy": { "accuracy": 98.3, "cost": 0.0054, - "robustness": 0 + "robustness": 42.9 }, "medium": { "accuracy": 47.1, @@ -2273,7 +2273,7 @@ "all": { "accuracy": 87.4, "cost": 0.0054, - "robustness": 0 + "robustness": 42.9 } } }, @@ -2282,12 +2282,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0045, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 82.4, "cost": 0.0051, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 15.4, @@ -2297,7 +2297,7 @@ "all": { "accuracy": 80.0, "cost": 0.0048, - "robustness": 0 + "robustness": 20.0 } } }, @@ -2306,12 +2306,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0045, - "robustness": 0 + "robustness": 0.0 }, "medium": { "accuracy": 69.0, "cost": 0.0057, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 27.3, @@ -2321,7 +2321,7 @@ "all": { "accuracy": 77.3, "cost": 0.0051, - "robustness": 0 + "robustness": 0.0 } } }, @@ -2330,7 +2330,7 @@ "easy": { "accuracy": 90.0, "cost": 0.0033, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 60.0, @@ -2345,7 +2345,7 @@ "all": { "accuracy": 84.0, "cost": 0.0032, - "robustness": 0 + "robustness": 50.0 } } }, @@ -2354,22 +2354,22 @@ "easy": { "accuracy": 96.5, "cost": 0.0049, - "robustness": 0 + "robustness": 87.5 }, "medium": { "accuracy": 69.5, "cost": 0.0052, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 15.4, "cost": 0.005, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 88.5, "cost": 0.005, - "robustness": 0 + "robustness": 80.0 } } } @@ -2380,22 +2380,22 @@ "easy": { "accuracy": 88.9, "cost": 0.0047, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 59.2, "cost": 0.0104, - "robustness": 0 + "robustness": 45.5 }, "hard": { "accuracy": 16.3, "cost": 0.0485, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 62.9, "cost": 0.016, - "robustness": 0 + "robustness": 41.7 } }, "subcategories": { @@ -2404,22 +2404,22 @@ "easy": { "accuracy": 93.2, "cost": 0.0036, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 76.9, "cost": 0.0041, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.9, "cost": 0.0043, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 58.4, "cost": 0.004, - "robustness": 0 + "robustness": 80.0 } } }, @@ -2428,22 +2428,22 @@ "easy": { "accuracy": 85.0, "cost": 0.005, - "robustness": 0 + "robustness": 0.0 }, "medium": { "accuracy": 74.0, "cost": 0.0178, - "robustness": 0 + "robustness": 25.0 }, "hard": { "accuracy": 18.2, "cost": 0.1022, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 67.0, "cost": 0.0301, - "robustness": 0 + "robustness": 37.5 } } }, @@ -2452,22 +2452,22 @@ "easy": { "accuracy": 91.9, "cost": 0.005, - "robustness": 0 + "robustness": 20.0 }, "medium": { "accuracy": 40.5, "cost": 0.0062, - "robustness": 0 + "robustness": 40.0 }, "hard": { "accuracy": 28.0, "cost": 0.0057, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 60.4, "cost": 0.0057, - "robustness": 0 + "robustness": 27.3 } } } @@ -2478,22 +2478,22 @@ "easy": { "accuracy": 96.7, "cost": 0.0053, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 57.4, "cost": 0.0041, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 24.6, "cost": 0.0092, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 39.4, "cost": 0.0081, - "robustness": 0 + "robustness": 62.1 } }, "subcategories": { @@ -2502,22 +2502,22 @@ "easy": { "accuracy": 96.7, "cost": 0.0053, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 57.4, "cost": 0.0041, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 24.6, "cost": 0.0092, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 39.4, "cost": 0.0081, - "robustness": 0 + "robustness": 62.1 } } } @@ -2528,22 +2528,22 @@ "easy": { "accuracy": 91.7, "cost": 0.0043, - "robustness": 0 + "robustness": 47.6 }, "medium": { "accuracy": 48.2, "cost": 0.0044, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 7.4, "cost": 0.0037, - "robustness": 0 + "robustness": 20.0 }, "all": { "accuracy": 61.4, "cost": 0.0042, - "robustness": 0 + "robustness": 43.2 } }, "subcategories": { @@ -2552,12 +2552,12 @@ "easy": { "accuracy": 89.0, "cost": 0.0025, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 40.0, "cost": 0.0025, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -2567,7 +2567,7 @@ "all": { "accuracy": 69.5, "cost": 0.0023, - "robustness": 0 + "robustness": 71.4 } } }, @@ -2576,22 +2576,22 @@ "easy": { "accuracy": 91.9, "cost": 0.005, - "robustness": 0 + "robustness": 27.3 }, "medium": { "accuracy": 47.5, "cost": 0.0048, - "robustness": 0 + "robustness": 60.0 }, "hard": { "accuracy": 8.1, "cost": 0.004, - "robustness": 0 + "robustness": 20.0 }, "all": { "accuracy": 56.6, "cost": 0.0047, - "robustness": 0 + "robustness": 30.8 } } }, @@ -2600,7 +2600,7 @@ "easy": { "accuracy": 95.2, "cost": 0.0037, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 77.8, @@ -2615,7 +2615,7 @@ "all": { "accuracy": 90.4, "cost": 0.0036, - "robustness": 0 + "robustness": 75.0 } } } @@ -2626,22 +2626,22 @@ "easy": { "accuracy": 78.9, "cost": 0.0144, - "robustness": 0 + "robustness": 81.8 }, "medium": { "accuracy": 43.7, "cost": 0.0118, - "robustness": 0 + "robustness": 55.6 }, "hard": { "accuracy": 54.1, "cost": 0.0427, - "robustness": 0 + "robustness": 88.0 }, "all": { "accuracy": 59.1, "cost": 0.0269, - "robustness": 0 + "robustness": 80.0 } }, "subcategories": { @@ -2650,22 +2650,22 @@ "easy": { "accuracy": 78.9, "cost": 0.0144, - "robustness": 0 + "robustness": 81.8 }, "medium": { "accuracy": 43.7, "cost": 0.0118, - "robustness": 0 + "robustness": 55.6 }, "hard": { "accuracy": 54.1, "cost": 0.0427, - "robustness": 0 + "robustness": 88.0 }, "all": { "accuracy": 59.1, "cost": 0.0269, - "robustness": 0 + "robustness": 80.0 } } } @@ -2676,22 +2676,22 @@ "easy": { "accuracy": 97.0, "cost": 0.0052, - "robustness": 0 + "robustness": 76.9 }, "medium": { "accuracy": 78.8, "cost": 0.0116, - "robustness": 0 + "robustness": 52.4 }, "hard": { "accuracy": 26.5, "cost": 0.0231, - "robustness": 0 + "robustness": 86.7 }, "all": { "accuracy": 78.6, "cost": 0.0104, - "robustness": 0 + "robustness": 71.0 } }, "subcategories": { @@ -2700,22 +2700,22 @@ "easy": { "accuracy": 97.1, "cost": 0.0054, - "robustness": 0 + "robustness": 61.5 }, "medium": { "accuracy": 82.2, "cost": 0.013, - "robustness": 0 + "robustness": 47.1 }, "hard": { "accuracy": 30.1, "cost": 0.0263, - "robustness": 0 + "robustness": 84.6 }, "all": { "accuracy": 77.6, "cost": 0.0126, - "robustness": 0 + "robustness": 62.8 } } }, @@ -2724,22 +2724,22 @@ "easy": { "accuracy": 96.7, "cost": 0.0048, - "robustness": 0 + "robustness": 92.3 }, "medium": { "accuracy": 63.5, "cost": 0.0053, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 5.6, "cost": 0.0047, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 81.3, "cost": 0.0049, - "robustness": 0 + "robustness": 89.5 } } } @@ -3684,22 +3684,22 @@ "easy": { "accuracy": 93.5, "cost": 0.0003, - "robustness": 0 + "robustness": 55.9 }, "medium": { "accuracy": 61.3, "cost": 0.0007, - "robustness": 0 + "robustness": 58.3 }, "hard": { "accuracy": 26.5, "cost": 0.0009, - "robustness": 0 + "robustness": 46.8 }, "all": { "accuracy": 68.1, "cost": 0.0005, - "robustness": 0 + "robustness": 54.1 } }, "categories": { @@ -3708,22 +3708,22 @@ "easy": { "accuracy": 93.7, "cost": 0.0003, - "robustness": 0 + "robustness": 61.5 }, "medium": { "accuracy": 70.3, "cost": 0.0007, - "robustness": 0 + "robustness": 68.4 }, "hard": { "accuracy": 23.9, "cost": 0.0007, - "robustness": 0 + "robustness": 57.1 }, "all": { "accuracy": 78.3, "cost": 0.0005, - "robustness": 0 + "robustness": 63.1 } }, "subcategories": { @@ -3732,22 +3732,22 @@ "easy": { "accuracy": 92.8, "cost": 0.0003, - "robustness": 0 + "robustness": 64.7 }, "medium": { "accuracy": 61.0, "cost": 0.0004, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 15.0, "cost": 0.0004, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 76.1, "cost": 0.0003, - "robustness": 0 + "robustness": 60.9 } } }, @@ -3756,22 +3756,22 @@ "easy": { "accuracy": 98.4, "cost": 0.0005, - "robustness": 0 + "robustness": 40.0 }, "medium": { "accuracy": 83.8, "cost": 0.001, - "robustness": 0 + "robustness": 76.9 }, "hard": { "accuracy": 47.6, "cost": 0.0015, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 84.8, "cost": 0.0009, - "robustness": 0 + "robustness": 68.4 } } } @@ -3782,22 +3782,22 @@ "easy": { "accuracy": 92.4, "cost": 0.0003, - "robustness": 0 + "robustness": 60.9 }, "medium": { "accuracy": 51.4, "cost": 0.0004, - "robustness": 0 + "robustness": 14.3 }, "hard": { "accuracy": 7.9, "cost": 0.0005, - "robustness": 0 + "robustness": 42.9 }, "all": { "accuracy": 69.5, "cost": 0.0003, - "robustness": 0 + "robustness": 48.6 } }, "subcategories": { @@ -3806,22 +3806,22 @@ "easy": { "accuracy": 86.5, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 62.5, "cost": 0.0005, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 11.1, "cost": 0.0007, - "robustness": 0 + "robustness": 33.3 }, "all": { "accuracy": 50.8, "cost": 0.0005, - "robustness": 0 + "robustness": 62.5 } } }, @@ -3830,7 +3830,7 @@ "easy": { "accuracy": 95.0, "cost": 0.0002, - "robustness": 0 + "robustness": 83.3 }, "medium": { "accuracy": 77.1, @@ -3840,12 +3840,12 @@ "hard": { "accuracy": 5.6, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 81.8, "cost": 0.0003, - "robustness": 0 + "robustness": 87.5 } } }, @@ -3854,22 +3854,22 @@ "easy": { "accuracy": 91.8, "cost": 0.0003, - "robustness": 0 + "robustness": 42.9 }, "medium": { "accuracy": 31.8, "cost": 0.0004, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 0.0, "cost": 0.0003, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 69.4, "cost": 0.0003, - "robustness": 0 + "robustness": 28.6 } } } @@ -3880,22 +3880,22 @@ "easy": { "accuracy": 93.3, "cost": 0.0003, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 65.1, "cost": 0.0007, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 6.7, "cost": 0.0006, - "robustness": 0 + "robustness": 14.3 }, "all": { "accuracy": 64.7, "cost": 0.0005, - "robustness": 0 + "robustness": 43.2 } }, "subcategories": { @@ -3909,17 +3909,17 @@ "medium": { "accuracy": 33.3, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 2.9, "cost": 0.0006, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 18.2, "cost": 0.0006, - "robustness": 0 + "robustness": 33.3 } } }, @@ -3928,22 +3928,22 @@ "easy": { "accuracy": 93.4, "cost": 0.0003, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 70.6, "cost": 0.0007, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 7.5, "cost": 0.0006, - "robustness": 0 + "robustness": 33.3 }, "all": { "accuracy": 68.5, "cost": 0.0005, - "robustness": 0 + "robustness": 42.9 } } }, @@ -3952,7 +3952,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 44.4, @@ -3962,12 +3962,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0002, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 83.6, "cost": 0.0002, - "robustness": 0 + "robustness": 50.0 } } }, @@ -3976,22 +3976,22 @@ "easy": { "accuracy": 87.7, "cost": 0.0004, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 62.9, "cost": 0.0006, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 10.0, "cost": 0.0005, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 63.1, "cost": 0.0005, - "robustness": 0 + "robustness": 44.4 } } } @@ -4002,22 +4002,22 @@ "easy": { "accuracy": 97.0, "cost": 0.0002, - "robustness": 0 + "robustness": 59.0 }, "medium": { "accuracy": 75.0, "cost": 0.0006, - "robustness": 0 + "robustness": 61.9 }, "hard": { "accuracy": 27.8, "cost": 0.0012, - "robustness": 0 + "robustness": 55.6 }, "all": { "accuracy": 80.6, "cost": 0.0005, - "robustness": 0 + "robustness": 59.4 } }, "subcategories": { @@ -4026,7 +4026,7 @@ "easy": { "accuracy": 96.7, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 20.0, @@ -4036,12 +4036,12 @@ "hard": { "accuracy": 3.4, "cost": 0.0006, - "robustness": 0 + "robustness": 33.3 }, "all": { "accuracy": 31.6, "cost": 0.0005, - "robustness": 0 + "robustness": 75.0 } } }, @@ -4050,22 +4050,22 @@ "easy": { "accuracy": 96.9, "cost": 0.0003, - "robustness": 0 + "robustness": 36.4 }, "medium": { "accuracy": 80.4, "cost": 0.0008, - "robustness": 0 + "robustness": 81.8 }, "hard": { "accuracy": 40.4, "cost": 0.0019, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 81.7, "cost": 0.0007, - "robustness": 0 + "robustness": 63.0 } } }, @@ -4074,7 +4074,7 @@ "easy": { "accuracy": 98.3, "cost": 0.0003, - "robustness": 0 + "robustness": 57.1 }, "medium": { "accuracy": 64.7, @@ -4089,7 +4089,7 @@ "all": { "accuracy": 89.5, "cost": 0.0003, - "robustness": 0 + "robustness": 57.1 } } }, @@ -4098,12 +4098,12 @@ "easy": { "accuracy": 93.0, "cost": 0.0004, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 91.2, "cost": 0.0006, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 30.8, @@ -4113,7 +4113,7 @@ "all": { "accuracy": 83.3, "cost": 0.0005, - "robustness": 0 + "robustness": 40.0 } } }, @@ -4122,12 +4122,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0003, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 73.8, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 54.5, @@ -4137,7 +4137,7 @@ "all": { "accuracy": 82.5, "cost": 0.0005, - "robustness": 0 + "robustness": 60.0 } } }, @@ -4146,7 +4146,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0002, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 60.0, @@ -4161,7 +4161,7 @@ "all": { "accuracy": 92.0, "cost": 0.0002, - "robustness": 0 + "robustness": 50.0 } } }, @@ -4170,22 +4170,22 @@ "easy": { "accuracy": 96.9, "cost": 0.0002, - "robustness": 0 + "robustness": 62.5 }, "medium": { "accuracy": 71.2, "cost": 0.0004, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 30.8, "cost": 0.0004, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 89.7, "cost": 0.0003, - "robustness": 0 + "robustness": 53.3 } } } @@ -4196,22 +4196,22 @@ "easy": { "accuracy": 90.9, "cost": 0.0006, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 52.9, "cost": 0.0012, - "robustness": 0 + "robustness": 45.5 }, "hard": { "accuracy": 10.9, "cost": 0.0017, - "robustness": 0 + "robustness": 40.0 }, "all": { "accuracy": 60.3, "cost": 0.001, - "robustness": 0 + "robustness": 54.2 } }, "subcategories": { @@ -4220,22 +4220,22 @@ "easy": { "accuracy": 94.9, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 51.3, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 5.4, "cost": 0.0009, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 51.3, "cost": 0.0006, - "robustness": 0 + "robustness": 80.0 } } }, @@ -4244,22 +4244,22 @@ "easy": { "accuracy": 87.9, "cost": 0.0009, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 52.0, "cost": 0.0016, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 15.2, "cost": 0.0025, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 60.5, "cost": 0.0015, - "robustness": 0 + "robustness": 12.5 } } }, @@ -4268,22 +4268,22 @@ "easy": { "accuracy": 92.9, "cost": 0.0004, - "robustness": 0 + "robustness": 80.0 }, "medium": { "accuracy": 54.3, "cost": 0.0011, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 12.0, "cost": 0.0012, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 65.8, "cost": 0.0008, - "robustness": 0 + "robustness": 72.7 } } } @@ -4294,22 +4294,22 @@ "easy": { "accuracy": 93.1, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 45.1, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 24.8, "cost": 0.0006, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 38.0, "cost": 0.0006, - "robustness": 0 + "robustness": 62.1 } }, "subcategories": { @@ -4318,22 +4318,22 @@ "easy": { "accuracy": 93.1, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 45.1, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 24.8, "cost": 0.0006, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 38.0, "cost": 0.0006, - "robustness": 0 + "robustness": 62.1 } } } @@ -4344,22 +4344,22 @@ "easy": { "accuracy": 97.0, "cost": 0.0003, - "robustness": 0 + "robustness": 57.1 }, "medium": { "accuracy": 55.5, "cost": 0.0004, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 4.0, "cost": 0.0006, - "robustness": 0 + "robustness": 40.0 }, "all": { "accuracy": 65.3, "cost": 0.0004, - "robustness": 0 + "robustness": 54.1 } }, "subcategories": { @@ -4368,12 +4368,12 @@ "easy": { "accuracy": 97.3, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 42.1, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.3, @@ -4383,7 +4383,7 @@ "all": { "accuracy": 76.9, "cost": 0.0003, - "robustness": 0 + "robustness": 71.4 } } }, @@ -4392,22 +4392,22 @@ "easy": { "accuracy": 96.4, "cost": 0.0003, - "robustness": 0 + "robustness": 45.5 }, "medium": { "accuracy": 56.6, "cost": 0.0005, - "robustness": 0 + "robustness": 60.0 }, "hard": { "accuracy": 3.7, "cost": 0.0006, - "robustness": 0 + "robustness": 40.0 }, "all": { "accuracy": 60.0, "cost": 0.0004, - "robustness": 0 + "robustness": 46.2 } } }, @@ -4416,7 +4416,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0002, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 66.7, @@ -4431,7 +4431,7 @@ "all": { "accuracy": 92.3, "cost": 0.0002, - "robustness": 0 + "robustness": 75.0 } } } @@ -4442,22 +4442,22 @@ "easy": { "accuracy": 84.7, "cost": 0.0003, - "robustness": 0 + "robustness": 45.5 }, "medium": { "accuracy": 41.2, "cost": 0.0003, - "robustness": 0 + "robustness": 55.6 }, "hard": { "accuracy": 56.4, "cost": 0.0006, - "robustness": 0 + "robustness": 28.0 }, "all": { "accuracy": 61.3, "cost": 0.0004, - "robustness": 0 + "robustness": 37.8 } }, "subcategories": { @@ -4466,22 +4466,22 @@ "easy": { "accuracy": 84.7, "cost": 0.0003, - "robustness": 0 + "robustness": 45.5 }, "medium": { "accuracy": 41.2, "cost": 0.0003, - "robustness": 0 + "robustness": 55.6 }, "hard": { "accuracy": 56.4, "cost": 0.0006, - "robustness": 0 + "robustness": 28.0 }, "all": { "accuracy": 61.3, "cost": 0.0004, - "robustness": 0 + "robustness": 37.8 } } } @@ -4492,22 +4492,22 @@ "easy": { "accuracy": 91.9, "cost": 0.0003, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 58.2, "cost": 0.0006, - "robustness": 0 + "robustness": 52.4 }, "hard": { "accuracy": 31.8, "cost": 0.0018, - "robustness": 0 + "robustness": 73.3 }, "all": { "accuracy": 70.3, "cost": 0.0007, - "robustness": 0 + "robustness": 56.5 } }, "subcategories": { @@ -4516,22 +4516,22 @@ "easy": { "accuracy": 89.7, "cost": 0.0003, - "robustness": 0 + "robustness": 53.8 }, "medium": { "accuracy": 58.6, "cost": 0.0007, - "robustness": 0 + "robustness": 64.7 }, "hard": { "accuracy": 36.4, "cost": 0.0021, - "robustness": 0 + "robustness": 76.9 }, "all": { "accuracy": 67.0, "cost": 0.0008, - "robustness": 0 + "robustness": 65.1 } } }, @@ -4540,22 +4540,22 @@ "easy": { "accuracy": 95.3, "cost": 0.0002, - "robustness": 0 + "robustness": 46.2 }, "medium": { "accuracy": 56.5, "cost": 0.0004, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 5.6, "cost": 0.0003, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 78.8, "cost": 0.0003, - "robustness": 0 + "robustness": 36.8 } } } @@ -5500,22 +5500,22 @@ "easy": { "accuracy": 96.0, "cost": 0.0001, - "robustness": 0 + "robustness": 55.6 }, "medium": { "accuracy": 57.3, "cost": 0.0002, - "robustness": 0 + "robustness": 62.8 }, "hard": { "accuracy": 19.6, "cost": 0.0002, - "robustness": 0 + "robustness": 69.7 }, "all": { "accuracy": 66.9, "cost": 0.0002, - "robustness": 0 + "robustness": 61.2 } }, "categories": { @@ -5524,22 +5524,22 @@ "easy": { "accuracy": 97.7, "cost": 0.0001, - "robustness": 0 + "robustness": 73.1 }, "medium": { "accuracy": 65.6, "cost": 0.0003, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 16.3, "cost": 0.0008, - "robustness": 0 + "robustness": 93.3 }, "all": { "accuracy": 72.9, "cost": 0.0003, - "robustness": 0 + "robustness": 75.8 } }, "subcategories": { @@ -5548,22 +5548,22 @@ "easy": { "accuracy": 97.1, "cost": 0.0001, - "robustness": 0 + "robustness": 76.9 }, "medium": { "accuracy": 50.6, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 78.3, "cost": 0.0001, - "robustness": 0 + "robustness": 78.9 } } }, @@ -5572,22 +5572,22 @@ "easy": { "accuracy": 98.1, "cost": 0.0001, - "robustness": 0 + "robustness": 69.2 }, "medium": { "accuracy": 69.0, "cost": 0.0003, - "robustness": 0 + "robustness": 58.8 }, "hard": { "accuracy": 19.1, "cost": 0.0009, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 70.7, "cost": 0.0004, - "robustness": 0 + "robustness": 74.4 } } } @@ -5598,22 +5598,22 @@ "easy": { "accuracy": 96.3, "cost": 0.0001, - "robustness": 0 + "robustness": 40.7 }, "medium": { "accuracy": 48.8, "cost": 0.0001, - "robustness": 0 + "robustness": 30.0 }, "hard": { "accuracy": 1.2, "cost": 0.0001, - "robustness": 0 + "robustness": 42.9 }, "all": { "accuracy": 71.0, "cost": 0.0001, - "robustness": 0 + "robustness": 38.6 } }, "subcategories": { @@ -5622,22 +5622,22 @@ "easy": { "accuracy": 96.2, "cost": 0.0001, - "robustness": 0 + "robustness": 14.3 }, "medium": { "accuracy": 32.9, "cost": 0.0001, - "robustness": 0 + "robustness": 20.0 }, "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 72.8, "cost": 0.0001, - "robustness": 0 + "robustness": 14.3 } } }, @@ -5646,22 +5646,22 @@ "easy": { "accuracy": 94.6, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 45.0, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 2.2, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 44.3, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 } } }, @@ -5670,7 +5670,7 @@ "easy": { "accuracy": 95.9, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 68.8, @@ -5680,12 +5680,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 79.7, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 } } }, @@ -5694,12 +5694,12 @@ "easy": { "accuracy": 98.5, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 63.2, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 0.0, @@ -5709,7 +5709,7 @@ "all": { "accuracy": 81.5, "cost": 0.0001, - "robustness": 0 + "robustness": 85.7 } } } @@ -5720,22 +5720,22 @@ "easy": { "accuracy": 96.3, "cost": 0.0001, - "robustness": 0 + "robustness": 88.9 }, "medium": { "accuracy": 53.8, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 2.7, "cost": 0.0002, - "robustness": 0 + "robustness": 57.1 }, "all": { "accuracy": 61.0, "cost": 0.0001, - "robustness": 0 + "robustness": 75.7 } }, "subcategories": { @@ -5744,22 +5744,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0001, - "robustness": 0 + "robustness": 83.3 }, "medium": { "accuracy": 54.6, "cost": 0.0002, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 1.2, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 62.9, "cost": 0.0002, - "robustness": 0 + "robustness": 81.0 } } }, @@ -5768,22 +5768,22 @@ "easy": { "accuracy": 93.8, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 52.4, "cost": 0.0001, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 6.7, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 60.8, "cost": 0.0001, - "robustness": 0 + "robustness": 88.9 } } }, @@ -5797,17 +5797,17 @@ "medium": { "accuracy": 66.7, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 2.9, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 29.1, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 } } }, @@ -5816,7 +5816,7 @@ "easy": { "accuracy": 95.7, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 22.2, @@ -5826,12 +5826,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 77.0, "cost": 0.0001, - "robustness": 0 + "robustness": 75.0 } } } @@ -5842,22 +5842,22 @@ "easy": { "accuracy": 86.6, "cost": 0.0001, - "robustness": 0 + "robustness": 36.4 }, "medium": { "accuracy": 34.1, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 51.4, "cost": 0.0001, - "robustness": 0 + "robustness": 76.0 }, "all": { "accuracy": 57.8, "cost": 0.0001, - "robustness": 0 + "robustness": 64.4 } }, "subcategories": { @@ -5866,22 +5866,22 @@ "easy": { "accuracy": 86.6, "cost": 0.0001, - "robustness": 0 + "robustness": 36.4 }, "medium": { "accuracy": 34.1, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 51.4, "cost": 0.0001, - "robustness": 0 + "robustness": 76.0 }, "all": { "accuracy": 57.8, "cost": 0.0001, - "robustness": 0 + "robustness": 64.4 } } } @@ -5892,22 +5892,22 @@ "easy": { "accuracy": 98.1, "cost": 0.0001, - "robustness": 0 + "robustness": 41.0 }, "medium": { "accuracy": 66.0, "cost": 0.0002, - "robustness": 0 + "robustness": 57.1 }, "hard": { "accuracy": 4.6, "cost": 0.0003, - "robustness": 0 + "robustness": 77.8 }, "all": { "accuracy": 75.2, "cost": 0.0001, - "robustness": 0 + "robustness": 50.7 } }, "subcategories": { @@ -5916,22 +5916,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0001, - "robustness": 0 + "robustness": 45.5 }, "medium": { "accuracy": 67.2, "cost": 0.0002, - "robustness": 0 + "robustness": 72.7 }, "hard": { "accuracy": 5.6, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 71.2, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 } } }, @@ -5940,22 +5940,22 @@ "easy": { "accuracy": 99.2, "cost": 0.0001, - "robustness": 0 + "robustness": 37.5 }, "medium": { "accuracy": 64.4, "cost": 0.0001, - "robustness": 0 + "robustness": 33.3 }, "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 89.1, "cost": 0.0001, - "robustness": 0 + "robustness": 33.3 } } }, @@ -5964,7 +5964,7 @@ "easy": { "accuracy": 98.3, "cost": 0.0001, - "robustness": 0 + "robustness": 42.9 }, "medium": { "accuracy": 70.6, @@ -5979,7 +5979,7 @@ "all": { "accuracy": 88.8, "cost": 0.0001, - "robustness": 0 + "robustness": 42.9 } } }, @@ -5988,12 +5988,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0001, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 73.5, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 0.0, @@ -6003,7 +6003,7 @@ "all": { "accuracy": 74.4, "cost": 0.0001, - "robustness": 0 + "robustness": 40.0 } } }, @@ -6012,12 +6012,12 @@ "easy": { "accuracy": 95.5, "cost": 0.0001, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 69.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 9.1, @@ -6027,7 +6027,7 @@ "all": { "accuracy": 74.2, "cost": 0.0001, - "robustness": 0 + "robustness": 40.0 } } }, @@ -6036,7 +6036,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 60.0, @@ -6051,7 +6051,7 @@ "all": { "accuracy": 92.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 } } }, @@ -6060,7 +6060,7 @@ "easy": { "accuracy": 96.7, "cost": 0.0001, - "robustness": 0 + "robustness": 40.0 }, "medium": { "accuracy": 40.0, @@ -6070,12 +6070,12 @@ "hard": { "accuracy": 5.1, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 36.8, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 } } } @@ -6086,22 +6086,22 @@ "easy": { "accuracy": 94.4, "cost": 0.0001, - "robustness": 0 + "robustness": 62.8 }, "medium": { "accuracy": 53.8, "cost": 0.0001, - "robustness": 0 + "robustness": 71.4 }, "hard": { "accuracy": 5.7, "cost": 0.0001, - "robustness": 0 + "robustness": 71.4 }, "all": { "accuracy": 72.3, "cost": 0.0001, - "robustness": 0 + "robustness": 66.2 } }, "subcategories": { @@ -6110,22 +6110,22 @@ "easy": { "accuracy": 97.6, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 68.9, "cost": 0.0002, - "robustness": 0 + "robustness": 76.9 }, "hard": { "accuracy": 7.1, "cost": 0.0002, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 71.9, "cost": 0.0002, - "robustness": 0 + "robustness": 78.9 } } }, @@ -6134,22 +6134,22 @@ "easy": { "accuracy": 93.2, "cost": 0.0001, - "robustness": 0 + "robustness": 58.8 }, "medium": { "accuracy": 43.8, "cost": 0.0001, - "robustness": 0 + "robustness": 57.1 }, "hard": { "accuracy": 5.3, "cost": 0.0001, - "robustness": 0 + "robustness": 83.3 }, "all": { "accuracy": 71.2, "cost": 0.0001, - "robustness": 0 + "robustness": 61.7 } } }, @@ -6158,12 +6158,12 @@ "easy": { "accuracy": 100.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 45.5, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -6173,7 +6173,7 @@ "all": { "accuracy": 89.0, "cost": 0.0001, - "robustness": 0 + "robustness": 60.0 } } } @@ -6184,22 +6184,22 @@ "easy": { "accuracy": 94.3, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 49.0, "cost": 0.0001, - "robustness": 0 + "robustness": 72.7 }, "hard": { "accuracy": 3.4, "cost": 0.0002, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 58.7, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 } }, "subcategories": { @@ -6208,22 +6208,22 @@ "easy": { "accuracy": 90.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 53.0, "cost": 0.0002, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 6.1, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 59.8, "cost": 0.0002, - "robustness": 0 + "robustness": 87.5 } } }, @@ -6232,22 +6232,22 @@ "easy": { "accuracy": 98.0, "cost": 0.0001, - "robustness": 0 + "robustness": 60.0 }, "medium": { "accuracy": 39.7, "cost": 0.0001, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 59.6, "cost": 0.0001, - "robustness": 0 + "robustness": 72.7 } } }, @@ -6256,22 +6256,22 @@ "easy": { "accuracy": 98.3, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "medium": { "accuracy": 66.7, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 1.8, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 55.2, "cost": 0.0001, - "robustness": 0 + "robustness": 20.0 } } } @@ -6282,22 +6282,22 @@ "easy": { "accuracy": 99.3, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 77.3, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 27.2, "cost": 0.0002, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 43.7, "cost": 0.0002, - "robustness": 0 + "robustness": 58.6 } }, "subcategories": { @@ -6306,22 +6306,22 @@ "easy": { "accuracy": 99.3, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 77.3, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 27.2, "cost": 0.0002, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 43.7, "cost": 0.0002, - "robustness": 0 + "robustness": 58.6 } } } @@ -6332,22 +6332,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0001, - "robustness": 0 + "robustness": 47.8 }, "medium": { "accuracy": 64.4, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 6.0, "cost": 0.0001, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 69.6, "cost": 0.0001, - "robustness": 0 + "robustness": 53.8 } }, "subcategories": { @@ -6356,12 +6356,12 @@ "easy": { "accuracy": 100.0, "cost": 0.0001, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 72.4, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 33.3, @@ -6371,7 +6371,7 @@ "all": { "accuracy": 89.3, "cost": 0.0001, - "robustness": 0 + "robustness": 33.3 } } }, @@ -6380,22 +6380,22 @@ "easy": { "accuracy": 96.9, "cost": 0.0001, - "robustness": 0 + "robustness": 72.7 }, "medium": { "accuracy": 62.6, "cost": 0.0001, - "robustness": 0 + "robustness": 60.0 }, "hard": { "accuracy": 3.7, "cost": 0.0001, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 61.8, "cost": 0.0001, - "robustness": 0 + "robustness": 65.4 } } }, @@ -6404,7 +6404,7 @@ "easy": { "accuracy": 92.9, "cost": 0.0001, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 66.7, @@ -6419,7 +6419,7 @@ "all": { "accuracy": 86.5, "cost": 0.0001, - "robustness": 0 + "robustness": 25.0 } } } @@ -6432,22 +6432,22 @@ "easy": { "accuracy": 93.5, "cost": 0.0005, - "robustness": 0 + "robustness": 49.0 }, "medium": { "accuracy": 46.4, "cost": 0.0007, - "robustness": 0 + "robustness": 56.6 }, "hard": { "accuracy": 17.4, "cost": 0.0009, - "robustness": 0 + "robustness": 42.2 }, "all": { "accuracy": 62.0, "cost": 0.0007, - "robustness": 0 + "robustness": 49.3 } }, "categories": { @@ -6456,22 +6456,22 @@ "easy": { "accuracy": 94.9, "cost": 0.0005, - "robustness": 0 + "robustness": 38.5 }, "medium": { "accuracy": 51.3, "cost": 0.0009, - "robustness": 0 + "robustness": 47.6 }, "hard": { "accuracy": 6.5, "cost": 0.0016, - "robustness": 0 + "robustness": 33.3 }, "all": { "accuracy": 65.1, "cost": 0.0008, - "robustness": 0 + "robustness": 40.3 } }, "subcategories": { @@ -6480,22 +6480,22 @@ "easy": { "accuracy": 96.0, "cost": 0.0004, - "robustness": 0 + "robustness": 38.5 }, "medium": { "accuracy": 45.9, "cost": 0.0003, - "robustness": 0 + "robustness": 25.0 }, "hard": { "accuracy": 2.8, "cost": 0.0003, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 76.8, "cost": 0.0003, - "robustness": 0 + "robustness": 36.8 } } }, @@ -6504,22 +6504,22 @@ "easy": { "accuracy": 94.3, "cost": 0.0006, - "robustness": 0 + "robustness": 38.5 }, "medium": { "accuracy": 52.5, "cost": 0.0011, - "robustness": 0 + "robustness": 52.9 }, "hard": { "accuracy": 7.2, "cost": 0.0018, - "robustness": 0 + "robustness": 30.8 }, "all": { "accuracy": 60.5, "cost": 0.001, - "robustness": 0 + "robustness": 41.9 } } } @@ -6530,22 +6530,22 @@ "easy": { "accuracy": 93.1, "cost": 0.0005, - "robustness": 0 + "robustness": 40.7 }, "medium": { "accuracy": 47.9, "cost": 0.0005, - "robustness": 0 + "robustness": 40.0 }, "hard": { "accuracy": 3.7, "cost": 0.0004, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 69.1, "cost": 0.0005, - "robustness": 0 + "robustness": 34.1 } }, "subcategories": { @@ -6554,22 +6554,22 @@ "easy": { "accuracy": 89.7, "cost": 0.0004, - "robustness": 0 + "robustness": 42.9 }, "medium": { "accuracy": 38.8, "cost": 0.0005, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 0.0, "cost": 0.0004, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 70.3, "cost": 0.0004, - "robustness": 0 + "robustness": 28.6 } } }, @@ -6578,22 +6578,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0005, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 47.5, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 4.4, "cost": 0.0004, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 46.7, "cost": 0.0005, - "robustness": 0 + "robustness": 25.0 } } }, @@ -6602,7 +6602,7 @@ "easy": { "accuracy": 98.3, "cost": 0.0005, - "robustness": 0 + "robustness": 16.7 }, "medium": { "accuracy": 56.2, @@ -6612,12 +6612,12 @@ "hard": { "accuracy": 5.6, "cost": 0.0007, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 78.6, "cost": 0.0005, - "robustness": 0 + "robustness": 12.5 } } }, @@ -6626,12 +6626,12 @@ "easy": { "accuracy": 90.8, "cost": 0.0007, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 57.9, "cost": 0.0006, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -6641,7 +6641,7 @@ "all": { "accuracy": 75.0, "cost": 0.0007, - "robustness": 0 + "robustness": 85.7 } } } @@ -6652,22 +6652,22 @@ "easy": { "accuracy": 93.3, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 47.4, "cost": 0.0008, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 5.4, "cost": 0.001, - "robustness": 0 + "robustness": 57.1 }, "all": { "accuracy": 58.0, "cost": 0.0008, - "robustness": 0 + "robustness": 56.8 } }, "subcategories": { @@ -6676,22 +6676,22 @@ "easy": { "accuracy": 94.0, "cost": 0.0006, - "robustness": 0 + "robustness": 58.3 }, "medium": { "accuracy": 48.5, "cost": 0.0009, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 7.5, "cost": 0.0014, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 60.3, "cost": 0.0009, - "robustness": 0 + "robustness": 66.7 } } }, @@ -6700,22 +6700,22 @@ "easy": { "accuracy": 89.2, "cost": 0.0007, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 50.8, "cost": 0.0009, - "robustness": 0 + "robustness": 25.0 }, "hard": { "accuracy": 6.7, "cost": 0.0007, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 58.2, "cost": 0.0008, - "robustness": 0 + "robustness": 33.3 } } }, @@ -6729,17 +6729,17 @@ "medium": { "accuracy": 44.4, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0004, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 20.0, "cost": 0.0004, - "robustness": 0 + "robustness": 66.7 } } }, @@ -6748,7 +6748,7 @@ "easy": { "accuracy": 95.7, "cost": 0.0003, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 11.1, @@ -6758,12 +6758,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 75.4, "cost": 0.0003, - "robustness": 0 + "robustness": 50.0 } } } @@ -6774,22 +6774,22 @@ "easy": { "accuracy": 84.7, "cost": 0.0003, - "robustness": 0 + "robustness": 45.5 }, "medium": { "accuracy": 26.3, "cost": 0.0006, - "robustness": 0 + "robustness": 33.3 }, "hard": { "accuracy": 51.9, "cost": 0.0004, - "robustness": 0 + "robustness": 36.0 }, "all": { "accuracy": 55.6, "cost": 0.0004, - "robustness": 0 + "robustness": 37.8 } }, "subcategories": { @@ -6798,22 +6798,22 @@ "easy": { "accuracy": 84.7, "cost": 0.0003, - "robustness": 0 + "robustness": 45.5 }, "medium": { "accuracy": 26.3, "cost": 0.0006, - "robustness": 0 + "robustness": 33.3 }, "hard": { "accuracy": 51.9, "cost": 0.0004, - "robustness": 0 + "robustness": 36.0 }, "all": { "accuracy": 55.6, "cost": 0.0004, - "robustness": 0 + "robustness": 37.8 } } } @@ -6824,22 +6824,22 @@ "easy": { "accuracy": 94.8, "cost": 0.0005, - "robustness": 0 + "robustness": 51.3 }, "medium": { "accuracy": 55.8, "cost": 0.0006, - "robustness": 0 + "robustness": 61.9 }, "hard": { "accuracy": 1.0, "cost": 0.0007, - "robustness": 0 + "robustness": 44.4 }, "all": { "accuracy": 69.8, "cost": 0.0006, - "robustness": 0 + "robustness": 53.6 } }, "subcategories": { @@ -6848,22 +6848,22 @@ "easy": { "accuracy": 91.2, "cost": 0.0005, - "robustness": 0 + "robustness": 63.6 }, "medium": { "accuracy": 57.6, "cost": 0.0006, - "robustness": 0 + "robustness": 72.7 }, "hard": { "accuracy": 1.1, "cost": 0.0008, - "robustness": 0 + "robustness": 40.0 }, "all": { "accuracy": 63.9, "cost": 0.0006, - "robustness": 0 + "robustness": 63.0 } } }, @@ -6872,22 +6872,22 @@ "easy": { "accuracy": 98.1, "cost": 0.0004, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 61.0, "cost": 0.0006, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 0.0, "cost": 0.0005, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 87.6, "cost": 0.0005, - "robustness": 0 + "robustness": 40.0 } } }, @@ -6896,7 +6896,7 @@ "easy": { "accuracy": 95.7, "cost": 0.0006, - "robustness": 0 + "robustness": 28.6 }, "medium": { "accuracy": 41.2, @@ -6911,7 +6911,7 @@ "all": { "accuracy": 83.2, "cost": 0.0007, - "robustness": 0 + "robustness": 28.6 } } }, @@ -6920,12 +6920,12 @@ "easy": { "accuracy": 88.4, "cost": 0.0006, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 55.9, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 0.0, @@ -6935,7 +6935,7 @@ "all": { "accuracy": 63.3, "cost": 0.0006, - "robustness": 0 + "robustness": 60.0 } } }, @@ -6944,12 +6944,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0005, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 57.1, "cost": 0.0008, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 0.0, @@ -6959,7 +6959,7 @@ "all": { "accuracy": 69.1, "cost": 0.0007, - "robustness": 0 + "robustness": 40.0 } } }, @@ -6968,7 +6968,7 @@ "easy": { "accuracy": 95.0, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 40.0, @@ -6983,7 +6983,7 @@ "all": { "accuracy": 84.0, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 } } }, @@ -6992,7 +6992,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0004, - "robustness": 0 + "robustness": 80.0 }, "medium": { "accuracy": 36.0, @@ -7002,12 +7002,12 @@ "hard": { "accuracy": 1.7, "cost": 0.0005, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 35.1, "cost": 0.0004, - "robustness": 0 + "robustness": 75.0 } } } @@ -7018,22 +7018,22 @@ "easy": { "accuracy": 93.3, "cost": 0.0006, - "robustness": 0 + "robustness": 48.8 }, "medium": { "accuracy": 47.1, "cost": 0.0007, - "robustness": 0 + "robustness": 71.4 }, "hard": { "accuracy": 5.1, "cost": 0.0007, - "robustness": 0 + "robustness": 71.4 }, "all": { "accuracy": 69.6, "cost": 0.0007, - "robustness": 0 + "robustness": 57.7 } }, "subcategories": { @@ -7042,22 +7042,22 @@ "easy": { "accuracy": 94.4, "cost": 0.0005, - "robustness": 0 + "robustness": 80.0 }, "medium": { "accuracy": 57.5, "cost": 0.0009, - "robustness": 0 + "robustness": 69.2 }, "hard": { "accuracy": 7.1, "cost": 0.0008, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 65.1, "cost": 0.0007, - "robustness": 0 + "robustness": 73.7 } } }, @@ -7066,22 +7066,22 @@ "easy": { "accuracy": 92.6, "cost": 0.0007, - "robustness": 0 + "robustness": 47.1 }, "medium": { "accuracy": 41.3, "cost": 0.0006, - "robustness": 0 + "robustness": 71.4 }, "hard": { "accuracy": 4.4, "cost": 0.0007, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 70.1, "cost": 0.0007, - "robustness": 0 + "robustness": 53.2 } } }, @@ -7090,12 +7090,12 @@ "easy": { "accuracy": 98.3, "cost": 0.0003, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 18.2, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -7105,7 +7105,7 @@ "all": { "accuracy": 83.6, "cost": 0.0003, - "robustness": 0 + "robustness": 40.0 } } } @@ -7116,22 +7116,22 @@ "easy": { "accuracy": 91.9, "cost": 0.0005, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 32.2, "cost": 0.0008, - "robustness": 0 + "robustness": 45.5 }, "hard": { "accuracy": 1.4, "cost": 0.001, - "robustness": 0 + "robustness": 20.0 }, "all": { "accuracy": 51.1, "cost": 0.0007, - "robustness": 0 + "robustness": 58.3 } }, "subcategories": { @@ -7140,22 +7140,22 @@ "easy": { "accuracy": 88.6, "cost": 0.0007, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 41.0, "cost": 0.0013, - "robustness": 0 + "robustness": 25.0 }, "hard": { "accuracy": 3.0, "cost": 0.0016, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 54.6, "cost": 0.0011, - "robustness": 0 + "robustness": 37.5 } } }, @@ -7164,22 +7164,22 @@ "easy": { "accuracy": 97.0, "cost": 0.0004, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 25.9, "cost": 0.0005, - "robustness": 0 + "robustness": 60.0 }, "hard": { "accuracy": 0.0, "cost": 0.0005, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 52.5, "cost": 0.0004, - "robustness": 0 + "robustness": 72.7 } } }, @@ -7188,22 +7188,22 @@ "easy": { "accuracy": 91.5, "cost": 0.0003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 28.2, "cost": 0.0004, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 0.0, "cost": 0.0006, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 42.2, "cost": 0.0004, - "robustness": 0 + "robustness": 60.0 } } } @@ -7214,22 +7214,22 @@ "easy": { "accuracy": 95.6, "cost": 0.0006, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 48.9, "cost": 0.0007, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 25.3, "cost": 0.0012, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 38.9, "cost": 0.001, - "robustness": 0 + "robustness": 58.6 } }, "subcategories": { @@ -7238,22 +7238,22 @@ "easy": { "accuracy": 95.6, "cost": 0.0006, - "robustness": 0 + "robustness": 33.3 }, "medium": { "accuracy": 48.9, "cost": 0.0007, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 25.3, "cost": 0.0012, - "robustness": 0 + "robustness": 58.3 }, "all": { "accuracy": 38.9, "cost": 0.001, - "robustness": 0 + "robustness": 58.6 } } } @@ -7264,22 +7264,22 @@ "easy": { "accuracy": 94.9, "cost": 0.0006, - "robustness": 0 + "robustness": 52.2 }, "medium": { "accuracy": 44.6, "cost": 0.0007, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 2.0, "cost": 0.0007, - "robustness": 0 + "robustness": 40.0 }, "all": { "accuracy": 62.4, "cost": 0.0006, - "robustness": 0 + "robustness": 51.3 } }, "subcategories": { @@ -7288,12 +7288,12 @@ "easy": { "accuracy": 96.3, "cost": 0.0003, - "robustness": 0 + "robustness": 37.5 }, "medium": { "accuracy": 44.8, "cost": 0.0003, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 0.0, @@ -7303,7 +7303,7 @@ "all": { "accuracy": 78.7, "cost": 0.0003, - "robustness": 0 + "robustness": 33.3 } } }, @@ -7312,22 +7312,22 @@ "easy": { "accuracy": 94.6, "cost": 0.0008, - "robustness": 0 + "robustness": 72.7 }, "medium": { "accuracy": 44.6, "cost": 0.0008, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 2.2, "cost": 0.0008, - "robustness": 0 + "robustness": 40.0 }, "all": { "accuracy": 55.4, "cost": 0.0008, - "robustness": 0 + "robustness": 61.5 } } }, @@ -7336,7 +7336,7 @@ "easy": { "accuracy": 92.9, "cost": 0.0003, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 44.4, @@ -7351,7 +7351,7 @@ "all": { "accuracy": 82.7, "cost": 0.0003, - "robustness": 0 + "robustness": 25.0 } } } @@ -8296,22 +8296,22 @@ "easy": { "accuracy": 89.3, "cost": 0.0031, - "robustness": 0 + "robustness": 83.3 }, "medium": { "accuracy": 42.2, "cost": 0.0051, - "robustness": 0 + "robustness": 84.1 }, "hard": { "accuracy": 17.1, "cost": 0.0057, - "robustness": 0 + "robustness": 82.6 }, "all": { "accuracy": 58.7, "cost": 0.0043, - "robustness": 0 + "robustness": 83.3 } }, "categories": { @@ -8320,22 +8320,22 @@ "easy": { "accuracy": 87.0, "cost": 0.0024, - "robustness": 0 + "robustness": 88.5 }, "medium": { "accuracy": 39.6, "cost": 0.0067, - "robustness": 0 + "robustness": 85.7 }, "hard": { "accuracy": 1.6, "cost": 0.0111, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 56.4, "cost": 0.0053, - "robustness": 0 + "robustness": 82.3 } }, "subcategories": { @@ -8344,22 +8344,22 @@ "easy": { "accuracy": 81.5, "cost": 0.0013, - "robustness": 0 + "robustness": 76.9 }, "medium": { "accuracy": 36.5, "cost": 0.0013, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 0.0, "cost": 0.0012, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 64.4, "cost": 0.0013, - "robustness": 0 + "robustness": 73.7 } } }, @@ -8368,22 +8368,22 @@ "easy": { "accuracy": 90.7, "cost": 0.0031, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 40.3, "cost": 0.0079, - "robustness": 0 + "robustness": 88.2 }, "hard": { "accuracy": 1.9, "cost": 0.0128, - "robustness": 0 + "robustness": 69.2 }, "all": { "accuracy": 53.3, "cost": 0.0069, - "robustness": 0 + "robustness": 86.0 } } } @@ -8394,22 +8394,22 @@ "easy": { "accuracy": 89.0, "cost": 0.0033, - "robustness": 0 + "robustness": 96.3 }, "medium": { "accuracy": 38.4, "cost": 0.0038, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 3.7, "cost": 0.0026, - "robustness": 0 + "robustness": 85.7 }, "all": { "accuracy": 63.9, "cost": 0.0034, - "robustness": 0 + "robustness": 95.5 } }, "subcategories": { @@ -8418,22 +8418,22 @@ "easy": { "accuracy": 84.3, "cost": 0.0032, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 30.6, "cost": 0.004, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 7.7, "cost": 0.0034, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 64.7, "cost": 0.0035, - "robustness": 0 + "robustness": 100.0 } } }, @@ -8442,22 +8442,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0024, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 37.5, "cost": 0.0032, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 4.4, "cost": 0.0018, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 43.4, "cost": 0.0024, - "robustness": 0 + "robustness": 87.5 } } }, @@ -8466,7 +8466,7 @@ "easy": { "accuracy": 91.7, "cost": 0.0035, - "robustness": 0 + "robustness": 83.3 }, "medium": { "accuracy": 56.2, @@ -8476,12 +8476,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0036, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 73.8, "cost": 0.0035, - "robustness": 0 + "robustness": 87.5 } } }, @@ -8490,12 +8490,12 @@ "easy": { "accuracy": 92.3, "cost": 0.0036, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 34.2, "cost": 0.0041, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -8505,7 +8505,7 @@ "all": { "accuracy": 67.6, "cost": 0.0038, - "robustness": 0 + "robustness": 100.0 } } } @@ -8516,22 +8516,22 @@ "easy": { "accuracy": 87.9, "cost": 0.0037, - "robustness": 0 + "robustness": 61.1 }, "medium": { "accuracy": 41.5, "cost": 0.0058, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 2.0, "cost": 0.0065, - "robustness": 0 + "robustness": 85.7 }, "all": { "accuracy": 52.9, "cost": 0.0051, - "robustness": 0 + "robustness": 67.6 } }, "subcategories": { @@ -8540,22 +8540,22 @@ "easy": { "accuracy": 84.7, "cost": 0.004, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 40.5, "cost": 0.0068, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 3.8, "cost": 0.0097, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 52.6, "cost": 0.0062, - "robustness": 0 + "robustness": 61.9 } } }, @@ -8564,22 +8564,22 @@ "easy": { "accuracy": 90.8, "cost": 0.0044, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 44.4, "cost": 0.005, - "robustness": 0 + "robustness": 25.0 }, "hard": { "accuracy": 0.0, "cost": 0.0048, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 55.1, "cost": 0.0047, - "robustness": 0 + "robustness": 66.7 } } }, @@ -8593,17 +8593,17 @@ "medium": { "accuracy": 50.0, "cost": 0.0011, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0011, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 21.8, "cost": 0.0011, - "robustness": 0 + "robustness": 100.0 } } }, @@ -8612,7 +8612,7 @@ "easy": { "accuracy": 95.7, "cost": 0.0018, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 22.2, @@ -8622,12 +8622,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 77.0, "cost": 0.0018, - "robustness": 0 + "robustness": 75.0 } } } @@ -8638,22 +8638,22 @@ "easy": { "accuracy": 86.1, "cost": 0.0021, - "robustness": 0 + "robustness": 81.8 }, "medium": { "accuracy": 46.7, "cost": 0.0034, - "robustness": 0 + "robustness": 88.9 }, "hard": { "accuracy": 49.4, "cost": 0.0037, - "robustness": 0 + "robustness": 92.0 }, "all": { "accuracy": 59.7, "cost": 0.0032, - "robustness": 0 + "robustness": 88.9 } }, "subcategories": { @@ -8662,22 +8662,22 @@ "easy": { "accuracy": 86.1, "cost": 0.0021, - "robustness": 0 + "robustness": 81.8 }, "medium": { "accuracy": 46.7, "cost": 0.0034, - "robustness": 0 + "robustness": 88.9 }, "hard": { "accuracy": 49.4, "cost": 0.0037, - "robustness": 0 + "robustness": 92.0 }, "all": { "accuracy": 59.7, "cost": 0.0032, - "robustness": 0 + "robustness": 88.9 } } } @@ -8688,22 +8688,22 @@ "easy": { "accuracy": 91.3, "cost": 0.0029, - "robustness": 0 + "robustness": 74.4 }, "medium": { "accuracy": 39.6, "cost": 0.0054, - "robustness": 0 + "robustness": 71.4 }, "hard": { "accuracy": 3.6, "cost": 0.0072, - "robustness": 0 + "robustness": 77.8 }, "all": { "accuracy": 63.2, "cost": 0.0043, - "robustness": 0 + "robustness": 73.9 } }, "subcategories": { @@ -8712,22 +8712,22 @@ "easy": { "accuracy": 89.3, "cost": 0.0039, - "robustness": 0 + "robustness": 90.9 }, "medium": { "accuracy": 40.4, "cost": 0.0069, - "robustness": 0 + "robustness": 63.6 }, "hard": { "accuracy": 4.5, "cost": 0.013, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 56.4, "cost": 0.0065, - "robustness": 0 + "robustness": 74.1 } } }, @@ -8736,22 +8736,22 @@ "easy": { "accuracy": 91.5, "cost": 0.0013, - "robustness": 0 + "robustness": 25.0 }, "medium": { "accuracy": 22.0, "cost": 0.001, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 15.4, "cost": 0.0009, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 76.1, "cost": 0.0012, - "robustness": 0 + "robustness": 46.7 } } }, @@ -8760,7 +8760,7 @@ "easy": { "accuracy": 93.2, "cost": 0.0043, - "robustness": 0 + "robustness": 85.7 }, "medium": { "accuracy": 29.4, @@ -8775,7 +8775,7 @@ "all": { "accuracy": 79.7, "cost": 0.0045, - "robustness": 0 + "robustness": 85.7 } } }, @@ -8784,12 +8784,12 @@ "easy": { "accuracy": 95.3, "cost": 0.0044, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 52.9, "cost": 0.0059, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -8799,7 +8799,7 @@ "all": { "accuracy": 65.6, "cost": 0.005, - "robustness": 0 + "robustness": 80.0 } } }, @@ -8808,12 +8808,12 @@ "easy": { "accuracy": 90.9, "cost": 0.003, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 50.0, "cost": 0.0056, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -8823,7 +8823,7 @@ "all": { "accuracy": 62.9, "cost": 0.0043, - "robustness": 0 + "robustness": 100.0 } } }, @@ -8832,7 +8832,7 @@ "easy": { "accuracy": 85.0, "cost": 0.0014, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 60.0, @@ -8847,7 +8847,7 @@ "all": { "accuracy": 80.0, "cost": 0.0014, - "robustness": 0 + "robustness": 50.0 } } }, @@ -8856,7 +8856,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 40.0, @@ -8866,12 +8866,12 @@ "hard": { "accuracy": 1.7, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 36.0, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 } } } @@ -8882,22 +8882,22 @@ "easy": { "accuracy": 87.2, "cost": 0.0039, - "robustness": 0 + "robustness": 83.7 }, "medium": { "accuracy": 33.8, "cost": 0.0051, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 3.2, "cost": 0.0047, - "robustness": 0 + "robustness": 57.1 }, "all": { "accuracy": 61.8, "cost": 0.0043, - "robustness": 0 + "robustness": 83.1 } }, "subcategories": { @@ -8906,22 +8906,22 @@ "easy": { "accuracy": 84.9, "cost": 0.0039, - "robustness": 0 + "robustness": 80.0 }, "medium": { "accuracy": 35.3, "cost": 0.0075, - "robustness": 0 + "robustness": 92.3 }, "hard": { "accuracy": 0.0, "cost": 0.0073, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 49.6, "cost": 0.0061, - "robustness": 0 + "robustness": 84.2 } } }, @@ -8930,22 +8930,22 @@ "easy": { "accuracy": 87.4, "cost": 0.0042, - "robustness": 0 + "robustness": 85.3 }, "medium": { "accuracy": 33.5, "cost": 0.0036, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 4.4, "cost": 0.0038, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 64.8, "cost": 0.004, - "robustness": 0 + "robustness": 85.1 } } }, @@ -8954,12 +8954,12 @@ "easy": { "accuracy": 90.0, "cost": 0.0012, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 18.2, "cost": 0.0009, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 0.0, @@ -8969,7 +8969,7 @@ "all": { "accuracy": 76.7, "cost": 0.0012, - "robustness": 0 + "robustness": 60.0 } } } @@ -8980,22 +8980,22 @@ "easy": { "accuracy": 90.9, "cost": 0.0028, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 47.8, "cost": 0.0046, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 7.5, "cost": 0.0059, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 57.7, "cost": 0.0041, - "robustness": 0 + "robustness": 87.5 } }, "subcategories": { @@ -9004,22 +9004,22 @@ "easy": { "accuracy": 89.3, "cost": 0.0039, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 61.0, "cost": 0.0082, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 6.1, "cost": 0.0107, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 62.1, "cost": 0.0068, - "robustness": 0 + "robustness": 100.0 } } }, @@ -9028,22 +9028,22 @@ "easy": { "accuracy": 89.9, "cost": 0.0021, - "robustness": 0 + "robustness": 60.0 }, "medium": { "accuracy": 30.2, "cost": 0.0027, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 16.0, "cost": 0.003, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 53.3, "cost": 0.0025, - "robustness": 0 + "robustness": 72.7 } } }, @@ -9052,22 +9052,22 @@ "easy": { "accuracy": 96.6, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 66.7, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 5.4, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 55.8, "cost": 0.0015, - "robustness": 0 + "robustness": 100.0 } } } @@ -9078,22 +9078,22 @@ "easy": { "accuracy": 95.0, "cost": 0.0027, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 71.7, "cost": 0.0029, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 25.6, "cost": 0.0051, - "robustness": 0 + "robustness": 87.5 }, "all": { "accuracy": 41.3, "cost": 0.0045, - "robustness": 0 + "robustness": 86.2 } }, "subcategories": { @@ -9102,22 +9102,22 @@ "easy": { "accuracy": 95.0, "cost": 0.0027, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 71.7, "cost": 0.0029, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 25.6, "cost": 0.0051, - "robustness": 0 + "robustness": 87.5 }, "all": { "accuracy": 41.3, "cost": 0.0045, - "robustness": 0 + "robustness": 86.2 } } } @@ -9128,22 +9128,22 @@ "easy": { "accuracy": 93.6, "cost": 0.0029, - "robustness": 0 + "robustness": 95.7 }, "medium": { "accuracy": 57.6, "cost": 0.0034, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 6.7, "cost": 0.0034, - "robustness": 0 + "robustness": 90.0 }, "all": { "accuracy": 66.0, "cost": 0.0031, - "robustness": 0 + "robustness": 92.3 } }, "subcategories": { @@ -9152,12 +9152,12 @@ "easy": { "accuracy": 89.9, "cost": 0.001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 58.6, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 25.0, @@ -9167,7 +9167,7 @@ "all": { "accuracy": 78.7, "cost": 0.0011, - "robustness": 0 + "robustness": 100.0 } } }, @@ -9176,22 +9176,22 @@ "easy": { "accuracy": 94.6, "cost": 0.004, - "robustness": 0 + "robustness": 90.9 }, "medium": { "accuracy": 57.6, "cost": 0.004, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 5.1, "cost": 0.0036, - "robustness": 0 + "robustness": 90.0 }, "all": { "accuracy": 59.8, "cost": 0.0039, - "robustness": 0 + "robustness": 88.5 } } }, @@ -9200,7 +9200,7 @@ "easy": { "accuracy": 97.6, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 55.6, @@ -9215,7 +9215,7 @@ "all": { "accuracy": 88.5, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 } } } @@ -9228,22 +9228,22 @@ "easy": { "accuracy": 93.5, "cost": 0.0036, - "robustness": 0 + "robustness": 84.8 }, "medium": { "accuracy": 45.5, "cost": 0.0057, - "robustness": 0 + "robustness": 85.0 }, "hard": { "accuracy": 17.0, "cost": 0.0064, - "robustness": 0 + "robustness": 66.1 }, "all": { "accuracy": 61.6, "cost": 0.0048, - "robustness": 0 + "robustness": 80.0 } }, "categories": { @@ -9252,22 +9252,22 @@ "easy": { "accuracy": 94.8, "cost": 0.0032, - "robustness": 0 + "robustness": 96.2 }, "medium": { "accuracy": 43.7, "cost": 0.0081, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 1.6, "cost": 0.0142, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 61.6, "cost": 0.0068, - "robustness": 0 + "robustness": 90.3 } }, "subcategories": { @@ -9276,22 +9276,22 @@ "easy": { "accuracy": 94.2, "cost": 0.0024, - "robustness": 0 + "robustness": 92.3 }, "medium": { "accuracy": 47.1, "cost": 0.0021, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0023, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 75.5, "cost": 0.0023, - "robustness": 0 + "robustness": 94.7 } } }, @@ -9300,22 +9300,22 @@ "easy": { "accuracy": 95.2, "cost": 0.0037, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 43.0, "cost": 0.0095, - "robustness": 0 + "robustness": 88.2 }, "hard": { "accuracy": 1.9, "cost": 0.0163, - "robustness": 0 + "robustness": 76.9 }, "all": { "accuracy": 56.2, "cost": 0.0085, - "robustness": 0 + "robustness": 88.4 } } } @@ -9326,22 +9326,22 @@ "easy": { "accuracy": 90.0, "cost": 0.0035, - "robustness": 0 + "robustness": 96.3 }, "medium": { "accuracy": 39.3, "cost": 0.0039, - "robustness": 0 + "robustness": 60.0 }, "hard": { "accuracy": 3.7, "cost": 0.0025, - "robustness": 0 + "robustness": 57.1 }, "all": { "accuracy": 64.7, "cost": 0.0035, - "robustness": 0 + "robustness": 81.8 } }, "subcategories": { @@ -9350,22 +9350,22 @@ "easy": { "accuracy": 84.3, "cost": 0.0032, - "robustness": 0 + "robustness": 92.9 }, "medium": { "accuracy": 30.6, "cost": 0.004, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 7.7, "cost": 0.0034, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 64.7, "cost": 0.0035, - "robustness": 0 + "robustness": 90.5 } } }, @@ -9374,22 +9374,22 @@ "easy": { "accuracy": 97.3, "cost": 0.0028, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 40.0, "cost": 0.0033, - "robustness": 0 + "robustness": 0.0 }, "hard": { "accuracy": 4.4, "cost": 0.0017, - "robustness": 0 + "robustness": 33.3 }, "all": { "accuracy": 44.3, "cost": 0.0026, - "robustness": 0 + "robustness": 50.0 } } }, @@ -9398,7 +9398,7 @@ "easy": { "accuracy": 95.0, "cost": 0.004, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 58.3, @@ -9408,12 +9408,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0035, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 76.5, "cost": 0.0039, - "robustness": 0 + "robustness": 87.5 } } }, @@ -9422,12 +9422,12 @@ "easy": { "accuracy": 92.3, "cost": 0.0036, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 34.2, "cost": 0.0041, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 0.0, @@ -9437,7 +9437,7 @@ "all": { "accuracy": 67.6, "cost": 0.0038, - "robustness": 0 + "robustness": 85.7 } } } @@ -9448,22 +9448,22 @@ "easy": { "accuracy": 93.3, "cost": 0.0043, - "robustness": 0 + "robustness": 94.4 }, "medium": { "accuracy": 46.6, "cost": 0.0063, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 3.4, "cost": 0.0067, - "robustness": 0 + "robustness": 85.7 }, "all": { "accuracy": 57.3, "cost": 0.0055, - "robustness": 0 + "robustness": 89.2 } }, "subcategories": { @@ -9472,22 +9472,22 @@ "easy": { "accuracy": 91.8, "cost": 0.0048, - "robustness": 0 + "robustness": 91.7 }, "medium": { "accuracy": 46.0, "cost": 0.0072, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 5.0, "cost": 0.01, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 58.0, "cost": 0.0067, - "robustness": 0 + "robustness": 90.5 } } }, @@ -9496,22 +9496,22 @@ "easy": { "accuracy": 93.8, "cost": 0.0048, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 49.2, "cost": 0.006, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 3.3, "cost": 0.0052, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 58.9, "cost": 0.0053, - "robustness": 0 + "robustness": 88.9 } } }, @@ -9525,17 +9525,17 @@ "medium": { "accuracy": 55.6, "cost": 0.0012, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 0.0, "cost": 0.0011, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 23.6, "cost": 0.0012, - "robustness": 0 + "robustness": 66.7 } } }, @@ -9544,7 +9544,7 @@ "easy": { "accuracy": 97.9, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 22.2, @@ -9554,12 +9554,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 78.7, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 } } } @@ -9570,22 +9570,22 @@ "easy": { "accuracy": 88.0, "cost": 0.0023, - "robustness": 0 + "robustness": 72.7 }, "medium": { "accuracy": 46.1, "cost": 0.0036, - "robustness": 0 + "robustness": 88.9 }, "hard": { "accuracy": 49.5, "cost": 0.0038, - "robustness": 0 + "robustness": 56.0 }, "all": { "accuracy": 60.2, "cost": 0.0033, - "robustness": 0 + "robustness": 66.7 } }, "subcategories": { @@ -9594,22 +9594,22 @@ "easy": { "accuracy": 88.0, "cost": 0.0023, - "robustness": 0 + "robustness": 72.7 }, "medium": { "accuracy": 46.1, "cost": 0.0036, - "robustness": 0 + "robustness": 88.9 }, "hard": { "accuracy": 49.5, "cost": 0.0038, - "robustness": 0 + "robustness": 56.0 }, "all": { "accuracy": 60.2, "cost": 0.0033, - "robustness": 0 + "robustness": 66.7 } } } @@ -9620,22 +9620,22 @@ "easy": { "accuracy": 95.9, "cost": 0.0036, - "robustness": 0 + "robustness": 79.5 }, "medium": { "accuracy": 44.7, "cost": 0.006, - "robustness": 0 + "robustness": 81.0 }, "hard": { "accuracy": 3.1, "cost": 0.0074, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 67.2, "cost": 0.0048, - "robustness": 0 + "robustness": 78.3 } }, "subcategories": { @@ -9644,22 +9644,22 @@ "easy": { "accuracy": 92.4, "cost": 0.0042, - "robustness": 0 + "robustness": 81.8 }, "medium": { "accuracy": 44.4, "cost": 0.0073, - "robustness": 0 + "robustness": 81.8 }, "hard": { "accuracy": 3.4, "cost": 0.0131, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 59.2, "cost": 0.0068, - "robustness": 0 + "robustness": 77.8 } } }, @@ -9668,22 +9668,22 @@ "easy": { "accuracy": 97.7, "cost": 0.0027, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 37.3, "cost": 0.0027, - "robustness": 0 + "robustness": 66.7 }, "hard": { "accuracy": 15.4, "cost": 0.0023, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 83.6, "cost": 0.0027, - "robustness": 0 + "robustness": 73.3 } } }, @@ -9692,7 +9692,7 @@ "easy": { "accuracy": 97.4, "cost": 0.0046, - "robustness": 0 + "robustness": 71.4 }, "medium": { "accuracy": 29.4, @@ -9707,7 +9707,7 @@ "all": { "accuracy": 83.2, "cost": 0.0048, - "robustness": 0 + "robustness": 71.4 } } }, @@ -9716,12 +9716,12 @@ "easy": { "accuracy": 97.7, "cost": 0.0043, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 55.9, "cost": 0.0061, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -9731,7 +9731,7 @@ "all": { "accuracy": 67.8, "cost": 0.0051, - "robustness": 0 + "robustness": 80.0 } } }, @@ -9740,12 +9740,12 @@ "easy": { "accuracy": 95.5, "cost": 0.0034, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 54.8, "cost": 0.0059, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -9755,7 +9755,7 @@ "all": { "accuracy": 67.0, "cost": 0.0046, - "robustness": 0 + "robustness": 80.0 } } }, @@ -9764,7 +9764,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 60.0, @@ -9779,7 +9779,7 @@ "all": { "accuracy": 92.0, "cost": 0.0016, - "robustness": 0 + "robustness": 100.0 } } }, @@ -9788,7 +9788,7 @@ "easy": { "accuracy": 100.0, "cost": 0.0017, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 40.0, @@ -9798,12 +9798,12 @@ "hard": { "accuracy": 1.7, "cost": 0.0013, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 36.0, "cost": 0.0014, - "robustness": 0 + "robustness": 87.5 } } } @@ -9814,22 +9814,22 @@ "easy": { "accuracy": 91.6, "cost": 0.0043, - "robustness": 0 + "robustness": 74.4 }, "medium": { "accuracy": 38.1, "cost": 0.0058, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 3.8, "cost": 0.0052, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 65.7, "cost": 0.0049, - "robustness": 0 + "robustness": 81.7 } }, "subcategories": { @@ -9838,22 +9838,22 @@ "easy": { "accuracy": 90.5, "cost": 0.0046, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 38.3, "cost": 0.0086, - "robustness": 0 + "robustness": 84.6 }, "hard": { "accuracy": 0.0, "cost": 0.0081, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 53.1, "cost": 0.007, - "robustness": 0 + "robustness": 89.5 } } }, @@ -9862,22 +9862,22 @@ "easy": { "accuracy": 91.4, "cost": 0.0045, - "robustness": 0 + "robustness": 70.6 }, "medium": { "accuracy": 37.6, "cost": 0.004, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 5.3, "cost": 0.0042, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 68.4, "cost": 0.0043, - "robustness": 0 + "robustness": 78.7 } } }, @@ -9886,12 +9886,12 @@ "easy": { "accuracy": 96.7, "cost": 0.0022, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 45.5, "cost": 0.0022, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -9901,7 +9901,7 @@ "all": { "accuracy": 86.3, "cost": 0.0022, - "robustness": 0 + "robustness": 80.0 } } } @@ -9912,22 +9912,22 @@ "easy": { "accuracy": 94.0, "cost": 0.0031, - "robustness": 0 + "robustness": 87.5 }, "medium": { "accuracy": 49.4, "cost": 0.0048, - "robustness": 0 + "robustness": 90.9 }, "hard": { "accuracy": 7.5, "cost": 0.0059, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 59.6, "cost": 0.0043, - "robustness": 0 + "robustness": 83.3 } }, "subcategories": { @@ -9936,22 +9936,22 @@ "easy": { "accuracy": 91.4, "cost": 0.004, - "robustness": 0 + "robustness": 50.0 }, "medium": { "accuracy": 63.0, "cost": 0.0083, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 6.1, "cost": 0.0107, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 63.7, "cost": 0.0069, - "robustness": 0 + "robustness": 87.5 } } }, @@ -9960,22 +9960,22 @@ "easy": { "accuracy": 93.9, "cost": 0.0027, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 30.2, "cost": 0.0029, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 16.0, "cost": 0.0033, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 55.0, "cost": 0.0029, - "robustness": 0 + "robustness": 90.9 } } }, @@ -9984,22 +9984,22 @@ "easy": { "accuracy": 100.0, "cost": 0.0018, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 71.8, "cost": 0.0014, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 5.4, "cost": 0.0014, - "robustness": 0 + "robustness": 0.0 }, "all": { "accuracy": 58.4, "cost": 0.0016, - "robustness": 0 + "robustness": 60.0 } } } @@ -10010,22 +10010,22 @@ "easy": { "accuracy": 95.4, "cost": 0.0028, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 65.3, "cost": 0.0024, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 24.7, "cost": 0.0057, - "robustness": 0 + "robustness": 45.8 }, "all": { "accuracy": 40.1, "cost": 0.0049, - "robustness": 0 + "robustness": 48.3 } }, "subcategories": { @@ -10034,22 +10034,22 @@ "easy": { "accuracy": 95.4, "cost": 0.0028, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 65.3, "cost": 0.0024, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 24.7, "cost": 0.0057, - "robustness": 0 + "robustness": 45.8 }, "all": { "accuracy": 40.1, "cost": 0.0049, - "robustness": 0 + "robustness": 48.3 } } } @@ -10060,22 +10060,22 @@ "easy": { "accuracy": 96.0, "cost": 0.0032, - "robustness": 0 + "robustness": 87.0 }, "medium": { "accuracy": 62.1, "cost": 0.0038, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 6.7, "cost": 0.0036, - "robustness": 0 + "robustness": 90.0 }, "all": { "accuracy": 68.4, "cost": 0.0035, - "robustness": 0 + "robustness": 89.7 } }, "subcategories": { @@ -10084,12 +10084,12 @@ "easy": { "accuracy": 95.4, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 58.6, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 25.0, @@ -10099,7 +10099,7 @@ "all": { "accuracy": 82.7, "cost": 0.0013, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10108,22 +10108,22 @@ "easy": { "accuracy": 96.0, "cost": 0.0044, - "robustness": 0 + "robustness": 72.7 }, "medium": { "accuracy": 62.6, "cost": 0.0045, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 5.1, "cost": 0.0038, - "robustness": 0 + "robustness": 90.0 }, "all": { "accuracy": 61.8, "cost": 0.0043, - "robustness": 0 + "robustness": 84.6 } } }, @@ -10132,7 +10132,7 @@ "easy": { "accuracy": 97.6, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 66.7, @@ -10147,7 +10147,7 @@ "all": { "accuracy": 90.4, "cost": 0.0019, - "robustness": 0 + "robustness": 100.0 } } } @@ -10160,22 +10160,22 @@ "easy": { "accuracy": 54.3, "cost": 0.0001, - "robustness": 0 + "robustness": 88.4 }, "medium": { "accuracy": 14.4, "cost": 0.0001, - "robustness": 0 + "robustness": 87.6 }, "hard": { "accuracy": 8.7, "cost": 0.0001, - "robustness": 0 + "robustness": 77.1 }, "all": { "accuracy": 32.0, "cost": 0.0001, - "robustness": 0 + "robustness": 85.2 } }, "categories": { @@ -10184,22 +10184,22 @@ "easy": { "accuracy": 55.3, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 11.0, "cost": 0.0001, - "robustness": 0 + "robustness": 76.2 }, "hard": { "accuracy": 1.2, "cost": 0.0002, - "robustness": 0 + "robustness": 86.7 }, "all": { "accuracy": 31.2, "cost": 0.0001, - "robustness": 0 + "robustness": 88.7 } }, "subcategories": { @@ -10208,22 +10208,22 @@ "easy": { "accuracy": 58.2, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 16.5, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 8.3, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 44.7, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10232,22 +10232,22 @@ "easy": { "accuracy": 53.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 9.8, "cost": 0.0001, - "robustness": 0 + "robustness": 70.6 }, "hard": { "accuracy": 0.0, "cost": 0.0002, - "robustness": 0 + "robustness": 84.6 }, "all": { "accuracy": 25.9, "cost": 0.0001, - "robustness": 0 + "robustness": 83.7 } } } @@ -10258,22 +10258,22 @@ "easy": { "accuracy": 52.9, "cost": 0.0, - "robustness": 0 + "robustness": 92.6 }, "medium": { "accuracy": 19.0, "cost": 0.0001, - "robustness": 0 + "robustness": 90.0 }, "hard": { "accuracy": 7.4, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 37.4, "cost": 0.0001, - "robustness": 0 + "robustness": 93.2 } }, "subcategories": { @@ -10282,22 +10282,22 @@ "easy": { "accuracy": 60.0, "cost": 0.0, - "robustness": 0 + "robustness": 85.7 }, "medium": { "accuracy": 30.6, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 7.7, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 48.8, "cost": 0.0, - "robustness": 0 + "robustness": 90.5 } } }, @@ -10306,22 +10306,22 @@ "easy": { "accuracy": 32.4, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 10.0, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 2.2, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 13.9, "cost": 0.0, - "robustness": 0 + "robustness": 87.5 } } }, @@ -10330,7 +10330,7 @@ "easy": { "accuracy": 49.6, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 18.8, @@ -10340,12 +10340,12 @@ "hard": { "accuracy": 11.1, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 38.0, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10354,12 +10354,12 @@ "easy": { "accuracy": 50.8, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 2.6, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 40.0, @@ -10369,7 +10369,7 @@ "all": { "accuracy": 33.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 } } } @@ -10380,22 +10380,22 @@ "easy": { "accuracy": 46.3, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 11.1, "cost": 0.0001, - "robustness": 0 + "robustness": 75.0 }, "hard": { "accuracy": 2.7, "cost": 0.0001, - "robustness": 0 + "robustness": 85.7 }, "all": { "accuracy": 24.3, "cost": 0.0001, - "robustness": 0 + "robustness": 73.0 } }, "subcategories": { @@ -10404,22 +10404,22 @@ "easy": { "accuracy": 47.0, "cost": 0.0001, - "robustness": 0 + "robustness": 75.0 }, "medium": { "accuracy": 10.4, "cost": 0.0001, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 3.8, "cost": 0.0002, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 24.9, "cost": 0.0001, - "robustness": 0 + "robustness": 76.2 } } }, @@ -10428,22 +10428,22 @@ "easy": { "accuracy": 29.2, "cost": 0.0001, - "robustness": 0 + "robustness": 0.0 }, "medium": { "accuracy": 6.3, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 3.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 15.2, "cost": 0.0001, - "robustness": 0 + "robustness": 44.4 } } }, @@ -10457,17 +10457,17 @@ "medium": { "accuracy": 5.6, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 3.6, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10476,7 +10476,7 @@ "easy": { "accuracy": 68.1, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 66.7, @@ -10486,12 +10486,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 62.3, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } } @@ -10502,22 +10502,22 @@ "easy": { "accuracy": 47.8, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 24.0, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 24.2, "cost": 0.0001, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 31.2, "cost": 0.0001, - "robustness": 0 + "robustness": 77.8 } }, "subcategories": { @@ -10526,22 +10526,22 @@ "easy": { "accuracy": 47.8, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 24.0, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 24.2, "cost": 0.0001, - "robustness": 0 + "robustness": 60.0 }, "all": { "accuracy": 31.2, "cost": 0.0001, - "robustness": 0 + "robustness": 77.8 } } } @@ -10552,22 +10552,22 @@ "easy": { "accuracy": 57.0, "cost": 0.0001, - "robustness": 0 + "robustness": 89.7 }, "medium": { "accuracy": 12.0, "cost": 0.0001, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 4.1, "cost": 0.0001, - "robustness": 0 + "robustness": 88.9 }, "all": { "accuracy": 35.8, "cost": 0.0001, - "robustness": 0 + "robustness": 89.9 } }, "subcategories": { @@ -10576,22 +10576,22 @@ "easy": { "accuracy": 50.4, "cost": 0.0001, - "robustness": 0 + "robustness": 90.9 }, "medium": { "accuracy": 11.6, "cost": 0.0001, - "robustness": 0 + "robustness": 90.9 }, "hard": { "accuracy": 2.2, "cost": 0.0001, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 27.1, "cost": 0.0001, - "robustness": 0 + "robustness": 88.9 } } }, @@ -10600,22 +10600,22 @@ "easy": { "accuracy": 67.4, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 8.5, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 15.4, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 54.8, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10624,7 +10624,7 @@ "easy": { "accuracy": 61.5, "cost": 0.0001, - "robustness": 0 + "robustness": 71.4 }, "medium": { "accuracy": 23.5, @@ -10639,7 +10639,7 @@ "all": { "accuracy": 53.8, "cost": 0.0001, - "robustness": 0 + "robustness": 71.4 } } }, @@ -10648,12 +10648,12 @@ "easy": { "accuracy": 44.2, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 23.5, "cost": 0.0001, - "robustness": 0 + "robustness": 50.0 }, "hard": { "accuracy": 15.4, @@ -10663,7 +10663,7 @@ "all": { "accuracy": 32.2, "cost": 0.0001, - "robustness": 0 + "robustness": 80.0 } } }, @@ -10672,12 +10672,12 @@ "easy": { "accuracy": 40.9, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "medium": { "accuracy": 9.5, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 9.1, @@ -10687,7 +10687,7 @@ "all": { "accuracy": 23.7, "cost": 0.0001, - "robustness": 0 + "robustness": 80.0 } } }, @@ -10696,7 +10696,7 @@ "easy": { "accuracy": 55.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 0.0, @@ -10711,7 +10711,7 @@ "all": { "accuracy": 44.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10720,7 +10720,7 @@ "easy": { "accuracy": 50.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 8.0, @@ -10730,12 +10730,12 @@ "hard": { "accuracy": 0.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 14.9, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } } @@ -10746,22 +10746,22 @@ "easy": { "accuracy": 57.4, "cost": 0.0001, - "robustness": 0 + "robustness": 76.7 }, "medium": { "accuracy": 18.6, "cost": 0.0001, - "robustness": 0 + "robustness": 90.5 }, "hard": { "accuracy": 1.9, "cost": 0.0001, - "robustness": 0 + "robustness": 85.7 }, "all": { "accuracy": 39.5, "cost": 0.0001, - "robustness": 0 + "robustness": 81.7 } }, "subcategories": { @@ -10770,22 +10770,22 @@ "easy": { "accuracy": 49.2, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 15.6, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 26.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10794,22 +10794,22 @@ "easy": { "accuracy": 57.6, "cost": 0.0001, - "robustness": 0 + "robustness": 70.6 }, "medium": { "accuracy": 20.2, "cost": 0.0001, - "robustness": 0 + "robustness": 71.4 }, "hard": { "accuracy": 2.7, "cost": 0.0001, - "robustness": 0 + "robustness": 83.3 }, "all": { "accuracy": 42.2, "cost": 0.0001, - "robustness": 0 + "robustness": 72.3 } } }, @@ -10818,12 +10818,12 @@ "easy": { "accuracy": 71.7, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 27.3, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -10833,7 +10833,7 @@ "all": { "accuracy": 63.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } } @@ -10844,22 +10844,22 @@ "easy": { "accuracy": 52.7, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 15.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.7, "cost": 0.0001, - "robustness": 0 + "robustness": 80.0 }, "all": { "accuracy": 28.1, "cost": 0.0001, - "robustness": 0 + "robustness": 95.8 } }, "subcategories": { @@ -10868,22 +10868,22 @@ "easy": { "accuracy": 43.6, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 11.0, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0002, - "robustness": 0 + "robustness": 50.0 }, "all": { "accuracy": 23.5, "cost": 0.0001, - "robustness": 0 + "robustness": 87.5 } } }, @@ -10892,22 +10892,22 @@ "easy": { "accuracy": 59.6, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 22.4, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 35.4, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } }, @@ -10916,22 +10916,22 @@ "easy": { "accuracy": 62.7, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 5.1, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 1.8, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "all": { "accuracy": 26.0, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } } @@ -10942,22 +10942,22 @@ "easy": { "accuracy": 57.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 9.5, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 12.4, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 19.4, "cost": 0.0001, - "robustness": 0 + "robustness": 72.4 } }, "subcategories": { @@ -10966,22 +10966,22 @@ "easy": { "accuracy": 57.3, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 9.5, "cost": 0.0001, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 12.4, "cost": 0.0001, - "robustness": 0 + "robustness": 66.7 }, "all": { "accuracy": 19.4, "cost": 0.0001, - "robustness": 0 + "robustness": 72.4 } } } @@ -10992,22 +10992,22 @@ "easy": { "accuracy": 51.6, "cost": 0.0001, - "robustness": 0 + "robustness": 95.7 }, "medium": { "accuracy": 10.2, "cost": 0.0001, - "robustness": 0 + "robustness": 83.3 }, "hard": { "accuracy": 2.7, "cost": 0.0001, - "robustness": 0 + "robustness": 90.0 }, "all": { "accuracy": 30.7, "cost": 0.0001, - "robustness": 0 + "robustness": 92.3 } }, "subcategories": { @@ -11016,12 +11016,12 @@ "easy": { "accuracy": 60.6, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 6.9, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "hard": { "accuracy": 0.0, @@ -11031,7 +11031,7 @@ "all": { "accuracy": 45.3, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } }, @@ -11040,22 +11040,22 @@ "easy": { "accuracy": 48.9, "cost": 0.0001, - "robustness": 0 + "robustness": 90.9 }, "medium": { "accuracy": 10.8, "cost": 0.0001, - "robustness": 0 + "robustness": 80.0 }, "hard": { "accuracy": 2.2, "cost": 0.0001, - "robustness": 0 + "robustness": 90.0 }, "all": { "accuracy": 25.5, "cost": 0.0001, - "robustness": 0 + "robustness": 88.5 } } }, @@ -11064,7 +11064,7 @@ "easy": { "accuracy": 42.9, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 }, "medium": { "accuracy": 11.1, @@ -11079,7 +11079,7 @@ "all": { "accuracy": 38.5, "cost": 0.0, - "robustness": 0 + "robustness": 100.0 } } } diff --git a/src/data/routerMetrics/leaderboard.json b/src/data/routerMetrics/leaderboard.json index bca1080..5324fd0 100644 --- a/src/data/routerMetrics/leaderboard.json +++ b/src/data/routerMetrics/leaderboard.json @@ -5,7 +5,7 @@ "Optimal Selection Score": 39.84, "Optimal Cost Score": 72.998, "Optimal Acc. Score": 49.05, - "Robustness Score": 97.6, + "Robustness Score": 85.24, "Latency Score": 10.7527, "Accuracy": 32.01, "Cost per 1k": 0.07 @@ -16,7 +16,7 @@ "Optimal Selection Score": 22.52, "Optimal Cost Score": 46.322, "Optimal Acc. Score": 81.96, - "Robustness Score": null, + "Robustness Score": 54.07, "Latency Score": null, "Accuracy": 68.09, "Cost per 1k": 0.54 @@ -27,7 +27,7 @@ "Optimal Selection Score": 2.68, "Optimal Cost Score": 6.7697, "Optimal Acc. Score": 78.63, - "Robustness Score": 93.6, + "Robustness Score": 89.05, "Latency Score": 1.4993, "Accuracy": 67.21, "Cost per 1k": 2.06 @@ -49,7 +49,7 @@ "Optimal Selection Score": 4.73, "Optimal Cost Score": 38.3347, "Optimal Acc. Score": 74.25, - "Robustness Score": 97.5, + "Robustness Score": 94.29, "Latency Score": 2.6954, "Accuracy": 57.0, "Cost per 1k": 0.34 @@ -60,7 +60,7 @@ "Optimal Selection Score": 3.44, "Optimal Cost Score": 19.6178, "Optimal Acc. Score": 78.18, - "Robustness Score": 94.5, + "Robustness Score": 61.19, "Latency Score": 27.027, "Accuracy": 66.88, "Cost per 1k": 0.15 @@ -71,7 +71,7 @@ "Optimal Selection Score": 3.83, "Optimal Cost Score": 14.039, "Optimal Acc. Score": 77.88, - "Robustness Score": 44.5, + "Robustness Score": 49.29, "Latency Score": 10.4167, "Accuracy": 66.34, "Cost per 1k": 0.21 @@ -82,7 +82,7 @@ "Optimal Selection Score": 1.55, "Optimal Cost Score": 2.1367, "Optimal Acc. Score": 76.81, - "Robustness Score": null, + "Robustness Score": 55.91, "Latency Score": null, "Accuracy": 60.83, "Cost per 1k": 4.1 @@ -93,7 +93,7 @@ "Optimal Selection Score": 99.72, "Optimal Cost Score": 99.6314, "Optimal Acc. Score": 68.76, - "Robustness Score": 99.8, + "Robustness Score": 100.00, "Latency Score": 0.4016, "Accuracy": 47.04, "Cost per 1k": 0.27 @@ -104,7 +104,7 @@ "Optimal Selection Score": 13.09, "Optimal Cost Score": 25.4887, "Optimal Acc. Score": 78.77, - "Robustness Score": 51.3, + "Robustness Score": 83.33, "Latency Score": 1.328, "Accuracy": 58.69, "Cost per 1k": 4.27 @@ -115,7 +115,7 @@ "Optimal Selection Score": 13.39, "Optimal Cost Score": 24.4499, "Optimal Acc. Score": 83.32, - "Robustness Score": 96.9, + "Robustness Score": 80.00, "Latency Score": 90.9091, "Accuracy": 61.62, "Cost per 1k": 4.83 @@ -126,7 +126,7 @@ "Optimal Selection Score": 4.79, "Optimal Cost Score": 12.5426, "Optimal Acc. Score": 79.33, - "Robustness Score": 100.0, + "Robustness Score": 35.00, "Latency Score": 0.1863, "Accuracy": 67.28, "Cost per 1k": 1.67 diff --git a/src/pages/LeaderboardPage.tsx b/src/pages/LeaderboardPage.tsx index 932e06f..e1d6bf0 100644 --- a/src/pages/LeaderboardPage.tsx +++ b/src/pages/LeaderboardPage.tsx @@ -859,7 +859,7 @@ platforms. synonyms, and typos) and measure how often the router selects the same model as for the original query.

- +

Higher values indicate greater stability under realistic input noise, reflecting robust model selection.