Upload results.json

results.json

@@ -0,0 +1 @@
+{"vidore/syntheticDocQA_government_reports_test": {"ndcg_at_1": 0.58, "ndcg_at_3": 0.70464, "ndcg_at_5": 0.73347, "ndcg_at_10": 0.74908, "ndcg_at_20": 0.7517, "ndcg_at_100": 0.76475, "ndcg_at_1000": 0.7662, "map_at_1": 0.58, "map_at_3": 0.675, "map_at_5": 0.691, "map_at_10": 0.69711, "map_at_20": 0.69788, "map_at_100": 0.6997, "map_at_1000": 0.69978, "recall_at_1": 0.58, "recall_at_3": 0.79, "recall_at_5": 0.86, "recall_at_10": 0.91, "recall_at_20": 0.92, "recall_at_100": 0.99, "recall_at_1000": 1.0, "precision_at_1": 0.58, "precision_at_3": 0.26333, "precision_at_5": 0.172, "precision_at_10": 0.091, "precision_at_20": 0.046, "precision_at_100": 0.0099, "precision_at_1000": 0.001, "mrr_at_1": 0.58, "mrr_at_3": 0.675, "mrr_at_5": 0.6910000000000001, "mrr_at_10": 0.6971111111111112, "mrr_at_20": 0.6978803418803422, "mrr_at_100": 0.699699094575387, "mrr_at_1000": 0.6997845646608574, "naucs_at_1_max": 0.35087210145979036, "naucs_at_1_std": -0.2717735148155674, "naucs_at_1_diff1": 0.5728589256174361, "naucs_at_3_max": 0.48760429942049166, "naucs_at_3_std": -0.1952052324043572, "naucs_at_3_diff1": 0.48204967898622125, "naucs_at_5_max": 0.6072814188721066, "naucs_at_5_std": -0.12238464736039634, "naucs_at_5_diff1": 0.458673964251073, "naucs_at_10_max": 0.6720095445585643, "naucs_at_10_std": -0.016184251478367565, "naucs_at_10_diff1": 0.4494760867309872, "naucs_at_20_max": 0.7112511671335223, "naucs_at_20_std": 0.19928804855275828, "naucs_at_20_diff1": 0.39700046685340834, "naucs_at_100_max": 0.7222222222222041, "naucs_at_100_std": 0.35807656395892007, "naucs_at_100_diff1": 1.0, "naucs_at_1000_max": NaN, "naucs_at_1000_std": NaN, "naucs_at_1000_diff1": NaN}, "vidore/syntheticDocQA_healthcare_industry_test": {"ndcg_at_1": 0.69, "ndcg_at_3": 0.77547, "ndcg_at_5": 0.79226, "ndcg_at_10": 0.79939, "ndcg_at_20": 0.81229, "ndcg_at_100": 0.81856, "ndcg_at_1000": 0.82102, "map_at_1": 0.69, "map_at_3": 0.75333, "map_at_5": 0.76283, "map_at_10": 0.76617, "map_at_20": 0.76983, "map_at_100": 0.77096, "map_at_1000": 0.77104, "recall_at_1": 0.69, "recall_at_3": 0.84, "recall_at_5": 0.88, "recall_at_10": 0.9, "recall_at_20": 0.95, "recall_at_100": 0.98, "recall_at_1000": 1.0, "precision_at_1": 0.69, "precision_at_3": 0.28, "precision_at_5": 0.176, "precision_at_10": 0.09, "precision_at_20": 0.0475, "precision_at_100": 0.0098, "precision_at_1000": 0.001, "mrr_at_1": 0.69, "mrr_at_3": 0.7533333333333334, "mrr_at_5": 0.7628333333333335, "mrr_at_10": 0.7661666666666669, "mrr_at_20": 0.7698333333333335, "mrr_at_100": 0.7709625000000002, "mrr_at_1000": 0.7710412411292813, "naucs_at_1_max": 0.2543168068857264, "naucs_at_1_std": -0.33253964054535495, "naucs_at_1_diff1": 0.5989029401908389, "naucs_at_3_max": 0.38576951399116305, "naucs_at_3_std": -0.2952565046637219, "naucs_at_3_diff1": 0.6520004909180163, "naucs_at_5_max": 0.45434298440980025, "naucs_at_5_std": -0.38613585746102075, "naucs_at_5_diff1": 0.5503102131721279, "naucs_at_10_max": 0.48968253968254033, "naucs_at_10_std": -0.4380952380952347, "naucs_at_10_diff1": 0.472128851540616, "naucs_at_20_max": 0.47824463118580945, "naucs_at_20_std": -0.4903828197945701, "naucs_at_20_diff1": 0.4944911297852529, "naucs_at_100_max": 0.4960317460317565, "naucs_at_100_std": 0.795751633986937, "naucs_at_100_diff1": 0.0793650793650745, "naucs_at_1000_max": NaN, "naucs_at_1000_std": NaN, "naucs_at_1000_diff1": NaN}, "vidore/syntheticDocQA_artificial_intelligence_test": {"ndcg_at_1": 0.54, "ndcg_at_3": 0.65202, "ndcg_at_5": 0.66363, "ndcg_at_10": 0.68366, "ndcg_at_20": 0.69864, "ndcg_at_100": 0.71298, "ndcg_at_1000": 0.71964, "map_at_1": 0.54, "map_at_3": 0.625, "map_at_5": 0.631, "map_at_10": 0.63967, "map_at_20": 0.64367, "map_at_100": 0.64618, "map_at_1000": 0.64647, "recall_at_1": 0.54, "recall_at_3": 0.73, "recall_at_5": 0.76, "recall_at_10": 0.82, "recall_at_20": 0.88, "recall_at_100": 0.95, "recall_at_1000": 1.0, "precision_at_1": 0.54, "precision_at_3": 0.24333, "precision_at_5": 0.152, "precision_at_10": 0.082, "precision_at_20": 0.044, "precision_at_100": 0.0095, "precision_at_1000": 0.001, "mrr_at_1": 0.54, "mrr_at_3": 0.6250000000000001, "mrr_at_5": 0.6310000000000001, "mrr_at_10": 0.6396666666666668, "mrr_at_20": 0.6436720085470088, "mrr_at_100": 0.6461795757311586, "mrr_at_1000": 0.6464676959289798, "naucs_at_1_max": 0.5435386624433889, "naucs_at_1_std": -0.11297268761634857, "naucs_at_1_diff1": 0.6860881887138451, "naucs_at_3_max": 0.5216143427234654, "naucs_at_3_std": -0.2744697163301808, "naucs_at_3_diff1": 0.7614264090113817, "naucs_at_5_max": 0.5886475728995417, "naucs_at_5_std": -0.18462836376222094, "naucs_at_5_diff1": 0.7462360474171499, "naucs_at_10_max": 0.5302070712633551, "naucs_at_10_std": -0.09805595329881955, "naucs_at_10_diff1": 0.7058596761757898, "naucs_at_20_max": 0.6711740375437496, "naucs_at_20_std": 0.10034202990773257, "naucs_at_20_diff1": 0.7342109449570478, "naucs_at_100_max": 0.9183006535947744, "naucs_at_100_std": 0.42586367880485737, "naucs_at_100_diff1": 0.802987861811396, "naucs_at_1000_max": NaN, "naucs_at_1000_std": NaN, "naucs_at_1000_diff1": NaN}, "vidore/arxivqa_test_subsampled": {"ndcg_at_1": 0.24, "ndcg_at_3": 0.30243, "ndcg_at_5": 0.32291, "ndcg_at_10": 0.34458, "ndcg_at_20": 0.36564, "ndcg_at_100": 0.41267, "ndcg_at_1000": 0.43866, "map_at_1": 0.24, "map_at_3": 0.28733, "map_at_5": 0.29863, "map_at_10": 0.30738, "map_at_20": 0.31308, "map_at_100": 0.31938, "map_at_1000": 0.32049, "recall_at_1": 0.24, "recall_at_3": 0.346, "recall_at_5": 0.396, "recall_at_10": 0.464, "recall_at_20": 0.548, "recall_at_100": 0.804, "recall_at_1000": 1.0, "precision_at_1": 0.24, "precision_at_3": 0.11533, "precision_at_5": 0.0792, "precision_at_10": 0.0464, "precision_at_20": 0.0274, "precision_at_100": 0.00804, "precision_at_1000": 0.001, "mrr_at_1": 0.24, "mrr_at_3": 0.2873333333333334, "mrr_at_5": 0.2986333333333333, "mrr_at_10": 0.30737936507936503, "mrr_at_20": 0.3130766691169169, "mrr_at_100": 0.31938006231140553, "mrr_at_1000": 0.3204929422065829, "naucs_at_1_max": 0.1842709082195064, "naucs_at_1_std": -0.03645159357776179, "naucs_at_1_diff1": 0.5877096812844477, "naucs_at_3_max": 0.18220512508883954, "naucs_at_3_std": -0.011046497803118133, "naucs_at_3_diff1": 0.48188728397769015, "naucs_at_5_max": 0.1944250978935477, "naucs_at_5_std": 0.08311020311298906, "naucs_at_5_diff1": 0.45229398502834905, "naucs_at_10_max": 0.15675660027790675, "naucs_at_10_std": 0.12465788033180361, "naucs_at_10_diff1": 0.4077882226620071, "naucs_at_20_max": 0.1248395648494292, "naucs_at_20_std": 0.12644811989620058, "naucs_at_20_diff1": 0.34071101495900136, "naucs_at_100_max": 0.2800392923308714, "naucs_at_100_std": 0.33033042203627977, "naucs_at_100_diff1": 0.3726103951627693, "naucs_at_1000_max": 1.0, "naucs_at_1000_std": 1.0, "naucs_at_1000_diff1": 1.0}, "vidore/tatdqa_test": {"ndcg_at_1": 0.14642, "ndcg_at_3": 0.19314, "ndcg_at_5": 0.22608, "ndcg_at_10": 0.25688, "ndcg_at_20": 0.28734, "ndcg_at_50": 0.31632, "ndcg_at_100": 0.34115, "map_at_1": 0.14642, "map_at_3": 0.18084, "map_at_5": 0.19907, "map_at_10": 0.21199, "map_at_20": 0.22029, "map_at_50": 0.22491, "map_at_100": 0.22713, "recall_at_1": 0.14642, "recall_at_3": 0.22904, "recall_at_5": 0.30923, "recall_at_10": 0.4034, "recall_at_20": 0.5243, "recall_at_50": 0.67072, "recall_at_100": 0.82321, "precision_at_1": 0.14642, "precision_at_3": 0.07635, "precision_at_5": 0.06185, "precision_at_10": 0.04034, "precision_at_20": 0.02622, "precision_at_50": 0.01341, "precision_at_100": 0.00823, "mrr_at_1": 0.14580801944106925, "mrr_at_3": 0.18114621304171735, "mrr_at_5": 0.19861279870392848, "mrr_at_10": 0.21173359177611886, "mrr_at_20": 0.2199985145171474, "mrr_at_50": 0.22464583725829063, "mrr_at_100": 0.22686018141811548, "naucs_at_1_max": 0.05465634086423098, "naucs_at_1_std": 0.004519764499792802, "naucs_at_1_diff1": 0.4408572289894211, "naucs_at_3_max": 0.03306416876433599, "naucs_at_3_std": 0.024442247534575856, "naucs_at_3_diff1": 0.3530547071977037, "naucs_at_5_max": 0.0033715746535325484, "naucs_at_5_std": 0.0427181601755046, "naucs_at_5_diff1": 0.2758768564262304, "naucs_at_10_max": 0.0058969125587405085, "naucs_at_10_std": 0.04710603380219742, "naucs_at_10_diff1": 0.2590246128249172, "naucs_at_20_max": 0.012299252373198462, "naucs_at_20_std": 0.06068471433562855, "naucs_at_20_diff1": 0.2063849542535864, "naucs_at_50_max": -0.016468959955682726, "naucs_at_50_std": 0.044299149229442544, "naucs_at_50_diff1": 0.21611081513849306, "naucs_at_100_max": 0.03791446961632374, "naucs_at_100_std": 0.12027513970629726, "naucs_at_100_diff1": 0.1810068035226886}, "vidore/tabfquad_test_subsampled": {"ndcg_at_1": 0.30357, "ndcg_at_3": 0.37861, "ndcg_at_5": 0.41274, "ndcg_at_10": 0.4481, "ndcg_at_20": 0.48265, "ndcg_at_100": 0.52889, "ndcg_at_1000": 0.52889, "map_at_1": 0.30357, "map_at_3": 0.36012, "map_at_5": 0.37923, "map_at_10": 0.39424, "map_at_20": 0.40385, "map_at_100": 0.41059, "map_at_1000": 0.41059, "recall_at_1": 0.30357, "recall_at_3": 0.43214, "recall_at_5": 0.51429, "recall_at_10": 0.62143, "recall_at_20": 0.75714, "recall_at_100": 1.0, "recall_at_1000": 1.0, "precision_at_1": 0.30357, "precision_at_3": 0.14405, "precision_at_5": 0.10286, "precision_at_10": 0.06214, "precision_at_20": 0.03786, "precision_at_100": 0.01, "precision_at_1000": 0.001, "mrr_at_1": 0.30357142857142855, "mrr_at_3": 0.3601190476190476, "mrr_at_5": 0.3792261904761905, "mrr_at_10": 0.3942361111111112, "mrr_at_20": 0.40384985476524066, "mrr_at_100": 0.4105879371168134, "mrr_at_1000": 0.4105879371168134, "naucs_at_1_max": 0.289573448300778, "naucs_at_1_std": 0.279626195235916, "naucs_at_1_diff1": 0.5516655025410053, "naucs_at_3_max": 0.34146613553224203, "naucs_at_3_std": 0.37747574801173195, "naucs_at_3_diff1": 0.45571891399433845, "naucs_at_5_max": 0.32754790604234285, "naucs_at_5_std": 0.31138444598980075, "naucs_at_5_diff1": 0.4243451553082986, "naucs_at_10_max": 0.3712364081039097, "naucs_at_10_std": 0.3112392724548589, "naucs_at_10_diff1": 0.3579281929250532, "naucs_at_20_max": 0.34740120603630337, "naucs_at_20_std": 0.39231381431938545, "naucs_at_20_diff1": 0.2868162478190339, "naucs_at_100_max": 1.0, "naucs_at_100_std": 1.0, "naucs_at_100_diff1": 1.0, "naucs_at_1000_max": NaN, "naucs_at_1000_std": NaN, "naucs_at_1000_diff1": NaN}, "vidore/shiftproject_test": {"ndcg_at_1": 0.19, "ndcg_at_3": 0.27917, "ndcg_at_5": 0.308, "ndcg_at_10": 0.32476, "ndcg_at_20": 0.35448, "ndcg_at_100": 0.40367, "ndcg_at_1000": 0.42026, "map_at_1": 0.19, "map_at_3": 0.255, "map_at_5": 0.271, "map_at_10": 0.27826, "map_at_20": 0.28609, "map_at_100": 0.29186, "map_at_1000": 0.29248, "recall_at_1": 0.19, "recall_at_3": 0.35, "recall_at_5": 0.42, "recall_at_10": 0.47, "recall_at_20": 0.59, "recall_at_100": 0.87, "recall_at_1000": 1.0, "precision_at_1": 0.19, "precision_at_3": 0.11667, "precision_at_5": 0.084, "precision_at_10": 0.047, "precision_at_20": 0.0295, "precision_at_100": 0.0087, "precision_at_1000": 0.001, "mrr_at_1": 0.19, "mrr_at_3": 0.25499999999999995, "mrr_at_5": 0.27099999999999996, "mrr_at_10": 0.27826190476190477, "mrr_at_20": 0.2860934096433322, "mrr_at_100": 0.2918602124507187, "mrr_at_1000": 0.29247856510791964, "naucs_at_1_max": -0.051418290652901454, "naucs_at_1_std": -0.14327509962504, "naucs_at_1_diff1": 0.15958502177458675, "naucs_at_3_max": 0.10773287626079489, "naucs_at_3_std": -0.08202254598193706, "naucs_at_3_diff1": 0.12007383479464677, "naucs_at_5_max": 0.14732871918500656, "naucs_at_5_std": 0.0508852425020092, "naucs_at_5_diff1": 0.060670866060087626, "naucs_at_10_max": 0.09290734318754254, "naucs_at_10_std": 0.06761692532865714, "naucs_at_10_diff1": -0.03843668824747108, "naucs_at_20_max": 0.12417274421471261, "naucs_at_20_std": 0.11425301187139605, "naucs_at_20_diff1": -0.05562974158803701, "naucs_at_100_max": 0.43323705843145927, "naucs_at_100_std": 0.4437532400207375, "naucs_at_100_diff1": 0.2544249426053473, "naucs_at_1000_max": NaN, "naucs_at_1000_std": NaN, "naucs_at_1000_diff1": NaN}, "vidore/docvqa_test_subsampled": {"ndcg_at_1": 0.1663, "ndcg_at_3": 0.22986, "ndcg_at_5": 0.25276, "ndcg_at_10": 0.27128, "ndcg_at_20": 0.2926, "ndcg_at_100": 0.33536, "ndcg_at_1000": 0.37215, "map_at_1": 0.1663, "map_at_3": 0.2136, "map_at_5": 0.22635, "map_at_10": 0.23393, "map_at_20": 0.23978, "map_at_100": 0.24541, "map_at_1000": 0.24673, "recall_at_1": 0.1663, "recall_at_3": 0.27716, "recall_at_5": 0.33259, "recall_at_10": 0.39024, "recall_at_20": 0.4745, "recall_at_100": 0.70953, "recall_at_1000": 1.0, "precision_at_1": 0.1663, "precision_at_3": 0.09239, "precision_at_5": 0.06652, "precision_at_10": 0.03902, "precision_at_20": 0.02373, "precision_at_100": 0.0071, "precision_at_1000": 0.001, "mrr_at_1": 0.1662971175166297, "mrr_at_3": 0.21359940872135988, "mrr_at_5": 0.22634885439763483, "mrr_at_10": 0.23393429064160767, "mrr_at_20": 0.23978410642617462, "mrr_at_100": 0.2454077496817927, "mrr_at_1000": 0.2467252465575737, "naucs_at_1_max": 0.2536078131048019, "naucs_at_1_std": 0.13257812819720316, "naucs_at_1_diff1": 0.5435028496811409, "naucs_at_3_max": 0.23532108899301044, "naucs_at_3_std": 0.20145505331646668, "naucs_at_3_diff1": 0.4377698835403377, "naucs_at_5_max": 0.25515712073517766, "naucs_at_5_std": 0.21471436043744227, "naucs_at_5_diff1": 0.39579960154277694, "naucs_at_10_max": 0.21340796167348994, "naucs_at_10_std": 0.22476673494330973, "naucs_at_10_diff1": 0.35864044100130466, "naucs_at_20_max": 0.2234889963975852, "naucs_at_20_std": 0.22595058231333293, "naucs_at_20_diff1": 0.3121050241108213, "naucs_at_100_max": 0.12773242951419447, "naucs_at_100_std": 0.31384285667429795, "naucs_at_100_diff1": 0.2837577093483005, "naucs_at_1000_max": 1.0, "naucs_at_1000_std": 1.0, "naucs_at_1000_diff1": 1.0}, "vidore/syntheticDocQA_energy_test": {"ndcg_at_1": 0.61, "ndcg_at_3": 0.68917, "ndcg_at_5": 0.70938, "ndcg_at_10": 0.72782, "ndcg_at_20": 0.74319, "ndcg_at_100": 0.75456, "ndcg_at_1000": 0.75707, "map_at_1": 0.61, "map_at_3": 0.66833, "map_at_5": 0.67933, "map_at_10": 0.68637, "map_at_20": 0.6907, "map_at_100": 0.69229, "map_at_1000": 0.69237, "recall_at_1": 0.61, "recall_at_3": 0.75, "recall_at_5": 0.8, "recall_at_10": 0.86, "recall_at_20": 0.92, "recall_at_100": 0.98, "recall_at_1000": 1.0, "precision_at_1": 0.61, "precision_at_3": 0.25, "precision_at_5": 0.16, "precision_at_10": 0.086, "precision_at_20": 0.046, "precision_at_100": 0.0098, "precision_at_1000": 0.001, "mrr_at_1": 0.61, "mrr_at_3": 0.6683333333333334, "mrr_at_5": 0.6793333333333333, "mrr_at_10": 0.686373015873016, "mrr_at_20": 0.6906960996519822, "mrr_at_100": 0.6922906137236967, "mrr_at_1000": 0.6923723077306243, "naucs_at_1_max": 0.420965963168311, "naucs_at_1_std": -0.32356895744263015, "naucs_at_1_diff1": 0.7714713036121644, "naucs_at_3_max": 0.4598011512297232, "naucs_at_3_std": -0.49823129251700493, "naucs_at_3_diff1": 0.8043328100470961, "naucs_at_5_max": 0.40549450549450505, "naucs_at_5_std": -0.6961788211788206, "naucs_at_5_diff1": 0.7687812187812181, "naucs_at_10_max": 0.3752598032423449, "naucs_at_10_std": -0.6594499099348732, "naucs_at_10_diff1": 0.7617084661216568, "naucs_at_20_max": 0.31506769374416554, "naucs_at_20_std": -0.7252567693744081, "naucs_at_20_diff1": 0.8211367880485521, "naucs_at_100_max": 0.3384687208216665, "naucs_at_100_std": -0.3674136321195118, "naucs_at_100_diff1": 0.7770774976657274, "naucs_at_1000_max": NaN, "naucs_at_1000_std": NaN, "naucs_at_1000_diff1": NaN}, "vidore/infovqa_test_subsampled": {"ndcg_at_1": 0.56275, "ndcg_at_3": 0.65975, "ndcg_at_5": 0.68083, "ndcg_at_10": 0.7028, "ndcg_at_20": 0.71708, "ndcg_at_100": 0.72735, "ndcg_at_1000": 0.73232, "map_at_1": 0.56275, "map_at_3": 0.63529, "map_at_5": 0.64713, "map_at_10": 0.65603, "map_at_20": 0.65992, "map_at_100": 0.66139, "map_at_1000": 0.66158, "recall_at_1": 0.56275, "recall_at_3": 0.73077, "recall_at_5": 0.78138, "recall_at_10": 0.8502, "recall_at_20": 0.90688, "recall_at_100": 0.96154, "recall_at_1000": 1.0, "precision_at_1": 0.56275, "precision_at_3": 0.24359, "precision_at_5": 0.15628, "precision_at_10": 0.08502, "precision_at_20": 0.04534, "precision_at_100": 0.00962, "precision_at_1000": 0.001, "mrr_at_1": 0.562753036437247, "mrr_at_3": 0.635290148448043, "mrr_at_5": 0.6471322537112009, "mrr_at_10": 0.6560278902384165, "mrr_at_20": 0.6599243439794445, "mrr_at_100": 0.6613935703360762, "mrr_at_1000": 0.6615847636380331, "naucs_at_1_max": 0.4119424574346777, "naucs_at_1_std": -0.003630549834122387, "naucs_at_1_diff1": 0.7221216077070339, "naucs_at_3_max": 0.3436446846643777, "naucs_at_3_std": 0.0001523597023308914, "naucs_at_3_diff1": 0.6382772325308235, "naucs_at_5_max": 0.30095593564308837, "naucs_at_5_std": 0.03696604927578493, "naucs_at_5_diff1": 0.5987084174053594, "naucs_at_10_max": 0.3218669741560863, "naucs_at_10_std": 0.07687153161157063, "naucs_at_10_diff1": 0.5646660565556654, "naucs_at_20_max": 0.370091823338698, "naucs_at_20_std": 0.35865427792958265, "naucs_at_20_diff1": 0.5205824045210238, "naucs_at_100_max": 0.4312402149606384, "naucs_at_100_std": 0.6976206389014427, "naucs_at_100_diff1": 0.5286804363103885, "naucs_at_1000_max": 1.0, "naucs_at_1000_std": 1.0, "naucs_at_1000_diff1": 1.0}}