vidore benchmark results

results.json

@@ -1 +1 @@
-{"validation_set": {"ndcg_at_1": 0.716, "ndcg_at_3": 0.79, "ndcg_at_5": 0.80162, "ndcg_at_10": 0.81647, "ndcg_at_20": 0.82376, "ndcg_at_50": 0.8278, "ndcg_at_100": 0.8304, "map_at_1": 0.716, "map_at_3": 0.772, "map_at_5": 0.7785, "map_at_10": 0.78462, "map_at_20": 0.78673, "map_at_50": 0.78741, "map_at_100": 0.78764, "recall_at_1": 0.716, "recall_at_3": 0.842, "recall_at_5": 0.87, "recall_at_10": 0.916, "recall_at_20": 0.944, "recall_at_50": 0.964, "recall_at_100": 0.98, "precision_at_1": 0.716, "precision_at_3": 0.28067, "precision_at_5": 0.174, "precision_at_10": 0.0916, "precision_at_20": 0.0472, "precision_at_50": 0.01928, "precision_at_100": 0.0098, "mrr_at_1": 0.718, "mrr_at_3": 0.7746666666666666, "mrr_at_5": 0.7806666666666666, "mrr_at_10": 0.7868285714285714, "mrr_at_20": 0.7888180706118786, "mrr_at_50": 0.7895996953694904, "mrr_at_100": 0.7898277337976115, "naucs_at_1_max": 0.8001312345855925, "naucs_at_1_std": 0.11250348827562291, "naucs_at_1_diff1": 0.8830428322535393, "naucs_at_3_max": 0.9284596953407236, "naucs_at_3_std": 0.3631205849855362, "naucs_at_3_diff1": 0.8605631354827493, "naucs_at_5_max": 0.9320521365622464, "naucs_at_5_std": 0.4144190179960025, "naucs_at_5_diff1": 0.8710805006294897, "naucs_at_10_max": 0.9234138099684321, "naucs_at_10_std": 0.517418078342447, "naucs_at_10_diff1": 0.8689586945889484, "naucs_at_20_max": 0.9554821928771539, "naucs_at_20_std": 0.6497932506335875, "naucs_at_20_diff1": 0.921051754034949, "naucs_at_50_max": 0.9700435729847456, "naucs_at_50_std": 0.7666511048864019, "naucs_at_50_diff1": 0.9173928830791599, "naucs_at_100_max": 1.0, "naucs_at_100_std": 1.0, "naucs_at_100_diff1": 0.9460784313725422}, "syntheticDocQA_energy": {"ndcg_at_1": 0.75, "ndcg_at_3": 0.82678, "ndcg_at_5": 0.82678, "ndcg_at_10": 0.84683, "ndcg_at_20": 0.8521, "ndcg_at_50": 0.85785, "ndcg_at_100": 0.85785, "map_at_1": 0.75, "map_at_3": 0.80833, "map_at_5": 0.80833, "map_at_10": 0.81696, "map_at_20": 0.81851, "map_at_50": 0.81935, "map_at_100": 0.81935, "recall_at_1": 0.75, "recall_at_3": 0.88, "recall_at_5": 0.88, "recall_at_10": 0.94, "recall_at_20": 0.96, "recall_at_50": 0.99, "recall_at_100": 0.99, "precision_at_1": 0.75, "precision_at_3": 0.29333, "precision_at_5": 0.176, "precision_at_10": 0.094, "precision_at_20": 0.048, "precision_at_50": 0.0198, "precision_at_100": 0.0099, "mrr_at_1": 0.75, "mrr_at_3": 0.8083333333333332, "mrr_at_5": 0.8083333333333332, "mrr_at_10": 0.8167261904761905, "mrr_at_20": 0.8182738095238095, "mrr_at_50": 0.8191393489763056, "mrr_at_100": 0.8191393489763056, "naucs_at_1_max": 0.4352485609628468, "naucs_at_1_std": -0.3026268969126114, "naucs_at_1_diff1": 0.764772370486656, "naucs_at_3_max": 0.3731705377028315, "naucs_at_3_std": -0.43238943684377823, "naucs_at_3_diff1": 0.6174435252943048, "naucs_at_5_max": 0.37317053770283104, "naucs_at_5_std": -0.4323894368437778, "naucs_at_5_diff1": 0.6174435252943049, "naucs_at_10_max": 0.6125894802365398, "naucs_at_10_std": -0.20728291316526842, "naucs_at_10_diff1": 0.8031434796140697, "naucs_at_20_max": 0.5303454715219444, "naucs_at_20_std": -0.406629318394016, "naucs_at_20_diff1": 0.7047152194211033, "naucs_at_50_max": 0.8692810457516374, "naucs_at_50_std": 0.554154995331464, "naucs_at_50_diff1": 0.8692810457516374, "naucs_at_100_max": 0.8692810457516374, "naucs_at_100_std": 0.554154995331464, "naucs_at_100_diff1": 0.8692810457516374}, "syntheticDocQA_healthcare_industry": {"ndcg_at_1": 0.8, "ndcg_at_3": 0.87678, "ndcg_at_5": 0.89788, "ndcg_at_10": 0.90445, "ndcg_at_20": 0.90445, "ndcg_at_50": 0.90445, "ndcg_at_100": 0.90445, "map_at_1": 0.8, "map_at_3": 0.85833, "map_at_5": 0.87033, "map_at_10": 0.87311, "map_at_20": 0.87311, "map_at_50": 0.87311, "map_at_100": 0.87311, "recall_at_1": 0.8, "recall_at_3": 0.93, "recall_at_5": 0.98, "recall_at_10": 1.0, "recall_at_20": 1.0, "recall_at_50": 1.0, "recall_at_100": 1.0, "precision_at_1": 0.8, "precision_at_3": 0.31, "precision_at_5": 0.196, "precision_at_10": 0.1, "precision_at_20": 0.05, "precision_at_50": 0.02, "precision_at_100": 0.01, "mrr_at_1": 0.8, "mrr_at_3": 0.8583333333333333, "mrr_at_5": 0.8703333333333333, "mrr_at_10": 0.8732500000000001, "mrr_at_20": 0.8732500000000001, "mrr_at_50": 0.8732500000000001, "mrr_at_100": 0.8732500000000001, "naucs_at_1_max": 0.32232767232767134, "naucs_at_1_std": -0.22445054945054957, "naucs_at_1_diff1": 0.6103646353646349, "naucs_at_3_max": 0.49893290649593114, "naucs_at_3_std": -0.6459250366813389, "naucs_at_3_diff1": 0.6492597038815514, "naucs_at_5_max": 0.6790382819794609, "naucs_at_5_std": -0.9556489262371534, "naucs_at_5_diff1": 0.861111111111116, "naucs_at_10_max": 1.0, "naucs_at_10_std": 1.0, "naucs_at_10_diff1": 1.0, "naucs_at_20_max": 1.0, "naucs_at_20_std": 1.0, "naucs_at_20_diff1": 1.0, "naucs_at_50_max": NaN, "naucs_at_50_std": NaN, "naucs_at_50_diff1": NaN, "naucs_at_100_max": NaN, "naucs_at_100_std": NaN, "naucs_at_100_diff1": NaN}, "syntheticDocQA_artificial_intelligence_test": {"ndcg_at_1": 0.71, "ndcg_at_3": 0.81071, "ndcg_at_5": 0.82276, "ndcg_at_10": 0.83243, "ndcg_at_20": 0.84238, "ndcg_at_50": 0.84624, "ndcg_at_100": 0.84624, "map_at_1": 0.71, "map_at_3": 0.78667, "map_at_5": 0.79317, "map_at_10": 0.79713, "map_at_20": 0.79978, "map_at_50": 0.80035, "map_at_100": 0.80035, "recall_at_1": 0.71, "recall_at_3": 0.88, "recall_at_5": 0.91, "recall_at_10": 0.94, "recall_at_20": 0.98, "recall_at_50": 1.0, "recall_at_100": 1.0, "precision_at_1": 0.71, "precision_at_3": 0.29333, "precision_at_5": 0.182, "precision_at_10": 0.094, "precision_at_20": 0.049, "precision_at_50": 0.02, "precision_at_100": 0.01, "mrr_at_1": 0.71, "mrr_at_3": 0.7866666666666667, "mrr_at_5": 0.7931666666666666, "mrr_at_10": 0.7971349206349206, "mrr_at_20": 0.7998128023936847, "mrr_at_50": 0.8003853830388461, "mrr_at_100": 0.8003853830388461, "naucs_at_1_max": 0.4346867297963423, "naucs_at_1_std": -0.26896206660118943, "naucs_at_1_diff1": 0.8395567333172729, "naucs_at_3_max": 0.5740136811963086, "naucs_at_3_std": -0.24773305758829017, "naucs_at_3_diff1": 0.7142857142857145, "naucs_at_5_max": 0.6865338728083817, "naucs_at_5_std": -0.06338831829027824, "naucs_at_5_diff1": 0.7004876024483871, "naucs_at_10_max": 0.8712262682850935, "naucs_at_10_std": 0.24198568316215255, "naucs_at_10_diff1": 0.6896202925614692, "naucs_at_20_max": 0.9346405228758136, "naucs_at_20_std": 0.5613912231559286, "naucs_at_20_diff1": 0.861111111111116, "naucs_at_50_max": NaN, "naucs_at_50_std": NaN, "naucs_at_50_diff1": NaN, "naucs_at_100_max": NaN, "naucs_at_100_std": NaN, "naucs_at_100_diff1": NaN}, "syntheticDocQA_government_reports": {"ndcg_at_1": 0.7, "ndcg_at_3": 0.80202, "ndcg_at_5": 0.83042, "ndcg_at_10": 0.84047, "ndcg_at_20": 0.84812, "ndcg_at_50": 0.84812, "ndcg_at_100": 0.84812, "map_at_1": 0.7, "map_at_3": 0.77833, "map_at_5": 0.79383, "map_at_10": 0.79818, "map_at_20": 0.80031, "map_at_50": 0.80031, "map_at_100": 0.80031, "recall_at_1": 0.7, "recall_at_3": 0.87, "recall_at_5": 0.94, "recall_at_10": 0.97, "recall_at_20": 1.0, "recall_at_50": 1.0, "recall_at_100": 1.0, "precision_at_1": 0.7, "precision_at_3": 0.29, "precision_at_5": 0.188, "precision_at_10": 0.097, "precision_at_20": 0.05, "precision_at_50": 0.02, "precision_at_100": 0.01, "mrr_at_1": 0.7, "mrr_at_3": 0.7783333333333333, "mrr_at_5": 0.7938333333333334, "mrr_at_10": 0.7981785714285714, "mrr_at_20": 0.8003052682611507, "mrr_at_50": 0.8003052682611507, "mrr_at_100": 0.8003052682611507, "naucs_at_1_max": 0.34667627454512717, "naucs_at_1_std": -0.18418303008467002, "naucs_at_1_diff1": 0.7120158529994597, "naucs_at_3_max": 0.4705991261201206, "naucs_at_3_std": -0.35103310375472296, "naucs_at_3_diff1": 0.44619714137598854, "naucs_at_5_max": 0.6545284780578933, "naucs_at_5_std": -0.7740429505135397, "naucs_at_5_diff1": 0.6678338001867429, "naucs_at_10_max": 0.5938375350140046, "naucs_at_10_std": -0.7231559290382794, "naucs_at_10_diff1": 0.6204481792717093, "naucs_at_20_max": 1.0, "naucs_at_20_std": 1.0, "naucs_at_20_diff1": 1.0, "naucs_at_50_max": NaN, "naucs_at_50_std": NaN, "naucs_at_50_diff1": NaN, "naucs_at_100_max": NaN, "naucs_at_100_std": NaN, "naucs_at_100_diff1": NaN}, "infovqa_subsampled": {"ndcg_at_1": 0.692, "ndcg_at_3": 0.76431, "ndcg_at_5": 0.78548, "ndcg_at_10": 0.80119, "ndcg_at_20": 0.80674, "ndcg_at_50": 0.81252, "ndcg_at_100": 0.81415, "map_at_1": 0.692, "map_at_3": 0.74767, "map_at_5": 0.75927, "map_at_10": 0.76587, "map_at_20": 0.76738, "map_at_50": 0.7684, "map_at_100": 0.76855, "recall_at_1": 0.692, "recall_at_3": 0.812, "recall_at_5": 0.864, "recall_at_10": 0.912, "recall_at_20": 0.934, "recall_at_50": 0.962, "recall_at_100": 0.972, "precision_at_1": 0.692, "precision_at_3": 0.27067, "precision_at_5": 0.1728, "precision_at_10": 0.0912, "precision_at_20": 0.0467, "precision_at_50": 0.01924, "precision_at_100": 0.00972, "mrr_at_1": 0.69, "mrr_at_3": 0.7473333333333334, "mrr_at_5": 0.7589333333333332, "mrr_at_10": 0.7651642857142857, "mrr_at_20": 0.7666584847299244, "mrr_at_50": 0.7676865192114355, "mrr_at_100": 0.7678320500103497, "naucs_at_1_max": 0.3834175233264141, "naucs_at_1_std": -0.27314752495589745, "naucs_at_1_diff1": 0.8045292563177913, "naucs_at_3_max": 0.577343039126478, "naucs_at_3_std": -0.008935184205513267, "naucs_at_3_diff1": 0.7278181010220708, "naucs_at_5_max": 0.675093872674517, "naucs_at_5_std": 0.18129515844569502, "naucs_at_5_diff1": 0.7001977015417865, "naucs_at_10_max": 0.8044414735591211, "naucs_at_10_std": 0.42251294457176947, "naucs_at_10_diff1": 0.7054579407520588, "naucs_at_20_max": 0.8796932914579985, "naucs_at_20_std": 0.6807854454913284, "naucs_at_20_diff1": 0.7039583510171761, "naucs_at_50_max": 0.9501203990368081, "naucs_at_50_std": 0.857265713302859, "naucs_at_50_diff1": 0.7446311858076538, "naucs_at_100_max": 0.9416433239962615, "naucs_at_100_std": 0.8673135921035019, "naucs_at_100_diff1": 0.6826063758836863}, "docvqa_subsampled": {"ndcg_at_1": 0.344, "ndcg_at_3": 0.41347, "ndcg_at_5": 0.44369, "ndcg_at_10": 0.46201, "ndcg_at_20": 0.4759, "ndcg_at_50": 0.4878, "ndcg_at_100": 0.5002, "map_at_1": 0.344, "map_at_3": 0.39667, "map_at_5": 0.41377, "map_at_10": 0.42144, "map_at_20": 0.42512, "map_at_50": 0.42702, "map_at_100": 0.42814, "recall_at_1": 0.344, "recall_at_3": 0.462, "recall_at_5": 0.534, "recall_at_10": 0.59, "recall_at_20": 0.646, "recall_at_50": 0.706, "recall_at_100": 0.782, "precision_at_1": 0.344, "precision_at_3": 0.154, "precision_at_5": 0.1068, "precision_at_10": 0.059, "precision_at_20": 0.0323, "precision_at_50": 0.01412, "precision_at_100": 0.00782, "mrr_at_1": 0.346, "mrr_at_3": 0.398, "mrr_at_5": 0.4155, "mrr_at_10": 0.42276111111111114, "mrr_at_20": 0.426310462043976, "mrr_at_50": 0.42830398638968425, "mrr_at_100": 0.4294428772908946, "naucs_at_1_max": 0.7371100485555001, "naucs_at_1_std": 0.33261575567803164, "naucs_at_1_diff1": 0.8130375220827029, "naucs_at_3_max": 0.8020023826617959, "naucs_at_3_std": 0.49953582802757135, "naucs_at_3_diff1": 0.7334337943494619, "naucs_at_5_max": 0.8117119579471634, "naucs_at_5_std": 0.5755264196835543, "naucs_at_5_diff1": 0.6923091789040089, "naucs_at_10_max": 0.8227097304356757, "naucs_at_10_std": 0.613846537631884, "naucs_at_10_diff1": 0.7031446725461138, "naucs_at_20_max": 0.8705519074674019, "naucs_at_20_std": 0.7222923846653254, "naucs_at_20_diff1": 0.7266076440815379, "naucs_at_50_max": 0.8427827590881949, "naucs_at_50_std": 0.7425625391614057, "naucs_at_50_diff1": 0.6988519982851428, "naucs_at_100_max": 0.8260205582772512, "naucs_at_100_std": 0.7179798527014848, "naucs_at_100_diff1": 0.683312862233389}, "arxivqa_subsampled": {"ndcg_at_1": 0.714, "ndcg_at_3": 0.77369, "ndcg_at_5": 0.79013, "ndcg_at_10": 0.79976, "ndcg_at_20": 0.8069, "ndcg_at_50": 0.8186, "ndcg_at_100": 0.82249, "map_at_1": 0.714, "map_at_3": 0.75967, "map_at_5": 0.76877, "map_at_10": 0.7727, "map_at_20": 0.77469, "map_at_50": 0.77682, "map_at_100": 0.77716, "recall_at_1": 0.714, "recall_at_3": 0.814, "recall_at_5": 0.854, "recall_at_10": 0.884, "recall_at_20": 0.912, "recall_at_50": 0.968, "recall_at_100": 0.992, "precision_at_1": 0.714, "precision_at_3": 0.27133, "precision_at_5": 0.1708, "precision_at_10": 0.0884, "precision_at_20": 0.0456, "precision_at_50": 0.01936, "precision_at_100": 0.00992, "mrr_at_1": 0.712, "mrr_at_3": 0.7586666666666666, "mrr_at_5": 0.7673666666666666, "mrr_at_10": 0.771681746031746, "mrr_at_20": 0.7736755779308411, "mrr_at_50": 0.7757993229634467, "mrr_at_100": 0.7761433854912912, "naucs_at_1_max": 0.7317309098271223, "naucs_at_1_std": -0.37215292384176624, "naucs_at_1_diff1": 0.8526437738310815, "naucs_at_3_max": 0.8049699197503837, "naucs_at_3_std": -0.23096075141852557, "naucs_at_3_diff1": 0.8225158415524194, "naucs_at_5_max": 0.8093326040846809, "naucs_at_5_std": -0.24028520736897105, "naucs_at_5_diff1": 0.7992676387017601, "naucs_at_10_max": 0.8183301091839064, "naucs_at_10_std": -0.2813370930194455, "naucs_at_10_diff1": 0.7861733171579385, "naucs_at_20_max": 0.8187866055513119, "naucs_at_20_std": -0.1553348612172151, "naucs_at_20_diff1": 0.7767910194380784, "naucs_at_50_max": 0.873511904761906, "naucs_at_50_std": 0.2547852474323104, "naucs_at_50_diff1": 0.7219887955182044, "naucs_at_100_max": 1.0, "naucs_at_100_std": 0.9673202614378978, "naucs_at_100_diff1": 0.8978758169934562}, "tabfquad_subsampled": {"ndcg_at_1": 0.74286, "ndcg_at_3": 0.79235, "ndcg_at_5": 0.81125, "ndcg_at_10": 0.82042, "ndcg_at_20": 0.83173, "ndcg_at_50": 0.83813, "ndcg_at_100": 0.84176, "map_at_1": 0.74286, "map_at_3": 0.78095, "map_at_5": 0.79131, "map_at_10": 0.79507, "map_at_20": 0.79795, "map_at_50": 0.79898, "map_at_100": 0.79935, "recall_at_1": 0.74286, "recall_at_3": 0.825, "recall_at_5": 0.87143, "recall_at_10": 0.9, "recall_at_20": 0.94643, "recall_at_50": 0.97857, "recall_at_100": 1.0, "precision_at_1": 0.74286, "precision_at_3": 0.275, "precision_at_5": 0.17429, "precision_at_10": 0.09, "precision_at_20": 0.04732, "precision_at_50": 0.01957, "precision_at_100": 0.01, "mrr_at_1": 0.7428571428571429, "mrr_at_3": 0.7803571428571429, "mrr_at_5": 0.7916071428571428, "mrr_at_10": 0.7947760770975056, "mrr_at_20": 0.7976513639285919, "mrr_at_50": 0.7986905677649363, "mrr_at_100": 0.799055515734078, "naucs_at_1_max": 0.7500073065232634, "naucs_at_1_std": 0.2514466916062664, "naucs_at_1_diff1": 0.8531827215337847, "naucs_at_3_max": 0.8343426391328044, "naucs_at_3_std": 0.4195381416661603, "naucs_at_3_diff1": 0.8111084196289781, "naucs_at_5_max": 0.8508788332086781, "naucs_at_5_std": 0.5882973608291492, "naucs_at_5_diff1": 0.8132145528368421, "naucs_at_10_max": 0.8186774709883973, "naucs_at_10_std": 0.5969054288382045, "naucs_at_10_diff1": 0.7904995331465913, "naucs_at_20_max": 0.8555866791160895, "naucs_at_20_std": 0.659228135698729, "naucs_at_20_diff1": 0.7743541861188911, "naucs_at_50_max": 0.7786336756925044, "naucs_at_50_std": 0.438297541238722, "naucs_at_50_diff1": 0.6545284780578803, "naucs_at_100_max": 1.0, "naucs_at_100_std": 1.0, "naucs_at_100_diff1": 1.0}, "tatdqa": {"ndcg_at_1": 0.44197, "ndcg_at_3": 0.55566, "ndcg_at_5": 0.58437, "ndcg_at_10": 0.62411, "ndcg_at_20": 0.64062, "ndcg_at_50": 0.65097, "ndcg_at_100": 0.65588, "map_at_1": 0.44197, "map_at_3": 0.52726, "map_at_5": 0.54304, "map_at_10": 0.55982, "map_at_20": 0.56448, "map_at_50": 0.56618, "map_at_100": 0.56662, "recall_at_1": 0.44197, "recall_at_3": 0.638, "recall_at_5": 0.70836, "recall_at_10": 0.82922, "recall_at_20": 0.89357, "recall_at_50": 0.94528, "recall_at_100": 0.97535, "precision_at_1": 0.44197, "precision_at_3": 0.21267, "precision_at_5": 0.14167, "precision_at_10": 0.08292, "precision_at_20": 0.04468, "precision_at_50": 0.01891, "precision_at_100": 0.00975, "mrr_at_1": 0.4395670475045099, "mrr_at_3": 0.525756664662257, "mrr_at_5": 0.5415714572058529, "mrr_at_10": 0.5584664834063511, "mrr_at_20": 0.5630899072695948, "mrr_at_50": 0.5647839013198243, "mrr_at_100": 0.5652361863084263, "naucs_at_1_max": 0.38576514235208315, "naucs_at_1_std": -0.08015654973630132, "naucs_at_1_diff1": 0.603618573280768, "naucs_at_3_max": 0.43656957234090016, "naucs_at_3_std": -0.021403883055059796, "naucs_at_3_diff1": 0.47048780566558884, "naucs_at_5_max": 0.4336016828393664, "naucs_at_5_std": -0.0005891945007471499, "naucs_at_5_diff1": 0.4256629282001986, "naucs_at_10_max": 0.5493338687113005, "naucs_at_10_std": 0.16083596394909894, "naucs_at_10_diff1": 0.38247082357061657, "naucs_at_20_max": 0.5584090277396198, "naucs_at_20_std": 0.2602657817525008, "naucs_at_20_diff1": 0.3980322374475721, "naucs_at_50_max": 0.633233633128742, "naucs_at_50_std": 0.4310547495591043, "naucs_at_50_diff1": 0.38621143845546824, "naucs_at_100_max": 0.8349677169227036, "naucs_at_100_std": 0.7325722103484296, "naucs_at_100_diff1": 0.4877781738863079}, "shift_project": {"ndcg_at_1": 0.52, "ndcg_at_3": 0.63702, "ndcg_at_5": 0.6779, "ndcg_at_10": 0.7029, "ndcg_at_20": 0.71283, "ndcg_at_50": 0.72072, "ndcg_at_100": 0.72246, "map_at_1": 0.52, "map_at_3": 0.60833, "map_at_5": 0.63083, "map_at_10": 0.64065, "map_at_20": 0.64327, "map_at_50": 0.64451, "map_at_100": 0.6447, "recall_at_1": 0.52, "recall_at_3": 0.72, "recall_at_5": 0.82, "recall_at_10": 0.9, "recall_at_20": 0.94, "recall_at_50": 0.98, "recall_at_100": 0.99, "precision_at_1": 0.52, "precision_at_3": 0.24, "precision_at_5": 0.164, "precision_at_10": 0.09, "precision_at_20": 0.047, "precision_at_50": 0.0196, "precision_at_100": 0.0099, "mrr_at_1": 0.55, "mrr_at_3": 0.6216666666666666, "mrr_at_5": 0.6451666666666667, "mrr_at_10": 0.6549841269841269, "mrr_at_20": 0.6576700244200243, "mrr_at_50": 0.6589199025112817, "mrr_at_100": 0.6591122102035896, "naucs_at_1_max": 0.05023396267479185, "naucs_at_1_std": -0.2630855868110345, "naucs_at_1_diff1": 0.5195467653299963, "naucs_at_3_max": 0.2670177045177047, "naucs_at_3_std": -0.04368894993894898, "naucs_at_3_diff1": 0.3609394078144081, "naucs_at_5_max": 0.5939530785328778, "naucs_at_5_std": 0.2729926203326362, "naucs_at_5_diff1": 0.2156625178984482, "naucs_at_10_max": 0.6814192343604119, "naucs_at_10_std": 0.3050420168067224, "naucs_at_10_diff1": 0.08445378151260835, "naucs_at_20_max": 0.7568471833177721, "naucs_at_20_std": 0.39285714285714246, "naucs_at_20_diff1": 0.20028011204481636, "naucs_at_50_max": 0.5401493930905731, "naucs_at_50_std": -0.17133520074695377, "naucs_at_50_diff1": 0.4960317460317565, "naucs_at_100_max": 0.35807656395892007, "naucs_at_100_std": -0.17133520074697067, "naucs_at_100_diff1": 0.12278244631185525}}
+{"../colpali/data_dir/eval_vidore/tatdqa_test": {"ndcg_at_1": 0.45504, "ndcg_at_3": 0.56872, "ndcg_at_5": 0.59711, "ndcg_at_10": 0.6331, "ndcg_at_20": 0.65064, "ndcg_at_50": 0.66009, "ndcg_at_100": 0.66538, "map_at_1": 0.45504, "map_at_3": 0.53989, "map_at_5": 0.55557, "map_at_10": 0.57078, "map_at_20": 0.57577, "map_at_50": 0.57736, "map_at_100": 0.57784, "recall_at_1": 0.45504, "recall_at_3": 0.65249, "recall_at_5": 0.72175, "recall_at_10": 0.83111, "recall_at_20": 0.89915, "recall_at_50": 0.94593, "recall_at_100": 0.97813, "precision_at_1": 0.45504, "precision_at_3": 0.2175, "precision_at_5": 0.14435, "precision_at_10": 0.08311, "precision_at_20": 0.04496, "precision_at_50": 0.01892, "precision_at_100": 0.00978, "mrr_at_1": 0.4526123936816525, "mrr_at_3": 0.5387808829485622, "mrr_at_5": 0.5541211016605914, "mrr_at_10": 0.5695155644274721, "mrr_at_20": 0.5744895949632595, "mrr_at_50": 0.5760732761650542, "mrr_at_100": 0.5765591764798109, "naucs_at_1_max": 0.4007616357070541, "naucs_at_1_std": -0.08607992151091595, "naucs_at_1_diff1": 0.6060499097590993, "naucs_at_3_max": 0.44870095096796053, "naucs_at_3_std": -0.008186482145121039, "naucs_at_3_diff1": 0.4878097052185952, "naucs_at_5_max": 0.43459142011431634, "naucs_at_5_std": -0.009447458561375376, "naucs_at_5_diff1": 0.4384749463893095, "naucs_at_10_max": 0.5620075425998711, "naucs_at_10_std": 0.16203801191956155, "naucs_at_10_diff1": 0.4322829810150347, "naucs_at_20_max": 0.5401898328924477, "naucs_at_20_std": 0.23026218741014237, "naucs_at_20_diff1": 0.4155676018698351, "naucs_at_50_max": 0.5491274432650216, "naucs_at_50_std": 0.30248595340618817, "naucs_at_50_diff1": 0.39369573233613886, "naucs_at_100_max": 0.7837692460321041, "naucs_at_100_std": 0.6807203771751222, "naucs_at_100_diff1": 0.3957565014112084}, "../colpali/data_dir/eval_vidore/shiftproject_test": {"ndcg_at_1": 0.52, "ndcg_at_3": 0.64833, "ndcg_at_5": 0.68446, "ndcg_at_10": 0.71008, "ndcg_at_20": 0.71264, "ndcg_at_50": 0.72455, "ndcg_at_100": 0.72624, "map_at_1": 0.52, "map_at_3": 0.61667, "map_at_5": 0.63617, "map_at_10": 0.64658, "map_at_20": 0.6473, "map_at_50": 0.64923, "map_at_100": 0.64939, "recall_at_1": 0.52, "recall_at_3": 0.74, "recall_at_5": 0.83, "recall_at_10": 0.91, "recall_at_20": 0.92, "recall_at_50": 0.98, "recall_at_100": 0.99, "precision_at_1": 0.52, "precision_at_3": 0.24667, "precision_at_5": 0.166, "precision_at_10": 0.091, "precision_at_20": 0.046, "precision_at_50": 0.0196, "precision_at_100": 0.0099, "mrr_at_1": 0.52, "mrr_at_3": 0.62, "mrr_at_5": 0.64, "mrr_at_10": 0.6504166666666668, "mrr_at_20": 0.6511309523809524, "mrr_at_50": 0.6530746753246752, "mrr_at_100": 0.6532413419913419, "naucs_at_1_max": 0.011072419609336812, "naucs_at_1_std": -0.28150334621034895, "naucs_at_1_diff1": 0.5199684422438654, "naucs_at_3_max": 0.30832083046105246, "naucs_at_3_std": -0.017132314180284387, "naucs_at_3_diff1": 0.3817768946920235, "naucs_at_5_max": 0.619753874731526, "naucs_at_5_std": 0.3159865327683294, "naucs_at_5_diff1": 0.32579671445985925, "naucs_at_10_max": 0.7622159975101153, "naucs_at_10_std": 0.4451187882560422, "naucs_at_10_diff1": 0.13990040460628678, "naucs_at_20_max": 0.7488328664799268, "naucs_at_20_std": 0.48541083099906845, "naucs_at_20_diff1": 0.03238795518207437, "naucs_at_50_max": 0.561391223155939, "naucs_at_50_std": 0.2754435107376223, "naucs_at_50_diff1": 0.4960317460317565, "naucs_at_100_max": 0.12278244631185525, "naucs_at_100_std": -0.17133520074697067, "naucs_at_100_diff1": 0.12278244631185525}, "../colpali/data_dir/eval_vidore/syntheticDocQA_artificial_intelligence_test": {"ndcg_at_1": 0.75, "ndcg_at_3": 0.82678, "ndcg_at_5": 0.84226, "ndcg_at_10": 0.84915, "ndcg_at_20": 0.85956, "ndcg_at_50": 0.86333, "ndcg_at_100": 0.86333, "map_at_1": 0.75, "map_at_3": 0.80833, "map_at_5": 0.81633, "map_at_10": 0.81943, "map_at_20": 0.82243, "map_at_50": 0.82296, "map_at_100": 0.82296, "recall_at_1": 0.75, "recall_at_3": 0.88, "recall_at_5": 0.92, "recall_at_10": 0.94, "recall_at_20": 0.98, "recall_at_50": 1.0, "recall_at_100": 1.0, "precision_at_1": 0.75, "precision_at_3": 0.29333, "precision_at_5": 0.184, "precision_at_10": 0.094, "precision_at_20": 0.049, "precision_at_50": 0.02, "precision_at_100": 0.01, "mrr_at_1": 0.75, "mrr_at_3": 0.8083333333333332, "mrr_at_5": 0.8163333333333334, "mrr_at_10": 0.8194285714285714, "mrr_at_20": 0.8224285714285715, "mrr_at_50": 0.8229552337063858, "mrr_at_100": 0.8229552337063858, "naucs_at_1_max": 0.40276295133437956, "naucs_at_1_std": -0.22528519099947694, "naucs_at_1_diff1": 0.8150915750915747, "naucs_at_3_max": 0.6308065542475333, "naucs_at_3_std": -0.2201718103722543, "naucs_at_3_diff1": 0.7729080496341069, "naucs_at_5_max": 0.701272175536881, "naucs_at_5_std": -0.12114845938375016, "naucs_at_5_diff1": 0.7619047619047611, "naucs_at_10_max": 0.9039060068471841, "naucs_at_10_std": 0.3204170557111718, "naucs_at_10_diff1": 0.7288359788359801, "naucs_at_20_max": 0.9346405228758136, "naucs_at_20_std": 0.6790382819794609, "naucs_at_20_diff1": 0.861111111111116, "naucs_at_50_max": NaN, "naucs_at_50_std": NaN, "naucs_at_50_diff1": NaN, "naucs_at_100_max": NaN, "naucs_at_100_std": NaN, "naucs_at_100_diff1": NaN}, "../colpali/data_dir/eval_vidore/syntheticDocQA_government_reports_test": {"ndcg_at_1": 0.73, "ndcg_at_3": 0.80678, "ndcg_at_5": 0.83175, "ndcg_at_10": 0.85149, "ndcg_at_20": 0.85647, "ndcg_at_50": 0.85647, "ndcg_at_100": 0.85647, "map_at_1": 0.73, "map_at_3": 0.78833, "map_at_5": 0.80233, "map_at_10": 0.81067, "map_at_20": 0.812, "map_at_50": 0.812, "map_at_100": 0.812, "recall_at_1": 0.73, "recall_at_3": 0.86, "recall_at_5": 0.92, "recall_at_10": 0.98, "recall_at_20": 1.0, "recall_at_50": 1.0, "recall_at_100": 1.0, "precision_at_1": 0.73, "precision_at_3": 0.28667, "precision_at_5": 0.184, "precision_at_10": 0.098, "precision_at_20": 0.05, "precision_at_50": 0.02, "precision_at_100": 0.01, "mrr_at_1": 0.73, "mrr_at_3": 0.79, "mrr_at_5": 0.8059999999999999, "mrr_at_10": 0.8126666666666666, "mrr_at_20": 0.8140000000000001, "mrr_at_50": 0.8140000000000001, "mrr_at_100": 0.8140000000000001, "naucs_at_1_max": 0.43661168884782514, "naucs_at_1_std": -0.10324559161768487, "naucs_at_1_diff1": 0.7004069275983408, "naucs_at_3_max": 0.5075169738118337, "naucs_at_3_std": -0.201711237356242, "naucs_at_3_diff1": 0.564361923236802, "naucs_at_5_max": 0.7782446311858078, "naucs_at_5_std": -0.36525443510737254, "naucs_at_5_diff1": 0.7284080298786184, "naucs_at_10_max": 0.7770774976657324, "naucs_at_10_std": -0.21475256769374024, "naucs_at_10_diff1": 0.27544351073763346, "naucs_at_20_max": 1.0, "naucs_at_20_std": 1.0, "naucs_at_20_diff1": 1.0, "naucs_at_50_max": NaN, "naucs_at_50_std": NaN, "naucs_at_50_diff1": NaN, "naucs_at_100_max": NaN, "naucs_at_100_std": NaN, "naucs_at_100_diff1": NaN}, "../colpali/data_dir/eval_vidore/arxivqa_test_subsampled": {"ndcg_at_1": 0.71, "ndcg_at_3": 0.76995, "ndcg_at_5": 0.78484, "ndcg_at_10": 0.79735, "ndcg_at_20": 0.80482, "ndcg_at_50": 0.81639, "ndcg_at_100": 0.81931, "map_at_1": 0.71, "map_at_3": 0.756, "map_at_5": 0.7643, "map_at_10": 0.76959, "map_at_20": 0.77158, "map_at_50": 0.77347, "map_at_100": 0.77372, "recall_at_1": 0.71, "recall_at_3": 0.81, "recall_at_5": 0.846, "recall_at_10": 0.884, "recall_at_20": 0.914, "recall_at_50": 0.972, "recall_at_100": 0.99, "precision_at_1": 0.71, "precision_at_3": 0.27, "precision_at_5": 0.1692, "precision_at_10": 0.0884, "precision_at_20": 0.0457, "precision_at_50": 0.01944, "precision_at_100": 0.0099, "mrr_at_1": 0.71, "mrr_at_3": 0.7573333333333334, "mrr_at_5": 0.7650333333333333, "mrr_at_10": 0.7700404761904762, "mrr_at_20": 0.7720276221714303, "mrr_at_50": 0.7738298194710532, "mrr_at_100": 0.7741667087707991, "naucs_at_1_max": 0.7299668291236584, "naucs_at_1_std": -0.3983062468728575, "naucs_at_1_diff1": 0.8380816485369605, "naucs_at_3_max": 0.7857509560479856, "naucs_at_3_std": -0.3172507726963166, "naucs_at_3_diff1": 0.7965215569175964, "naucs_at_5_max": 0.7825180779544266, "naucs_at_5_std": -0.35381320929759713, "naucs_at_5_diff1": 0.7722114199295939, "naucs_at_10_max": 0.7812469261287255, "naucs_at_10_std": -0.36191350536083366, "naucs_at_10_diff1": 0.7579428833732241, "naucs_at_20_max": 0.7745640892015706, "naucs_at_20_std": -0.35226803899854764, "naucs_at_20_diff1": 0.7307884394067706, "naucs_at_50_max": 0.8487728424703169, "naucs_at_50_std": -0.014705882352948679, "naucs_at_50_diff1": 0.764639189008937, "naucs_at_100_max": 1.0, "naucs_at_100_std": 0.8585434173669326, "naucs_at_100_diff1": 0.9215686274509768}, "../colpali/data_dir/eval_vidore/docvqa_test_subsampled": {"ndcg_at_1": 0.38803, "ndcg_at_3": 0.46109, "ndcg_at_5": 0.49029, "ndcg_at_10": 0.50319, "ndcg_at_20": 0.51979, "ndcg_at_50": 0.5325, "ndcg_at_100": 0.54495, "map_at_1": 0.38803, "map_at_3": 0.44272, "map_at_5": 0.45891, "map_at_10": 0.46423, "map_at_20": 0.46867, "map_at_50": 0.47069, "map_at_100": 0.47175, "recall_at_1": 0.38803, "recall_at_3": 0.51441, "recall_at_5": 0.58537, "recall_at_10": 0.62528, "recall_at_20": 0.6918, "recall_at_50": 0.7561, "recall_at_100": 0.8337, "precision_at_1": 0.38803, "precision_at_3": 0.17147, "precision_at_5": 0.11707, "precision_at_10": 0.06253, "precision_at_20": 0.03459, "precision_at_50": 0.01512, "precision_at_100": 0.00834, "mrr_at_1": 0.3858093126385809, "mrr_at_3": 0.4416112342941611, "mrr_at_5": 0.4577974870657798, "mrr_at_10": 0.46294125928272273, "mrr_at_20": 0.4675986755215576, "mrr_at_50": 0.4696136690706579, "mrr_at_100": 0.47067284991598984, "naucs_at_1_max": 0.6880912118081786, "naucs_at_1_std": 0.1227003172324307, "naucs_at_1_diff1": 0.7549324423028185, "naucs_at_3_max": 0.7563655154362492, "naucs_at_3_std": 0.2935640829770719, "naucs_at_3_diff1": 0.6672805174710129, "naucs_at_5_max": 0.7828606305856601, "naucs_at_5_std": 0.3598978437045625, "naucs_at_5_diff1": 0.6234834202744584, "naucs_at_10_max": 0.7933866938151406, "naucs_at_10_std": 0.3972366212280104, "naucs_at_10_diff1": 0.6243158334404361, "naucs_at_20_max": 0.8125506614708696, "naucs_at_20_std": 0.4573186675647723, "naucs_at_20_diff1": 0.6070464733977256, "naucs_at_50_max": 0.8001630369453867, "naucs_at_50_std": 0.5106491209009716, "naucs_at_50_diff1": 0.5700401539611133, "naucs_at_100_max": 0.8046502127635046, "naucs_at_100_std": 0.5681011886365621, "naucs_at_100_diff1": 0.5652815021362786}, "../colpali/data_dir/eval_vidore/syntheticDocQA_healthcare_industry_test": {"ndcg_at_1": 0.8, "ndcg_at_3": 0.88678, "ndcg_at_5": 0.89496, "ndcg_at_10": 0.90446, "ndcg_at_20": 0.90446, "ndcg_at_50": 0.90446, "ndcg_at_100": 0.90446, "map_at_1": 0.8, "map_at_3": 0.865, "map_at_5": 0.8695, "map_at_10": 0.87329, "map_at_20": 0.87329, "map_at_50": 0.87329, "map_at_100": 0.87329, "recall_at_1": 0.8, "recall_at_3": 0.95, "recall_at_5": 0.97, "recall_at_10": 1.0, "recall_at_20": 1.0, "recall_at_50": 1.0, "recall_at_100": 1.0, "precision_at_1": 0.8, "precision_at_3": 0.31667, "precision_at_5": 0.194, "precision_at_10": 0.1, "precision_at_20": 0.05, "precision_at_50": 0.02, "precision_at_100": 0.01, "mrr_at_1": 0.8, "mrr_at_3": 0.865, "mrr_at_5": 0.8695, "mrr_at_10": 0.8732896825396825, "mrr_at_20": 0.8732896825396825, "mrr_at_50": 0.8732896825396825, "mrr_at_100": 0.8732896825396825, "naucs_at_1_max": 0.3517732267732264, "naucs_at_1_std": -0.1562687312687319, "naucs_at_1_diff1": 0.6350899100899094, "naucs_at_3_max": 0.38347338935574254, "naucs_at_3_std": -0.5926237161531261, "naucs_at_3_diff1": 0.8627450980392143, "naucs_at_5_max": 0.7860255213196407, "naucs_at_5_std": -0.21988795518207335, "naucs_at_5_diff1": 0.8638344226579531, "naucs_at_10_max": 1.0, "naucs_at_10_std": 1.0, "naucs_at_10_diff1": 1.0, "naucs_at_20_max": 1.0, "naucs_at_20_std": 1.0, "naucs_at_20_diff1": 1.0, "naucs_at_50_max": NaN, "naucs_at_50_std": NaN, "naucs_at_50_diff1": NaN, "naucs_at_100_max": NaN, "naucs_at_100_std": NaN, "naucs_at_100_diff1": NaN}, "../colpali/data_dir/eval_vidore/infovqa_test_subsampled": {"ndcg_at_1": 0.70648, "ndcg_at_3": 0.77988, "ndcg_at_5": 0.79791, "ndcg_at_10": 0.81178, "ndcg_at_20": 0.81896, "ndcg_at_50": 0.82317, "ndcg_at_100": 0.82486, "map_at_1": 0.70648, "map_at_3": 0.76248, "map_at_5": 0.7723, "map_at_10": 0.7781, "map_at_20": 0.78007, "map_at_50": 0.78083, "map_at_100": 0.78099, "recall_at_1": 0.70648, "recall_at_3": 0.82996, "recall_at_5": 0.87449, "recall_at_10": 0.917, "recall_at_20": 0.94534, "recall_at_50": 0.96559, "recall_at_100": 0.97571, "precision_at_1": 0.70648, "precision_at_3": 0.27665, "precision_at_5": 0.1749, "precision_at_10": 0.0917, "precision_at_20": 0.04727, "precision_at_50": 0.01931, "precision_at_100": 0.00976, "mrr_at_1": 0.708502024291498, "mrr_at_3": 0.7641700404858299, "mrr_at_5": 0.7731781376518219, "mrr_at_10": 0.7793160786581839, "mrr_at_20": 0.7812885331505837, "mrr_at_50": 0.78204531049572, "mrr_at_100": 0.7822104132561041, "naucs_at_1_max": 0.35210693912642194, "naucs_at_1_std": -0.25981339755457183, "naucs_at_1_diff1": 0.8006069446320773, "naucs_at_3_max": 0.5563684521448528, "naucs_at_3_std": -0.010153976919273033, "naucs_at_3_diff1": 0.7128786278945445, "naucs_at_5_max": 0.6558564747123251, "naucs_at_5_std": 0.1663990479174898, "naucs_at_5_diff1": 0.6930629953864437, "naucs_at_10_max": 0.8006339864253019, "naucs_at_10_std": 0.4302090584095774, "naucs_at_10_diff1": 0.6775921451523413, "naucs_at_20_max": 0.8681033015554591, "naucs_at_20_std": 0.6931262081290481, "naucs_at_20_diff1": 0.6412583755689819, "naucs_at_50_max": 0.9606153470616298, "naucs_at_50_std": 0.8639551994341516, "naucs_at_50_diff1": 0.7053052116711681, "naucs_at_100_max": 0.9550885295093762, "naucs_at_100_std": 0.8934242475636182, "naucs_at_100_diff1": 0.6165437325193877}, "../colpali/data_dir/eval_vidore/syntheticDocQA_energy_test": {"ndcg_at_1": 0.76, "ndcg_at_3": 0.82547, "ndcg_at_5": 0.82934, "ndcg_at_10": 0.84907, "ndcg_at_20": 0.85426, "ndcg_at_50": 0.85988, "ndcg_at_100": 0.85988, "map_at_1": 0.76, "map_at_3": 0.81, "map_at_5": 0.812, "map_at_10": 0.82031, "map_at_20": 0.8218, "map_at_50": 0.82256, "map_at_100": 0.82256, "recall_at_1": 0.76, "recall_at_3": 0.87, "recall_at_5": 0.88, "recall_at_10": 0.94, "recall_at_20": 0.96, "recall_at_50": 0.99, "recall_at_100": 0.99, "precision_at_1": 0.76, "precision_at_3": 0.29, "precision_at_5": 0.176, "precision_at_10": 0.094, "precision_at_20": 0.048, "precision_at_50": 0.0198, "precision_at_100": 0.0099, "mrr_at_1": 0.77, "mrr_at_3": 0.8166666666666668, "mrr_at_5": 0.8186666666666667, "mrr_at_10": 0.8269801587301587, "mrr_at_20": 0.8284636752136753, "mrr_at_50": 0.8292760517772038, "mrr_at_100": 0.8292760517772038, "naucs_at_1_max": 0.4600242277407624, "naucs_at_1_std": -0.3603876438522109, "naucs_at_1_diff1": 0.8055723803755296, "naucs_at_3_max": 0.47600533214841056, "naucs_at_3_std": -0.3254832259497905, "naucs_at_3_diff1": 0.7038065615048492, "naucs_at_5_max": 0.4260658606426989, "naucs_at_5_std": -0.4347756920139959, "naucs_at_5_diff1": 0.681872414890232, "naucs_at_10_max": 0.6507158418923127, "naucs_at_10_std": -0.23996265172735637, "naucs_at_10_diff1": 0.8031434796140697, "naucs_at_20_max": 0.6365546218487388, "naucs_at_20_std": -0.406629318394016, "naucs_at_20_diff1": 0.7373949579831963, "naucs_at_50_max": 1.0, "naucs_at_50_std": 0.554154995331464, "naucs_at_50_diff1": 1.0, "naucs_at_100_max": 1.0, "naucs_at_100_std": 0.554154995331464, "naucs_at_100_diff1": 1.0}, "../colpali/data_dir/eval_vidore/tabfquad_test_subsampled": {"ndcg_at_1": 0.71429, "ndcg_at_3": 0.77185, "ndcg_at_5": 0.79075, "ndcg_at_10": 0.80106, "ndcg_at_20": 0.8125, "ndcg_at_50": 0.82107, "ndcg_at_100": 0.82468, "map_at_1": 0.71429, "map_at_3": 0.75833, "map_at_5": 0.76869, "map_at_10": 0.7729, "map_at_20": 0.77588, "map_at_50": 0.77729, "map_at_100": 0.77764, "recall_at_1": 0.71429, "recall_at_3": 0.81071, "recall_at_5": 0.85714, "recall_at_10": 0.88929, "recall_at_20": 0.93571, "recall_at_50": 0.97857, "recall_at_100": 1.0, "precision_at_1": 0.71429, "precision_at_3": 0.27024, "precision_at_5": 0.17143, "precision_at_10": 0.08893, "precision_at_20": 0.04679, "precision_at_50": 0.01957, "precision_at_100": 0.01, "mrr_at_1": 0.7142857142857143, "mrr_at_3": 0.7589285714285714, "mrr_at_5": 0.7696428571428572, "mrr_at_10": 0.7738577097505669, "mrr_at_20": 0.7768215657475112, "mrr_at_50": 0.7782313410735464, "mrr_at_100": 0.7785880748819538, "naucs_at_1_max": 0.7914299115518632, "naucs_at_1_std": 0.3343473599571155, "naucs_at_1_diff1": 0.8793621013133214, "naucs_at_3_max": 0.8434654450556667, "naucs_at_3_std": 0.43567767493381904, "naucs_at_3_diff1": 0.8138482623960349, "naucs_at_5_max": 0.857142857142856, "naucs_at_5_std": 0.5951409135082594, "naucs_at_5_diff1": 0.8132896015549081, "naucs_at_10_max": 0.8319910924010837, "naucs_at_10_std": 0.6162411152801917, "naucs_at_10_diff1": 0.7920289191909943, "naucs_at_20_max": 0.8185755783795007, "naucs_at_20_std": 0.6212522045855406, "naucs_at_20_diff1": 0.7624234879136864, "naucs_at_50_max": 0.7786336756925044, "naucs_at_50_std": 0.438297541238722, "naucs_at_50_diff1": 0.676314970432611, "naucs_at_100_max": 1.0, "naucs_at_100_std": 1.0, "naucs_at_100_diff1": 1.0}}