{ "results": { "anli_r1": { "acc": 0.356, "acc_stderr": 0.01514904265930662 }, "anli_r2": { "acc": 0.366, "acc_stderr": 0.01524061272640576 }, "anli_r3": { "acc": 0.35583333333333333, "acc_stderr": 0.013826518748493315 }, "cb": { "acc": 0.5, "acc_stderr": 0.06741998624632421, "f1": 0.34491725768321513 }, "copa": { "acc": 0.71, "acc_stderr": 0.045604802157206845 }, "hellaswag": { "acc": 0.4587731527584147, "acc_stderr": 0.004972790690640181, "acc_norm": 0.5902210714997013, "acc_norm_stderr": 0.004907877144720023 }, "rte": { "acc": 0.4729241877256318, "acc_stderr": 0.030052303463143706 }, "winogrande": { "acc": 0.5406471981057617, "acc_stderr": 0.014005973823825136 }, "storycloze_2016": { "acc": 0.689470871191876, "acc_stderr": 0.010700112173178448 }, "boolq": { "acc": 0.5131498470948013, "acc_stderr": 0.008742030090044975 }, "arc_easy": { "acc": 0.5231481481481481, "acc_stderr": 0.010248782484554471, "acc_norm": 0.5046296296296297, "acc_norm_stderr": 0.010259343705889728 }, "arc_challenge": { "acc": 0.2713310580204778, "acc_stderr": 0.012993807727545792, "acc_norm": 0.310580204778157, "acc_norm_stderr": 0.013522292098053054 }, "sciq": { "acc": 0.795, "acc_stderr": 0.012772554096113109, "acc_norm": 0.783, "acc_norm_stderr": 0.01304151375727071 }, "piqa": { "acc": 0.7236126224156693, "acc_stderr": 0.010434162388275624, "acc_norm": 0.7312295973884657, "acc_norm_stderr": 0.010343392940090011 } }, "versions": { "anli_r1": 0, "anli_r2": 0, "anli_r3": 0, "cb": 1, "copa": 0, "hellaswag": 0, "rte": 0, "winogrande": 0, "storycloze_2016": 0, "boolq": 1, "arc_easy": 0, "arc_challenge": 0, "sciq": 0, "piqa": 0 } }