{ "results": { "anli_r1": { "acc": 0.336, "acc_stderr": 0.014944140233795023 }, "anli_r2": { "acc": 0.331, "acc_stderr": 0.014888272588203938 }, "anli_r3": { "acc": 0.3425, "acc_stderr": 0.013704669762934727 }, "cb": { "acc": 0.44642857142857145, "acc_stderr": 0.06703189227942398, "f1": 0.2956393200295639 }, "copa": { "acc": 0.78, "acc_stderr": 0.04163331998932261 }, "hellaswag": { "acc": 0.44901414060944034, "acc_stderr": 0.004963771168672082, "acc_norm": 0.5796654052977495, "acc_norm_stderr": 0.004926038197714521 }, "rte": { "acc": 0.516245487364621, "acc_stderr": 0.030080573208738064 }, "winogrande": { "acc": 0.569060773480663, "acc_stderr": 0.013917796623335962 }, "storycloze_2016": { "acc": 0.7071084981293426, "acc_stderr": 0.010523873293246305 }, "boolq": { "acc": 0.5923547400611621, "acc_stderr": 0.008594580270731615 }, "arc_easy": { "acc": 0.5883838383838383, "acc_stderr": 0.01009821864671491, "acc_norm": 0.5193602693602694, "acc_norm_stderr": 0.010252089491165522 }, "arc_challenge": { "acc": 0.257679180887372, "acc_stderr": 0.012780770562768403, "acc_norm": 0.28498293515358364, "acc_norm_stderr": 0.013191348179838793 }, "sciq": { "acc": 0.816, "acc_stderr": 0.012259457340938584, "acc_norm": 0.734, "acc_norm_stderr": 0.013979965645145153 }, "piqa": { "acc": 0.7274211099020674, "acc_stderr": 0.01038925680329602, "acc_norm": 0.7372143634385201, "acc_norm_stderr": 0.010269354068140777 } }, "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 } }