{ "results": { "anli_r1": { "acc": 0.32, "acc_stderr": 0.014758652303574886 }, "anli_r2": { "acc": 0.346, "acc_stderr": 0.015050266127564443 }, "anli_r3": { "acc": 0.35, "acc_stderr": 0.013774667009018554 }, "cb": { "acc": 0.39285714285714285, "acc_stderr": 0.0658538889806635, "f1": 0.3130977130977131 }, "copa": { "acc": 0.77, "acc_stderr": 0.04229525846816506 }, "hellaswag": { "acc": 0.4643497311292571, "acc_stderr": 0.004977081808179426, "acc_norm": 0.603963353913563, "acc_norm_stderr": 0.004880726787988643 }, "rte": { "acc": 0.5379061371841155, "acc_stderr": 0.030009848912529117 }, "winogrande": { "acc": 0.5659037095501184, "acc_stderr": 0.013929882555694058 }, "storycloze_2016": { "acc": 0.706574024585783, "acc_stderr": 0.010529489334744466 }, "boolq": { "acc": 0.6238532110091743, "acc_stderr": 0.008472516562330725 }, "arc_easy": { "acc": 0.5976430976430976, "acc_stderr": 0.010062244711011518, "acc_norm": 0.5913299663299664, "acc_norm_stderr": 0.010087174498762886 }, "arc_challenge": { "acc": 0.2858361774744027, "acc_stderr": 0.01320319608853737, "acc_norm": 0.3122866894197952, "acc_norm_stderr": 0.013542598541688065 }, "sciq": { "acc": 0.866, "acc_stderr": 0.01077776229836968, "acc_norm": 0.859, "acc_norm_stderr": 0.011010914595992436 }, "piqa": { "acc": 0.749183895538629, "acc_stderr": 0.010113869547069044, "acc_norm": 0.7453754080522307, "acc_norm_stderr": 0.010164432237060499 } }, "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 } }