{ "results": { "anli_r1": { "acc": 0.303, "acc_stderr": 0.014539683710535259 }, "anli_r2": { "acc": 0.335, "acc_stderr": 0.014933117490932572 }, "anli_r3": { "acc": 0.3408333333333333, "acc_stderr": 0.013688600793296936 }, "cb": { "acc": 0.4642857142857143, "acc_stderr": 0.06724777654937658, "f1": 0.3162578162578163 }, "copa": { "acc": 0.68, "acc_stderr": 0.04688261722621505 }, "hellaswag": { "acc": 0.302230631348337, "acc_stderr": 0.004582861219020889, "acc_norm": 0.3355905198167696, "acc_norm_stderr": 0.004712314511950981 }, "rte": { "acc": 0.4729241877256318, "acc_stderr": 0.0300523034631437 }, "winogrande": { "acc": 0.5169692186266772, "acc_stderr": 0.014044390401612974 }, "storycloze_2016": { "acc": 0.5905932656333511, "acc_stderr": 0.01137105952719707 }, "boolq": { "acc": 0.5342507645259938, "acc_stderr": 0.008724512941821092 }, "arc_easy": { "acc": 0.492003367003367, "acc_stderr": 0.010258471289841975, "acc_norm": 0.4642255892255892, "acc_norm_stderr": 0.010233488709726549 }, "arc_challenge": { "acc": 0.22696245733788395, "acc_stderr": 0.012240491536132866, "acc_norm": 0.26023890784982934, "acc_norm_stderr": 0.01282193022511256 }, "sciq": { "acc": 0.843, "acc_stderr": 0.0115101469792302, "acc_norm": 0.825, "acc_norm_stderr": 0.012021627157731968 }, "piqa": { "acc": 0.6485310119695321, "acc_stderr": 0.011139207691931191, "acc_norm": 0.6485310119695321, "acc_norm_stderr": 0.011139207691931193 } }, "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 } }