{ "results": { "anli_r1": { "acc": 0.332, "acc_stderr": 0.01489959724281149 }, "anli_r2": { "acc": 0.362, "acc_stderr": 0.015204840912919503 }, "anli_r3": { "acc": 0.33416666666666667, "acc_stderr": 0.013622434813136788 }, "cb": { "acc": 0.4642857142857143, "acc_stderr": 0.06724777654937658, "f1": 0.4217687074829932 }, "copa": { "acc": 0.71, "acc_stderr": 0.045604802157206845 }, "hellaswag": { "acc": 0.40420235012945627, "acc_stderr": 0.004897340793314381, "acc_norm": 0.5269866560446126, "acc_norm_stderr": 0.004982508198584267 }, "rte": { "acc": 0.5776173285198556, "acc_stderr": 0.02973162264649588 }, "winogrande": { "acc": 0.5351223362273086, "acc_stderr": 0.014017773120881585 }, "storycloze_2016": { "acc": 0.6675574559059326, "acc_stderr": 0.01089386077834354 }, "boolq": { "acc": 0.5688073394495413, "acc_stderr": 0.008661853128165595 }, "arc_easy": { "acc": 0.6014309764309764, "acc_stderr": 0.010046455400477943, "acc_norm": 0.585016835016835, "acc_norm_stderr": 0.01011038315196114 }, "arc_challenge": { "acc": 0.28071672354948807, "acc_stderr": 0.013131238126975578, "acc_norm": 0.3046075085324232, "acc_norm_stderr": 0.013449522109932489 }, "sciq": { "acc": 0.918, "acc_stderr": 0.008680515615523727, "acc_norm": 0.908, "acc_norm_stderr": 0.009144376393151098 }, "piqa": { "acc": 0.7274211099020674, "acc_stderr": 0.010389256803296023, "acc_norm": 0.7290533188248096, "acc_norm_stderr": 0.010369718937426844 } }, "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 } }