{ "results": { "anli_r1": { "acc": 0.315, "acc_stderr": 0.014696631960792506 }, "anli_r2": { "acc": 0.307, "acc_stderr": 0.014593284892852628 }, "anli_r3": { "acc": 0.3375, "acc_stderr": 0.013655897185463657 }, "cb": { "acc": 0.48214285714285715, "acc_stderr": 0.06737697508644648, "f1": 0.3421052631578947 }, "copa": { "acc": 0.66, "acc_stderr": 0.04760952285695237 }, "hellaswag": { "acc": 0.36168094005178253, "acc_stderr": 0.00479505103791772, "acc_norm": 0.44363672575184226, "acc_norm_stderr": 0.004957976789260528 }, "rte": { "acc": 0.5270758122743683, "acc_stderr": 0.0300523034631437 }, "winogrande": { "acc": 0.5248618784530387, "acc_stderr": 0.01403510288362775 }, "storycloze_2016": { "acc": 0.632816675574559, "acc_stderr": 0.011147041781368648 }, "boolq": { "acc": 0.5152905198776758, "acc_stderr": 0.008740964832224338 }, "arc_easy": { "acc": 0.5269360269360269, "acc_stderr": 0.010244884740620108, "acc_norm": 0.4962121212121212, "acc_norm_stderr": 0.010259489101351845 }, "arc_challenge": { "acc": 0.22696245733788395, "acc_stderr": 0.012240491536132866, "acc_norm": 0.2645051194539249, "acc_norm_stderr": 0.012889272949313368 }, "sciq": { "acc": 0.87, "acc_stderr": 0.010640169792499344, "acc_norm": 0.861, "acc_norm_stderr": 0.010945263761042965 }, "piqa": { "acc": 0.690424374319913, "acc_stderr": 0.010786656752183345, "acc_norm": 0.6936887921653971, "acc_norm_stderr": 0.010754970032367321 } }, "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 } }