{ "results": { "anli_r1": { "acc": 0.325, "acc_stderr": 0.014818724459095526 }, "anli_r2": { "acc": 0.318, "acc_stderr": 0.0147340793093119 }, "anli_r3": { "acc": 0.3408333333333333, "acc_stderr": 0.013688600793296936 }, "cb": { "acc": 0.5, "acc_stderr": 0.06741998624632421, "f1": 0.35276653171390016 }, "copa": { "acc": 0.72, "acc_stderr": 0.04512608598542128 }, "hellaswag": { "acc": 0.43766182035451107, "acc_stderr": 0.004950848456984546, "acc_norm": 0.5681139215295757, "acc_norm_stderr": 0.004943264339868656 }, "rte": { "acc": 0.5451263537906137, "acc_stderr": 0.029973636495415252 }, "winogrande": { "acc": 0.5453827940015785, "acc_stderr": 0.01399448102706599 }, "storycloze_2016": { "acc": 0.6953500801710315, "acc_stderr": 0.010643426988646804 }, "boolq": { "acc": 0.5969418960244648, "acc_stderr": 0.008579113210566456 }, "arc_easy": { "acc": 0.5702861952861953, "acc_stderr": 0.010157908005763676, "acc_norm": 0.5244107744107744, "acc_norm_stderr": 0.010247548905242269 }, "arc_challenge": { "acc": 0.25853242320819114, "acc_stderr": 0.01279455375428868, "acc_norm": 0.2713310580204778, "acc_norm_stderr": 0.012993807727545792 }, "sciq": { "acc": 0.844, "acc_stderr": 0.011480235006122358, "acc_norm": 0.799, "acc_norm_stderr": 0.012679107214617326 }, "piqa": { "acc": 0.7334058759521219, "acc_stderr": 0.010316749863541367, "acc_norm": 0.735038084874864, "acc_norm_stderr": 0.010296557993316044 } }, "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 } }