{ "results": { "anli_r1": { "acc": 0.335, "acc_stderr": 0.014933117490932575 }, "anli_r2": { "acc": 0.335, "acc_stderr": 0.014933117490932573 }, "anli_r3": { "acc": 0.3325, "acc_stderr": 0.01360541734571053 }, "cb": { "acc": 0.4107142857142857, "acc_stderr": 0.0663363415035954, "f1": 0.1940928270042194 }, "copa": { "acc": 0.7, "acc_stderr": 0.046056618647183814 }, "hellaswag": { "acc": 0.3520215096594304, "acc_stderr": 0.004766245539606642, "acc_norm": 0.4331806413065126, "acc_norm_stderr": 0.0049450236570322765 }, "rte": { "acc": 0.5126353790613718, "acc_stderr": 0.030086851767188564 }, "winogrande": { "acc": 0.5272296764009471, "acc_stderr": 0.014031631629827696 }, "storycloze_2016": { "acc": 0.6424371993586317, "acc_stderr": 0.011083341168827782 }, "boolq": { "acc": 0.5966360856269113, "acc_stderr": 0.00858016855488973 }, "arc_easy": { "acc": 0.5210437710437711, "acc_stderr": 0.010250692602022559, "acc_norm": 0.4692760942760943, "acc_norm_stderr": 0.01024039558481524 }, "arc_challenge": { "acc": 0.22525597269624573, "acc_stderr": 0.0122078399954073, "acc_norm": 0.24658703071672355, "acc_norm_stderr": 0.012595726268790134 }, "sciq": { "acc": 0.825, "acc_stderr": 0.012021627157731975, "acc_norm": 0.71, "acc_norm_stderr": 0.014356395999905687 }, "piqa": { "acc": 0.6942328618063112, "acc_stderr": 0.010749627366141629, "acc_norm": 0.6942328618063112, "acc_norm_stderr": 0.010749627366141639 } }, "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 } }