{ "results": { "anli_r1": { "acc": 0.306, "acc_stderr": 0.014580006055436967 }, "anli_r2": { "acc": 0.361, "acc_stderr": 0.015195720118175122 }, "anli_r3": { "acc": 0.31583333333333335, "acc_stderr": 0.01342456883035645 }, "cb": { "acc": 0.35714285714285715, "acc_stderr": 0.06460957383809221, "f1": 0.2902724515627741 }, "copa": { "acc": 0.81, "acc_stderr": 0.03942772444036623 }, "hellaswag": { "acc": 0.4742083250348536, "acc_stderr": 0.0049831384796043795, "acc_norm": 0.6274646484763992, "acc_norm_stderr": 0.004824917516374194 }, "rte": { "acc": 0.5667870036101083, "acc_stderr": 0.029826764082138277 }, "winogrande": { "acc": 0.580110497237569, "acc_stderr": 0.013870943986310391 }, "storycloze_2016": { "acc": 0.7167290219134153, "acc_stderr": 0.010419760409155363 }, "boolq": { "acc": 0.6042813455657492, "acc_stderr": 0.008552742471459792 }, "arc_easy": { "acc": 0.622895622895623, "acc_stderr": 0.009945041946366515, "acc_norm": 0.6035353535353535, "acc_norm_stderr": 0.01003741276306453 }, "arc_challenge": { "acc": 0.30119453924914674, "acc_stderr": 0.01340674176784762, "acc_norm": 0.3250853242320819, "acc_norm_stderr": 0.013688147309729122 }, "sciq": { "acc": 0.897, "acc_stderr": 0.009616833339695798, "acc_norm": 0.882, "acc_norm_stderr": 0.010206869264381791 }, "piqa": { "acc": 0.7573449401523396, "acc_stderr": 0.0100020025697087, "acc_norm": 0.764417845484222, "acc_norm_stderr": 0.009901067586473893 } }, "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 } }