{ "results": { "anli_r1": { "acc": 0.329, "acc_stderr": 0.014865395385928355 }, "anli_r2": { "acc": 0.322, "acc_stderr": 0.014782913600996667 }, "anli_r3": { "acc": 0.3516666666666667, "acc_stderr": 0.013789711695404794 }, "cb": { "acc": 0.5, "acc_stderr": 0.06741998624632421, "f1": 0.34486817325800373 }, "copa": { "acc": 0.59, "acc_stderr": 0.049431107042371025 }, "hellaswag": { "acc": 0.2810197171878112, "acc_stderr": 0.00448578446857668, "acc_norm": 0.29645488946425014, "acc_norm_stderr": 0.004557606227194299 }, "rte": { "acc": 0.5306859205776173, "acc_stderr": 0.030039730592197816 }, "winogrande": { "acc": 0.510655090765588, "acc_stderr": 0.014049294536290393 }, "storycloze_2016": { "acc": 0.5622661678246926, "acc_stderr": 0.01147242507417594 }, "boolq": { "acc": 0.5626911314984709, "acc_stderr": 0.008676043429497423 }, "arc_easy": { "acc": 0.4006734006734007, "acc_stderr": 0.010055304474255558, "acc_norm": 0.3611111111111111, "acc_norm_stderr": 0.009856013425811244 }, "arc_challenge": { "acc": 0.17832764505119453, "acc_stderr": 0.011186139406711289, "acc_norm": 0.21843003412969283, "acc_norm_stderr": 0.012074291605700983 }, "sciq": { "acc": 0.684, "acc_stderr": 0.014709193056057125, "acc_norm": 0.644, "acc_norm_stderr": 0.015149042659306623 }, "piqa": { "acc": 0.6245919477693145, "acc_stderr": 0.011297839589776662, "acc_norm": 0.6213275299238302, "acc_norm_stderr": 0.011317163404516854 } }, "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 } }