{ "results": { "anli_r1": { "acc": 0.327, "acc_stderr": 0.014842213153411242 }, "anli_r2": { "acc": 0.338, "acc_stderr": 0.014965960710224487 }, "anli_r3": { "acc": 0.3333333333333333, "acc_stderr": 0.0136139500102256 }, "cb": { "acc": 0.5, "acc_stderr": 0.06741998624632421, "f1": 0.34164884770729387 }, "copa": { "acc": 0.77, "acc_stderr": 0.04229525846816506 }, "hellaswag": { "acc": 0.46634136626170086, "acc_stderr": 0.00497846269096693, "acc_norm": 0.6101374228241386, "acc_norm_stderr": 0.004867221634461266 }, "rte": { "acc": 0.5776173285198556, "acc_stderr": 0.029731622646495887 }, "winogrande": { "acc": 0.5895816890292028, "acc_stderr": 0.013825107120035861 }, "storycloze_2016": { "acc": 0.7167290219134153, "acc_stderr": 0.010419760409155363 }, "boolq": { "acc": 0.6186544342507645, "acc_stderr": 0.008495245917063564 }, "arc_easy": { "acc": 0.5892255892255892, "acc_stderr": 0.01009510134934865, "acc_norm": 0.5361952861952862, "acc_norm_stderr": 0.01023286555034674 }, "arc_challenge": { "acc": 0.26706484641638223, "acc_stderr": 0.01292893319649636, "acc_norm": 0.30631399317406144, "acc_norm_stderr": 0.013470584417276513 }, "sciq": { "acc": 0.847, "acc_stderr": 0.011389500459665532, "acc_norm": 0.777, "acc_norm_stderr": 0.013169830843425673 }, "piqa": { "acc": 0.7459194776931447, "acc_stderr": 0.010157271999135041, "acc_norm": 0.7524483133841132, "acc_norm_stderr": 0.010069703966857114 } }, "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 } }