{ "results": { "anli_r1": { "acc": 0.332, "acc_stderr": 0.014899597242811471 }, "anli_r2": { "acc": 0.334, "acc_stderr": 0.014922019523732961 }, "anli_r3": { "acc": 0.3275, "acc_stderr": 0.013553211167251951 }, "cb": { "acc": 0.30357142857142855, "acc_stderr": 0.06199938655510754, "f1": 0.24454009245974814 }, "copa": { "acc": 0.76, "acc_stderr": 0.04292346959909283 }, "hellaswag": { "acc": 0.4312885879306911, "acc_stderr": 0.004942440746328495, "acc_norm": 0.55646285600478, "acc_norm_stderr": 0.004957863944093132 }, "rte": { "acc": 0.516245487364621, "acc_stderr": 0.030080573208738064 }, "winogrande": { "acc": 0.5477505919494869, "acc_stderr": 0.013988256216606008 }, "storycloze_2016": { "acc": 0.6958845537145911, "acc_stderr": 0.01063817265519479 }, "boolq": { "acc": 0.5889908256880734, "acc_stderr": 0.008605429733982185 }, "arc_easy": { "acc": 0.5740740740740741, "acc_stderr": 0.010146568651002255, "acc_norm": 0.553030303030303, "acc_norm_stderr": 0.01020191492779168 }, "arc_challenge": { "acc": 0.2593856655290102, "acc_stderr": 0.012808273573927102, "acc_norm": 0.28924914675767915, "acc_norm_stderr": 0.013250012579393443 }, "sciq": { "acc": 0.88, "acc_stderr": 0.01028132801274739, "acc_norm": 0.843, "acc_norm_stderr": 0.011510146979230189 }, "piqa": { "acc": 0.7301414581066377, "acc_stderr": 0.010356595421852197, "acc_norm": 0.735038084874864, "acc_norm_stderr": 0.010296557993316049 } }, "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 } }