{ "results": { "anli_r1": { "acc": 0.337, "acc_stderr": 0.014955087918653605 }, "anli_r2": { "acc": 0.336, "acc_stderr": 0.014944140233795027 }, "anli_r3": { "acc": 0.33416666666666667, "acc_stderr": 0.013622434813136781 }, "cb": { "acc": 0.4107142857142857, "acc_stderr": 0.0663363415035954, "f1": 0.1940928270042194 }, "copa": { "acc": 0.76, "acc_stderr": 0.04292346959909283 }, "hellaswag": { "acc": 0.4558852818163712, "acc_stderr": 0.004970322156997941, "acc_norm": 0.5975901214897431, "acc_norm_stderr": 0.004893814890208305 }, "rte": { "acc": 0.51985559566787, "acc_stderr": 0.030072723167317184 }, "winogrande": { "acc": 0.5777426992896606, "acc_stderr": 0.013881582030658554 }, "storycloze_2016": { "acc": 0.7113842864778194, "acc_stderr": 0.01047831178564294 }, "boolq": { "acc": 0.5464831804281346, "acc_stderr": 0.008707182331111646 }, "arc_easy": { "acc": 0.5782828282828283, "acc_stderr": 0.010133255284012325, "acc_norm": 0.5155723905723906, "acc_norm_stderr": 0.010254806331961897 }, "arc_challenge": { "acc": 0.2440273037542662, "acc_stderr": 0.012551447627856257, "acc_norm": 0.2841296928327645, "acc_norm_stderr": 0.013179442447653886 }, "sciq": { "acc": 0.818, "acc_stderr": 0.012207580637662153, "acc_norm": 0.743, "acc_norm_stderr": 0.013825416526895045 }, "piqa": { "acc": 0.7540805223068553, "acc_stderr": 0.010047331865625194, "acc_norm": 0.7622415669205659, "acc_norm_stderr": 0.009932525779525492 } }, "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 } }