{ "results": { "anli_r1": { "acc": 0.346, "acc_stderr": 0.015050266127564434 }, "anli_r2": { "acc": 0.363, "acc_stderr": 0.015213890444671281 }, "anli_r3": { "acc": 0.3458333333333333, "acc_stderr": 0.013736245342311012 }, "cb": { "acc": 0.3392857142857143, "acc_stderr": 0.06384226561930827, "f1": 0.3177045177045177 }, "copa": { "acc": 0.83, "acc_stderr": 0.03775251680686371 }, "hellaswag": { "acc": 0.47938657637920734, "acc_stderr": 0.004985539159783413, "acc_norm": 0.633240390360486, "acc_norm_stderr": 0.004809352075008956 }, "rte": { "acc": 0.5523465703971119, "acc_stderr": 0.02993107036293953 }, "winogrande": { "acc": 0.6029992107340174, "acc_stderr": 0.0137510925198067 }, "storycloze_2016": { "acc": 0.729021913415286, "acc_stderr": 0.010278188399635044 }, "boolq": { "acc": 0.5740061162079511, "acc_stderr": 0.008648732832949143 }, "arc_easy": { "acc": 0.6397306397306397, "acc_stderr": 0.009851002584732383, "acc_norm": 0.627104377104377, "acc_norm_stderr": 0.009922743197129241 }, "arc_challenge": { "acc": 0.29948805460750855, "acc_stderr": 0.013385021637313576, "acc_norm": 0.3250853242320819, "acc_norm_stderr": 0.013688147309729122 }, "sciq": { "acc": 0.914, "acc_stderr": 0.008870325962594766, "acc_norm": 0.906, "acc_norm_stderr": 0.009233052000787733 }, "piqa": { "acc": 0.7535364526659413, "acc_stderr": 0.010054810789671822, "acc_norm": 0.7704026115342764, "acc_norm_stderr": 0.009812682950815183 } }, "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 } }