{ "results": { "anli_r1": { "acc": 0.318, "acc_stderr": 0.014734079309311901 }, "anli_r2": { "acc": 0.347, "acc_stderr": 0.015060472031706618 }, "anli_r3": { "acc": 0.3275, "acc_stderr": 0.013553211167251953 }, "cb": { "acc": 0.35714285714285715, "acc_stderr": 0.0646095738380922, "f1": 0.26868521549372615 }, "copa": { "acc": 0.79, "acc_stderr": 0.040936018074033256 }, "hellaswag": { "acc": 0.47470623381796456, "acc_stderr": 0.004983392650570956, "acc_norm": 0.6330412268472416, "acc_norm_stderr": 0.004809901151234834 }, "rte": { "acc": 0.51985559566787, "acc_stderr": 0.030072723167317184 }, "winogrande": { "acc": 0.5816890292028414, "acc_stderr": 0.013863669961195918 }, "storycloze_2016": { "acc": 0.7188669160876536, "acc_stderr": 0.010395836091628113 }, "boolq": { "acc": 0.636697247706422, "acc_stderr": 0.008411885836787163 }, "arc_easy": { "acc": 0.6254208754208754, "acc_stderr": 0.00993175882041061, "acc_norm": 0.5993265993265994, "acc_norm_stderr": 0.010055304474255585 }, "arc_challenge": { "acc": 0.30716723549488056, "acc_stderr": 0.013481034054980945, "acc_norm": 0.33276450511945393, "acc_norm_stderr": 0.013769863046192314 }, "sciq": { "acc": 0.904, "acc_stderr": 0.009320454434783222, "acc_norm": 0.893, "acc_norm_stderr": 0.009779910359847165 }, "piqa": { "acc": 0.749727965179543, "acc_stderr": 0.010106561880089786, "acc_norm": 0.76550598476605, "acc_norm_stderr": 0.00988520314324054 } }, "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 } }