{ "results": { "anli_r1": { "acc": 0.306, "acc_stderr": 0.014580006055436967 }, "anli_r2": { "acc": 0.359, "acc_stderr": 0.01517726422479859 }, "anli_r3": { "acc": 0.35583333333333333, "acc_stderr": 0.01382651874849331 }, "cb": { "acc": 0.5714285714285714, "acc_stderr": 0.06672848092813058, "f1": 0.5178689064558629 }, "copa": { "acc": 0.83, "acc_stderr": 0.03775251680686371 }, "hellaswag": { "acc": 0.47540330611431986, "acc_stderr": 0.004983740145218613, "acc_norm": 0.630551682931687, "acc_norm_stderr": 0.004816690123209743 }, "rte": { "acc": 0.5415162454873647, "acc_stderr": 0.029992535385373314 }, "winogrande": { "acc": 0.5919494869771112, "acc_stderr": 0.013812822643745028 }, "storycloze_2016": { "acc": 0.7258150721539284, "acc_stderr": 0.010316062787590011 }, "boolq": { "acc": 0.634862385321101, "acc_stderr": 0.008420941009417815 }, "arc_easy": { "acc": 0.6430976430976431, "acc_stderr": 0.009830630210347012, "acc_norm": 0.622895622895623, "acc_norm_stderr": 0.00994504194636652 }, "arc_challenge": { "acc": 0.2773037542662116, "acc_stderr": 0.013082095839059376, "acc_norm": 0.3174061433447099, "acc_norm_stderr": 0.01360223908803817 }, "sciq": { "acc": 0.914, "acc_stderr": 0.008870325962594766, "acc_norm": 0.906, "acc_norm_stderr": 0.009233052000787733 }, "piqa": { "acc": 0.7573449401523396, "acc_stderr": 0.010002002569708698, "acc_norm": 0.766050054406964, "acc_norm_stderr": 0.00987723689513744 } }, "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 } }