{ "results": { "anli_r1": { "acc": 0.337, "acc_stderr": 0.014955087918653586 }, "anli_r2": { "acc": 0.369, "acc_stderr": 0.015266698139154619 }, "anli_r3": { "acc": 0.3541666666666667, "acc_stderr": 0.013811933499570956 }, "cb": { "acc": 0.4642857142857143, "acc_stderr": 0.0672477765493766, "f1": 0.2330246913580247 }, "copa": { "acc": 0.71, "acc_stderr": 0.045604802157206845 }, "hellaswag": { "acc": 0.39603664608643696, "acc_stderr": 0.004880726787988637, "acc_norm": 0.49422425811591314, "acc_norm_stderr": 0.004989448490164425 }, "rte": { "acc": 0.48375451263537905, "acc_stderr": 0.030080573208738064 }, "winogrande": { "acc": 0.5146014206787688, "acc_stderr": 0.014046492383275835 }, "storycloze_2016": { "acc": 0.6611437733832175, "acc_stderr": 0.010945481921408545 }, "boolq": { "acc": 0.4712538226299694, "acc_stderr": 0.00873059018871715 }, "arc_easy": { "acc": 0.4970538720538721, "acc_stderr": 0.010259605416237575, "acc_norm": 0.4730639730639731, "acc_norm_stderr": 0.010244884740620087 }, "arc_challenge": { "acc": 0.24488054607508533, "acc_stderr": 0.012566273985131354, "acc_norm": 0.2619453924914676, "acc_norm_stderr": 0.012849054826858117 }, "sciq": { "acc": 0.742, "acc_stderr": 0.013842963108656603, "acc_norm": 0.694, "acc_norm_stderr": 0.014580006055436969 }, "piqa": { "acc": 0.6926006528835691, "acc_stderr": 0.01076560250693907, "acc_norm": 0.6991294885745375, "acc_norm_stderr": 0.010700745724145973 } }, "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 } }