{ "results": { "anli_r1": { "acc": 0.316, "acc_stderr": 0.014709193056057134 }, "anli_r2": { "acc": 0.344, "acc_stderr": 0.015029633724408945 }, "anli_r3": { "acc": 0.335, "acc_stderr": 0.013630871843821482 }, "cb": { "acc": 0.48214285714285715, "acc_stderr": 0.0673769750864465, "f1": 0.32495309568480296 }, "copa": { "acc": 0.68, "acc_stderr": 0.04688261722621504 }, "hellaswag": { "acc": 0.38179645488946423, "acc_stderr": 0.004848341560492138, "acc_norm": 0.4785899223262298, "acc_norm_stderr": 0.004985204766555062 }, "rte": { "acc": 0.5054151624548736, "acc_stderr": 0.030094698123239966 }, "winogrande": { "acc": 0.5288082083662194, "acc_stderr": 0.014029141615909617 }, "storycloze_2016": { "acc": 0.6541956173169428, "acc_stderr": 0.010998874799044323 }, "boolq": { "acc": 0.5660550458715596, "acc_stderr": 0.008668405003744127 }, "arc_easy": { "acc": 0.5698653198653199, "acc_stderr": 0.010159130445178502, "acc_norm": 0.5517676767676768, "acc_norm_stderr": 0.010204645126856942 }, "arc_challenge": { "acc": 0.23720136518771331, "acc_stderr": 0.012430399829260844, "acc_norm": 0.2883959044368601, "acc_norm_stderr": 0.01323839442242817 }, "sciq": { "acc": 0.882, "acc_stderr": 0.010206869264381791, "acc_norm": 0.879, "acc_norm_stderr": 0.010318210380946097 }, "piqa": { "acc": 0.7187159956474428, "acc_stderr": 0.010490509832327423, "acc_norm": 0.7127312295973884, "acc_norm_stderr": 0.010557291761528637 } }, "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 } }