{ "results": { "anli_r1": { "acc": 0.354, "acc_stderr": 0.015129868238451773 }, "anli_r2": { "acc": 0.331, "acc_stderr": 0.01488827258820394 }, "anli_r3": { "acc": 0.3458333333333333, "acc_stderr": 0.01373624534231101 }, "cb": { "acc": 0.2857142857142857, "acc_stderr": 0.06091449038731724, "f1": 0.30952380952380953 }, "copa": { "acc": 0.7, "acc_stderr": 0.046056618647183814 }, "hellaswag": { "acc": 0.4360685122485561, "acc_stderr": 0.004948824501355485, "acc_norm": 0.5632344154550887, "acc_norm_stderr": 0.004949716368890496 }, "rte": { "acc": 0.5306859205776173, "acc_stderr": 0.03003973059219781 }, "winogrande": { "acc": 0.5666929755327546, "acc_stderr": 0.013926915052757347 }, "storycloze_2016": { "acc": 0.6969535008017104, "acc_stderr": 0.010627613073376715 }, "boolq": { "acc": 0.5501529051987768, "acc_stderr": 0.008700950643028801 }, "arc_easy": { "acc": 0.5707070707070707, "acc_stderr": 0.010156678075911087, "acc_norm": 0.5172558922558923, "acc_norm_stderr": 0.010253671674754631 }, "arc_challenge": { "acc": 0.2721843003412969, "acc_stderr": 0.013006600406423707, "acc_norm": 0.2832764505119454, "acc_norm_stderr": 0.013167478735134575 }, "sciq": { "acc": 0.848, "acc_stderr": 0.011358918303475282, "acc_norm": 0.758, "acc_norm_stderr": 0.013550631705555958 }, "piqa": { "acc": 0.7225244831338411, "acc_stderr": 0.010446818281039959, "acc_norm": 0.7317736670293797, "acc_norm_stderr": 0.010336761992404485 } }, "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 } }