{ "results": { "anli_r1": { "acc": 0.331, "acc_stderr": 0.014888272588203934 }, "anli_r2": { "acc": 0.352, "acc_stderr": 0.015110404505648668 }, "anli_r3": { "acc": 0.35583333333333333, "acc_stderr": 0.01382651874849331 }, "cb": { "acc": 0.2857142857142857, "acc_stderr": 0.06091449038731724, "f1": 0.1717171717171717 }, "copa": { "acc": 0.8, "acc_stderr": 0.040201512610368445 }, "hellaswag": { "acc": 0.5319657438757219, "acc_stderr": 0.004979573765575866, "acc_norm": 0.7045409281019717, "acc_norm_stderr": 0.004553164013379556 }, "rte": { "acc": 0.5667870036101083, "acc_stderr": 0.029826764082138277 }, "winogrande": { "acc": 0.6243093922651933, "acc_stderr": 0.013611257508380437 }, "storycloze_2016": { "acc": 0.7525387493319081, "acc_stderr": 0.009979234591920141 }, "boolq": { "acc": 0.6284403669724771, "acc_stderr": 0.008451598145076575 }, "arc_easy": { "acc": 0.6637205387205387, "acc_stderr": 0.009694178072725206, "acc_norm": 0.5896464646464646, "acc_norm_stderr": 0.010093531255765452 }, "arc_challenge": { "acc": 0.32081911262798635, "acc_stderr": 0.013640943091946526, "acc_norm": 0.33447098976109213, "acc_norm_stderr": 0.013787460322441374 }, "sciq": { "acc": 0.89, "acc_stderr": 0.00989939381972444, "acc_norm": 0.815, "acc_norm_stderr": 0.012285191326386684 }, "piqa": { "acc": 0.7731229597388466, "acc_stderr": 0.009771584259215172, "acc_norm": 0.7829162132752993, "acc_norm_stderr": 0.009618708415756788 } }, "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 } }