{ "results": { "anli_r1": { "acc": 0.314, "acc_stderr": 0.014683991951087966 }, "anli_r2": { "acc": 0.319, "acc_stderr": 0.014746404865473477 }, "anli_r3": { "acc": 0.3475, "acc_stderr": 0.013751753243291852 }, "cb": { "acc": 0.4107142857142857, "acc_stderr": 0.0663363415035954, "f1": 0.3282195387458545 }, "copa": { "acc": 0.76, "acc_stderr": 0.04292346959909283 }, "hellaswag": { "acc": 0.47321250746863175, "acc_stderr": 0.004982615233057104, "acc_norm": 0.6222863971320454, "acc_norm_stderr": 0.004838246410786253 }, "rte": { "acc": 0.5487364620938628, "acc_stderr": 0.029953149241808946 }, "winogrande": { "acc": 0.5706393054459353, "acc_stderr": 0.013911537499969179 }, "storycloze_2016": { "acc": 0.7049706039551042, "acc_stderr": 0.010546232606962287 }, "boolq": { "acc": 0.5957186544342508, "acc_stderr": 0.008583313811372076 }, "arc_easy": { "acc": 0.617003367003367, "acc_stderr": 0.009974920384536469, "acc_norm": 0.5744949494949495, "acc_norm_stderr": 0.010145271182591026 }, "arc_challenge": { "acc": 0.29436860068259385, "acc_stderr": 0.013318528460539426, "acc_norm": 0.32337883959044367, "acc_norm_stderr": 0.013669421630012132 }, "sciq": { "acc": 0.884, "acc_stderr": 0.010131468138756995, "acc_norm": 0.859, "acc_norm_stderr": 0.011010914595992443 }, "piqa": { "acc": 0.7540805223068553, "acc_stderr": 0.010047331865625191, "acc_norm": 0.7584330794341676, "acc_norm_stderr": 0.009986718001804454 } }, "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 } }