| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.313, | |
| "acc_stderr": 0.014671272822977885 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.014818724459095527 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3491666666666667, | |
| "acc_stderr": 0.01376707539507725 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.0663363415035954, | |
| "f1": 0.2576489533011272 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45200159330810596, | |
| "acc_stderr": 0.004966736811010493, | |
| "acc_norm": 0.5935072694682334, | |
| "acc_norm_stderr": 0.004901747426331751 | |
| }, | |
| "rte": { | |
| "acc": 0.516245487364621, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5666929755327546, | |
| "acc_stderr": 0.01392691505275734 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6953500801710315, | |
| "acc_stderr": 0.0106434269886468 | |
| }, | |
| "boolq": { | |
| "acc": 0.618348623853211, | |
| "acc_stderr": 0.008496550741178254 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6346801346801347, | |
| "acc_stderr": 0.009880576614806928, | |
| "acc_norm": 0.6254208754208754, | |
| "acc_norm_stderr": 0.009931758820410629 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2764505119453925, | |
| "acc_stderr": 0.013069662474252425, | |
| "acc_norm": 0.2986348122866894, | |
| "acc_norm_stderr": 0.013374078615068752 | |
| }, | |
| "sciq": { | |
| "acc": 0.898, | |
| "acc_stderr": 0.009575368801653892, | |
| "acc_norm": 0.902, | |
| "acc_norm_stderr": 0.009406619184621228 | |
| }, | |
| "piqa": { | |
| "acc": 0.7486398258977149, | |
| "acc_stderr": 0.010121156016819255, | |
| "acc_norm": 0.750816104461371, | |
| "acc_norm_stderr": 0.01009188277012021 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |