| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.342, | |
| "acc_stderr": 0.01500870618212173 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.323, | |
| "acc_stderr": 0.014794927843348642 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33166666666666667, | |
| "acc_stderr": 0.013596836729485163 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.23907547851209823 | |
| }, | |
| "copa": { | |
| "acc": 0.86, | |
| "acc_stderr": 0.03487350880197772 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5319657438757219, | |
| "acc_stderr": 0.0049795737655758615, | |
| "acc_norm": 0.7148974307906791, | |
| "acc_norm_stderr": 0.00450540617660685 | |
| }, | |
| "rte": { | |
| "acc": 0.5992779783393501, | |
| "acc_stderr": 0.029497229237163143 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6290449881610103, | |
| "acc_stderr": 0.013576399902231568 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7648316408337787, | |
| "acc_stderr": 0.009807347513356905 | |
| }, | |
| "boolq": { | |
| "acc": 0.6235474006116208, | |
| "acc_stderr": 0.008473882279194586 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6919191919191919, | |
| "acc_stderr": 0.009473887075826332, | |
| "acc_norm": 0.6830808080808081, | |
| "acc_norm_stderr": 0.009547254611446373 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3583617747440273, | |
| "acc_stderr": 0.014012883334859864, | |
| "acc_norm": 0.3677474402730375, | |
| "acc_norm_stderr": 0.01409099561816847 | |
| }, | |
| "sciq": { | |
| "acc": 0.938, | |
| "acc_stderr": 0.007629823996280307, | |
| "acc_norm": 0.93, | |
| "acc_norm_stderr": 0.008072494358323499 | |
| }, | |
| "piqa": { | |
| "acc": 0.7812840043525572, | |
| "acc_stderr": 0.009644731932667556, | |
| "acc_norm": 0.7840043525571273, | |
| "acc_norm_stderr": 0.009601236303553544 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |