| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.352, | |
| "acc_stderr": 0.015110404505648664 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.354, | |
| "acc_stderr": 0.015129868238451773 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3433333333333333, | |
| "acc_stderr": 0.01371263383046586 | |
| }, | |
| "cb": { | |
| "acc": 0.5, | |
| "acc_stderr": 0.06741998624632421, | |
| "f1": 0.3770850423844681 | |
| }, | |
| "copa": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.0368452949177471 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.530372435769767, | |
| "acc_stderr": 0.004980566907790448, | |
| "acc_norm": 0.7117108145787692, | |
| "acc_norm_stderr": 0.00452040633108404 | |
| }, | |
| "rte": { | |
| "acc": 0.5090252707581228, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6511444356748224, | |
| "acc_stderr": 0.013395059320137327 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7669695350080171, | |
| "acc_stderr": 0.009776301898548037 | |
| }, | |
| "boolq": { | |
| "acc": 0.6614678899082569, | |
| "acc_stderr": 0.008276502626477437 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6893939393939394, | |
| "acc_stderr": 0.009495260551195608, | |
| "acc_norm": 0.6750841750841751, | |
| "acc_norm_stderr": 0.00961020360450482 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.35665529010238906, | |
| "acc_stderr": 0.013998056902620199, | |
| "acc_norm": 0.3677474402730375, | |
| "acc_norm_stderr": 0.014090995618168468 | |
| }, | |
| "sciq": { | |
| "acc": 0.938, | |
| "acc_stderr": 0.007629823996280308, | |
| "acc_norm": 0.928, | |
| "acc_norm_stderr": 0.008178195576218681 | |
| }, | |
| "piqa": { | |
| "acc": 0.7872687704026116, | |
| "acc_stderr": 0.00954822312304734, | |
| "acc_norm": 0.7889009793253536, | |
| "acc_norm_stderr": 0.00952137737873415 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |