| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811485 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.0149550879186536 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.355, | |
| "acc_stderr": 0.013819249004047296 | |
| }, | |
| "cb": { | |
| "acc": 0.48214285714285715, | |
| "acc_stderr": 0.0673769750864465, | |
| "f1": 0.4347442680776014 | |
| }, | |
| "copa": { | |
| "acc": 0.76, | |
| "acc_stderr": 0.04292346959909283 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4841665006970723, | |
| "acc_stderr": 0.004987278910505115, | |
| "acc_norm": 0.6352320254929297, | |
| "acc_norm_stderr": 0.004803812631994966 | |
| }, | |
| "rte": { | |
| "acc": 0.5306859205776173, | |
| "acc_stderr": 0.03003973059219781 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5990528808208366, | |
| "acc_stderr": 0.013773974554948033 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7151256012827365, | |
| "acc_stderr": 0.010437513986611718 | |
| }, | |
| "boolq": { | |
| "acc": 0.5669724770642202, | |
| "acc_stderr": 0.008666251305518059 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.617003367003367, | |
| "acc_stderr": 0.009974920384536469, | |
| "acc_norm": 0.5462962962962963, | |
| "acc_norm_stderr": 0.010215708295494117 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.28668941979522183, | |
| "acc_stderr": 0.013214986329274757, | |
| "acc_norm": 0.30631399317406144, | |
| "acc_norm_stderr": 0.013470584417276513 | |
| }, | |
| "sciq": { | |
| "acc": 0.845, | |
| "acc_stderr": 0.011450157470799475, | |
| "acc_norm": 0.757, | |
| "acc_norm_stderr": 0.013569640199177458 | |
| }, | |
| "piqa": { | |
| "acc": 0.7578890097932536, | |
| "acc_stderr": 0.00999437126910438, | |
| "acc_norm": 0.7676822633297062, | |
| "acc_norm_stderr": 0.009853201384168243 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |