| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.333, | |
| "acc_stderr": 0.01491084616422986 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811478 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33916666666666667, | |
| "acc_stderr": 0.013672343491681822 | |
| }, | |
| "cb": { | |
| "acc": 0.14285714285714285, | |
| "acc_stderr": 0.04718416136255829, | |
| "f1": 0.14017094017094014 | |
| }, | |
| "copa": { | |
| "acc": 0.86, | |
| "acc_stderr": 0.034873508801977725 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5324636526588329, | |
| "acc_stderr": 0.004979252954977319, | |
| "acc_norm": 0.7127066321449911, | |
| "acc_norm_stderr": 0.004515748192605716 | |
| }, | |
| "rte": { | |
| "acc": 0.4981949458483754, | |
| "acc_stderr": 0.030096267148976626 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6195737963693765, | |
| "acc_stderr": 0.013644727908656833 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7546766435061465, | |
| "acc_stderr": 0.009950137914623096 | |
| }, | |
| "boolq": { | |
| "acc": 0.6486238532110091, | |
| "acc_stderr": 0.00834978197660316 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.7028619528619529, | |
| "acc_stderr": 0.009377397867796849, | |
| "acc_norm": 0.6771885521885522, | |
| "acc_norm_stderr": 0.009593950220366737 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.35665529010238906, | |
| "acc_stderr": 0.013998056902620203, | |
| "acc_norm": 0.37627986348122866, | |
| "acc_norm_stderr": 0.014157022555407173 | |
| }, | |
| "sciq": { | |
| "acc": 0.938, | |
| "acc_stderr": 0.007629823996280306, | |
| "acc_norm": 0.93, | |
| "acc_norm_stderr": 0.008072494358323508 | |
| }, | |
| "piqa": { | |
| "acc": 0.764961915125136, | |
| "acc_stderr": 0.00989314668880531, | |
| "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 | |
| } | |
| } |