| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811487 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928357 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3541666666666667, | |
| "acc_stderr": 0.013811933499570954 | |
| }, | |
| "cb": { | |
| "acc": 0.5535714285714286, | |
| "acc_stderr": 0.06703189227942395, | |
| "f1": 0.38376730002345766 | |
| }, | |
| "copa": { | |
| "acc": 0.81, | |
| "acc_stderr": 0.03942772444036623 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.47400916152160927, | |
| "acc_stderr": 0.004983035420235716, | |
| "acc_norm": 0.619896434973113, | |
| "acc_norm_stderr": 0.004844199910173026 | |
| }, | |
| "rte": { | |
| "acc": 0.516245487364621, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5722178374112076, | |
| "acc_stderr": 0.013905134013839944 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7177979690005345, | |
| "acc_stderr": 0.010407834479647675 | |
| }, | |
| "boolq": { | |
| "acc": 0.5648318042813456, | |
| "acc_stderr": 0.008671229580582118 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5997474747474747, | |
| "acc_stderr": 0.010053550119896127, | |
| "acc_norm": 0.569023569023569, | |
| "acc_norm_stderr": 0.010161552863493746 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.27559726962457337, | |
| "acc_stderr": 0.01305716965576184, | |
| "acc_norm": 0.31569965870307165, | |
| "acc_norm_stderr": 0.013582571095815291 | |
| }, | |
| "sciq": { | |
| "acc": 0.844, | |
| "acc_stderr": 0.01148023500612236, | |
| "acc_norm": 0.794, | |
| "acc_norm_stderr": 0.012795613612786551 | |
| }, | |
| "piqa": { | |
| "acc": 0.7399347116430903, | |
| "acc_stderr": 0.0102348932490613, | |
| "acc_norm": 0.7595212187159956, | |
| "acc_norm_stderr": 0.009971345364651064 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |