| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.014955087918653605 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.336, | |
| "acc_stderr": 0.014944140233795018 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3375, | |
| "acc_stderr": 0.013655897185463658 | |
| }, | |
| "cb": { | |
| "acc": 0.125, | |
| "acc_stderr": 0.04459412925079224, | |
| "f1": 0.10899594232927566 | |
| }, | |
| "copa": { | |
| "acc": 0.85, | |
| "acc_stderr": 0.03588702812826373 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5310695080661223, | |
| "acc_stderr": 0.004980138679161042, | |
| "acc_norm": 0.7102170882294364, | |
| "acc_norm_stderr": 0.004527343651130806 | |
| }, | |
| "rte": { | |
| "acc": 0.48375451263537905, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6227308602999211, | |
| "acc_stderr": 0.013622567928799501 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7589524318546232, | |
| "acc_stderr": 0.00989094649057693 | |
| }, | |
| "boolq": { | |
| "acc": 0.6474006116207951, | |
| "acc_stderr": 0.00835641249356212 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6999158249158249, | |
| "acc_stderr": 0.00940400055851335, | |
| "acc_norm": 0.6746632996632996, | |
| "acc_norm_stderr": 0.009613427708996187 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3515358361774744, | |
| "acc_stderr": 0.01395241369960094, | |
| "acc_norm": 0.3660409556313993, | |
| "acc_norm_stderr": 0.014077223108470139 | |
| }, | |
| "sciq": { | |
| "acc": 0.934, | |
| "acc_stderr": 0.00785529793869759, | |
| "acc_norm": 0.93, | |
| "acc_norm_stderr": 0.008072494358323508 | |
| }, | |
| "piqa": { | |
| "acc": 0.7698585418933623, | |
| "acc_stderr": 0.009820832826839815, | |
| "acc_norm": 0.7780195865070729, | |
| "acc_norm_stderr": 0.009696120744662022 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |