| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932575 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.014842213153411242 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3458333333333333, | |
| "acc_stderr": 0.013736245342311012 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.28503144654088053 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.040201512610368445 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4422425811591316, | |
| "acc_stderr": 0.004956378590571537, | |
| "acc_norm": 0.5832503485361482, | |
| "acc_norm_stderr": 0.004920130733271772 | |
| }, | |
| "rte": { | |
| "acc": 0.5415162454873647, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5659037095501184, | |
| "acc_stderr": 0.01392988255569405 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7220737573490112, | |
| "acc_stderr": 0.010359403651225854 | |
| }, | |
| "boolq": { | |
| "acc": 0.5908256880733945, | |
| "acc_stderr": 0.008599563442397349 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6153198653198653, | |
| "acc_stderr": 0.009983171707009008, | |
| "acc_norm": 0.5980639730639731, | |
| "acc_norm_stderr": 0.010060521220920566 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2687713310580205, | |
| "acc_stderr": 0.012955065963710691, | |
| "acc_norm": 0.30204778156996587, | |
| "acc_norm_stderr": 0.013417519144716417 | |
| }, | |
| "sciq": { | |
| "acc": 0.889, | |
| "acc_stderr": 0.009938701010583726, | |
| "acc_norm": 0.862, | |
| "acc_norm_stderr": 0.010912152632504394 | |
| }, | |
| "piqa": { | |
| "acc": 0.7230685527747551, | |
| "acc_stderr": 0.010440499969334535, | |
| "acc_norm": 0.733949945593036, | |
| "acc_norm_stderr": 0.010310039263352826 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |