| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.328, | |
| "acc_stderr": 0.014853842487270333 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.325, | |
| "acc_stderr": 0.014818724459095527 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.013680495725767789 | |
| }, | |
| "cb": { | |
| "acc": 0.42857142857142855, | |
| "acc_stderr": 0.06672848092813058, | |
| "f1": 0.30414746543778803 | |
| }, | |
| "copa": { | |
| "acc": 0.66, | |
| "acc_stderr": 0.04760952285695237 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.28141804421429994, | |
| "acc_stderr": 0.00448771884333028, | |
| "acc_norm": 0.29934276040629354, | |
| "acc_norm_stderr": 0.004570342034463229 | |
| }, | |
| "rte": { | |
| "acc": 0.5487364620938628, | |
| "acc_stderr": 0.029953149241808943 | |
| }, | |
| "winogrande": { | |
| "acc": 0.4988161010260458, | |
| "acc_stderr": 0.014052446290529019 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.5831106360235169, | |
| "acc_stderr": 0.011401581234266751 | |
| }, | |
| "boolq": { | |
| "acc": 0.5541284403669725, | |
| "acc_stderr": 0.008693659886486843 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.4065656565656566, | |
| "acc_stderr": 0.010079056419223503, | |
| "acc_norm": 0.36952861952861954, | |
| "acc_norm_stderr": 0.009904325878447317 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.18430034129692832, | |
| "acc_stderr": 0.011330517933037411, | |
| "acc_norm": 0.23122866894197952, | |
| "acc_norm_stderr": 0.012320858834772278 | |
| }, | |
| "sciq": { | |
| "acc": 0.685, | |
| "acc_stderr": 0.014696631960792496, | |
| "acc_norm": 0.595, | |
| "acc_norm_stderr": 0.015531136990453047 | |
| }, | |
| "piqa": { | |
| "acc": 0.6376496191512514, | |
| "acc_stderr": 0.011215040215104565, | |
| "acc_norm": 0.6289445048966268, | |
| "acc_norm_stderr": 0.011271222398600525 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |