| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.014922019523732968 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.339, | |
| "acc_stderr": 0.014976758771620344 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3308333333333333, | |
| "acc_stderr": 0.013588208070709006 | |
| }, | |
| "cb": { | |
| "acc": 0.35714285714285715, | |
| "acc_stderr": 0.06460957383809221, | |
| "f1": 0.182648401826484 | |
| }, | |
| "copa": { | |
| "acc": 0.87, | |
| "acc_stderr": 0.03379976689896309 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5327623979286995, | |
| "acc_stderr": 0.004979058078478698, | |
| "acc_norm": 0.7144991037641903, | |
| "acc_norm_stderr": 0.004507296196227816 | |
| }, | |
| "rte": { | |
| "acc": 0.5667870036101083, | |
| "acc_stderr": 0.029826764082138274 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6345698500394633, | |
| "acc_stderr": 0.013533965097638793 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7696419027258151, | |
| "acc_stderr": 0.009737002698356936 | |
| }, | |
| "boolq": { | |
| "acc": 0.6165137614678899, | |
| "acc_stderr": 0.008504304838837023 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6952861952861953, | |
| "acc_stderr": 0.009444871667360213, | |
| "acc_norm": 0.6797138047138047, | |
| "acc_norm_stderr": 0.009574152668739419 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.35494880546075086, | |
| "acc_stderr": 0.013983036904094099, | |
| "acc_norm": 0.36689419795221845, | |
| "acc_norm_stderr": 0.014084133118104298 | |
| }, | |
| "sciq": { | |
| "acc": 0.942, | |
| "acc_stderr": 0.007395315455792942, | |
| "acc_norm": 0.934, | |
| "acc_norm_stderr": 0.007855297938697596 | |
| }, | |
| "piqa": { | |
| "acc": 0.7774755168661589, | |
| "acc_stderr": 0.009704600975718238, | |
| "acc_norm": 0.79379760609358, | |
| "acc_norm_stderr": 0.009439460331609514 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |