| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.338, | |
| "acc_stderr": 0.014965960710224489 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.352, | |
| "acc_stderr": 0.015110404505648673 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.013680495725767804 | |
| }, | |
| "cb": { | |
| "acc": 0.39285714285714285, | |
| "acc_stderr": 0.0658538889806635, | |
| "f1": 0.28048780487804875 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.03861229196653694 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5388368850826528, | |
| "acc_stderr": 0.004974706428434288, | |
| "acc_norm": 0.7166899024098785, | |
| "acc_norm_stderr": 0.004496847773250643 | |
| }, | |
| "rte": { | |
| "acc": 0.5415162454873647, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6416732438831886, | |
| "acc_stderr": 0.013476581172567528 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.774986638161411, | |
| "acc_stderr": 0.009656738215290533 | |
| }, | |
| "boolq": { | |
| "acc": 0.6639143730886851, | |
| "acc_stderr": 0.008261778456573672 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6898148148148148, | |
| "acc_stderr": 0.009491721291998515, | |
| "acc_norm": 0.6851851851851852, | |
| "acc_norm_stderr": 0.009530150430975602 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3515358361774744, | |
| "acc_stderr": 0.013952413699600942, | |
| "acc_norm": 0.36006825938566556, | |
| "acc_norm_stderr": 0.014027516814585183 | |
| }, | |
| "sciq": { | |
| "acc": 0.937, | |
| "acc_stderr": 0.007687007876286423, | |
| "acc_norm": 0.939, | |
| "acc_norm_stderr": 0.007572076091557425 | |
| }, | |
| "piqa": { | |
| "acc": 0.7720348204570185, | |
| "acc_stderr": 0.009788093832324908, | |
| "acc_norm": 0.7899891186071817, | |
| "acc_norm_stderr": 0.009503353305818581 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |