| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.29, | |
| "acc_stderr": 0.014356395999905697 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.333, | |
| "acc_stderr": 0.014910846164229868 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3358333333333333, | |
| "acc_stderr": 0.013639261190932873 | |
| }, | |
| "cb": { | |
| "acc": 0.17857142857142858, | |
| "acc_stderr": 0.051642771820087224, | |
| "f1": 0.1770273087346258 | |
| }, | |
| "copa": { | |
| "acc": 0.83, | |
| "acc_stderr": 0.037752516806863715 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5314678350926111, | |
| "acc_stderr": 0.004979889597551663, | |
| "acc_norm": 0.7079267078271261, | |
| "acc_norm_stderr": 0.0045378651714140265 | |
| }, | |
| "rte": { | |
| "acc": 0.5126353790613718, | |
| "acc_stderr": 0.030086851767188564 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6495659037095501, | |
| "acc_stderr": 0.013409047676670187 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.757883484767504, | |
| "acc_stderr": 0.009905870033193868 | |
| }, | |
| "boolq": { | |
| "acc": 0.6584097859327217, | |
| "acc_stderr": 0.008294560677768499 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6830808080808081, | |
| "acc_stderr": 0.009547254611446381, | |
| "acc_norm": 0.6654040404040404, | |
| "acc_norm_stderr": 0.009682137724327909 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3361774744027304, | |
| "acc_stderr": 0.013804855026205756, | |
| "acc_norm": 0.3583617747440273, | |
| "acc_norm_stderr": 0.014012883334859868 | |
| }, | |
| "sciq": { | |
| "acc": 0.943, | |
| "acc_stderr": 0.0073351758537068225, | |
| "acc_norm": 0.934, | |
| "acc_norm_stderr": 0.007855297938697587 | |
| }, | |
| "piqa": { | |
| "acc": 0.7720348204570185, | |
| "acc_stderr": 0.00978809383232491, | |
| "acc_norm": 0.7780195865070729, | |
| "acc_norm_stderr": 0.00969612074466202 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |