| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.338, | |
| "acc_stderr": 0.014965960710224489 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.332, | |
| "acc_stderr": 0.014899597242811476 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3325, | |
| "acc_stderr": 0.013605417345710526 | |
| }, | |
| "cb": { | |
| "acc": 0.08928571428571429, | |
| "acc_stderr": 0.038450387280282494, | |
| "f1": 0.0871517027863777 | |
| }, | |
| "copa": { | |
| "acc": 0.83, | |
| "acc_stderr": 0.03775251680686371 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4525990838478391, | |
| "acc_stderr": 0.0049673082544257514, | |
| "acc_norm": 0.5948018323043218, | |
| "acc_norm_stderr": 0.004899270310557971 | |
| }, | |
| "rte": { | |
| "acc": 0.49458483754512633, | |
| "acc_stderr": 0.03009469812323996 | |
| }, | |
| "winogrande": { | |
| "acc": 0.585635359116022, | |
| "acc_stderr": 0.01384484623226856 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7135221806520577, | |
| "acc_stderr": 0.01045510591863303 | |
| }, | |
| "boolq": { | |
| "acc": 0.6079510703363914, | |
| "acc_stderr": 0.008538802914911992 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5765993265993266, | |
| "acc_stderr": 0.010138671005289047, | |
| "acc_norm": 0.5585016835016835, | |
| "acc_norm_stderr": 0.010189314382749929 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.27986348122866894, | |
| "acc_stderr": 0.01311904089772592, | |
| "acc_norm": 0.29180887372013653, | |
| "acc_norm_stderr": 0.013284525292403506 | |
| }, | |
| "sciq": { | |
| "acc": 0.908, | |
| "acc_stderr": 0.009144376393151086, | |
| "acc_norm": 0.906, | |
| "acc_norm_stderr": 0.009233052000787738 | |
| }, | |
| "piqa": { | |
| "acc": 0.7257889009793254, | |
| "acc_stderr": 0.010408618664933382, | |
| "acc_norm": 0.7334058759521219, | |
| "acc_norm_stderr": 0.010316749863541365 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |