| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.01492201952373297 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.336, | |
| "acc_stderr": 0.01494414023379502 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33916666666666667, | |
| "acc_stderr": 0.013672343491681819 | |
| }, | |
| "cb": { | |
| "acc": 0.19642857142857142, | |
| "acc_stderr": 0.05357142857142859, | |
| "f1": 0.18920723969812325 | |
| }, | |
| "copa": { | |
| "acc": 0.85, | |
| "acc_stderr": 0.03588702812826373 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5288787094204341, | |
| "acc_stderr": 0.004981451704451047, | |
| "acc_norm": 0.7052380003983271, | |
| "acc_norm_stderr": 0.0045500389685506236 | |
| }, | |
| "rte": { | |
| "acc": 0.4729241877256318, | |
| "acc_stderr": 0.030052303463143706 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6172059984214681, | |
| "acc_stderr": 0.013660946109442013 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7402458578300374, | |
| "acc_stderr": 0.010140244588689848 | |
| }, | |
| "boolq": { | |
| "acc": 0.6351681957186545, | |
| "acc_stderr": 0.00841944098496365 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6856060606060606, | |
| "acc_stderr": 0.009526702423162905, | |
| "acc_norm": 0.6426767676767676, | |
| "acc_norm_stderr": 0.009833205612463107 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.33361774744027306, | |
| "acc_stderr": 0.01377868705417654, | |
| "acc_norm": 0.34897610921501704, | |
| "acc_norm_stderr": 0.013928933461382496 | |
| }, | |
| "sciq": { | |
| "acc": 0.922, | |
| "acc_stderr": 0.008484573530118581, | |
| "acc_norm": 0.897, | |
| "acc_norm_stderr": 0.00961683333969579 | |
| }, | |
| "piqa": { | |
| "acc": 0.7627856365614799, | |
| "acc_stderr": 0.009924694933586364, | |
| "acc_norm": 0.7742110990206746, | |
| "acc_norm_stderr": 0.009754980670917315 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |