| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.326, | |
| "acc_stderr": 0.014830507204541042 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.368, | |
| "acc_stderr": 0.015258073561521802 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.013579531277800922 | |
| }, | |
| "cb": { | |
| "acc": 0.44642857142857145, | |
| "acc_stderr": 0.06703189227942398, | |
| "f1": 0.3407114624505929 | |
| }, | |
| "copa": { | |
| "acc": 0.75, | |
| "acc_stderr": 0.04351941398892446 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.44124676359290976, | |
| "acc_stderr": 0.004955212787832385, | |
| "acc_norm": 0.572495518820952, | |
| "acc_norm_stderr": 0.004937054233711573 | |
| }, | |
| "rte": { | |
| "acc": 0.5415162454873647, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5445935280189423, | |
| "acc_stderr": 0.013996485037729794 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6916087653661144, | |
| "acc_stderr": 0.010679734445487796 | |
| }, | |
| "boolq": { | |
| "acc": 0.6079510703363914, | |
| "acc_stderr": 0.008538802914911997 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5812289562289562, | |
| "acc_stderr": 0.010123487160167812, | |
| "acc_norm": 0.5614478114478114, | |
| "acc_norm_stderr": 0.010182010275471115 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2645051194539249, | |
| "acc_stderr": 0.012889272949313368, | |
| "acc_norm": 0.2901023890784983, | |
| "acc_norm_stderr": 0.013261573677520767 | |
| }, | |
| "sciq": { | |
| "acc": 0.868, | |
| "acc_stderr": 0.010709373963528033, | |
| "acc_norm": 0.848, | |
| "acc_norm_stderr": 0.01135891830347529 | |
| }, | |
| "piqa": { | |
| "acc": 0.7219804134929271, | |
| "acc_stderr": 0.010453117358332811, | |
| "acc_norm": 0.7323177366702938, | |
| "acc_norm_stderr": 0.010330111189370423 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |