| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.348, | |
| "acc_stderr": 0.015070604603768408 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.333, | |
| "acc_stderr": 0.014910846164229857 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33916666666666667, | |
| "acc_stderr": 0.013672343491681815 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.1986111111111111 | |
| }, | |
| "copa": { | |
| "acc": 0.76, | |
| "acc_stderr": 0.04292346959909282 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.42182832105158335, | |
| "acc_stderr": 0.0049284209030265504, | |
| "acc_norm": 0.5407289384584744, | |
| "acc_norm_stderr": 0.004973199296339958 | |
| }, | |
| "rte": { | |
| "acc": 0.5415162454873647, | |
| "acc_stderr": 0.029992535385373314 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5359116022099447, | |
| "acc_stderr": 0.01401619343395831 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6841261357562801, | |
| "acc_stderr": 0.010749892827011111 | |
| }, | |
| "boolq": { | |
| "acc": 0.563914373088685, | |
| "acc_stderr": 0.00867331277632493 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5521885521885522, | |
| "acc_stderr": 0.010203742451111532, | |
| "acc_norm": 0.49242424242424243, | |
| "acc_norm_stderr": 0.010258605792153321 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.24573378839590443, | |
| "acc_stderr": 0.012581033453730107, | |
| "acc_norm": 0.2781569965870307, | |
| "acc_norm_stderr": 0.013094469919538805 | |
| }, | |
| "sciq": { | |
| "acc": 0.797, | |
| "acc_stderr": 0.012726073744598283, | |
| "acc_norm": 0.719, | |
| "acc_norm_stderr": 0.01422115470843493 | |
| }, | |
| "piqa": { | |
| "acc": 0.7257889009793254, | |
| "acc_stderr": 0.010408618664933382, | |
| "acc_norm": 0.7377584330794341, | |
| "acc_norm_stderr": 0.010262502565172442 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |