| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.319, | |
| "acc_stderr": 0.014746404865473487 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.331, | |
| "acc_stderr": 0.014888272588203931 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33666666666666667, | |
| "acc_stderr": 0.013647602942406387 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.0663363415035954, | |
| "f1": 0.22313546423135466 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.040201512610368445 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5329615614419438, | |
| "acc_stderr": 0.004978927164792885, | |
| "acc_norm": 0.707329217287393, | |
| "acc_norm_stderr": 0.004540586983229989 | |
| }, | |
| "rte": { | |
| "acc": 0.5090252707581228, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6227308602999211, | |
| "acc_stderr": 0.013622567928799501 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7562800641368252, | |
| "acc_stderr": 0.009928097887358885 | |
| }, | |
| "boolq": { | |
| "acc": 0.637308868501529, | |
| "acc_stderr": 0.008408838061823175 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6923400673400674, | |
| "acc_stderr": 0.009470292575831181, | |
| "acc_norm": 0.6531986531986532, | |
| "acc_norm_stderr": 0.009766326091716002 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.33361774744027306, | |
| "acc_stderr": 0.013778687054176545, | |
| "acc_norm": 0.34726962457337884, | |
| "acc_norm_stderr": 0.01391303452962044 | |
| }, | |
| "sciq": { | |
| "acc": 0.929, | |
| "acc_stderr": 0.008125578442487933, | |
| "acc_norm": 0.922, | |
| "acc_norm_stderr": 0.008484573530118588 | |
| }, | |
| "piqa": { | |
| "acc": 0.7834602829162133, | |
| "acc_stderr": 0.009609984714384606, | |
| "acc_norm": 0.7878128400435256, | |
| "acc_norm_stderr": 0.009539299828174056 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |