| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.364, | |
| "acc_stderr": 0.01522286884052202 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.348, | |
| "acc_stderr": 0.01507060460376841 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3566666666666667, | |
| "acc_stderr": 0.013833742805050715 | |
| }, | |
| "cb": { | |
| "acc": 0.5357142857142857, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.3263888888888889 | |
| }, | |
| "copa": { | |
| "acc": 0.64, | |
| "acc_stderr": 0.04824181513244218 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.3935471021708823, | |
| "acc_stderr": 0.004875379352079819, | |
| "acc_norm": 0.4924317864967138, | |
| "acc_norm_stderr": 0.00498920977074323 | |
| }, | |
| "rte": { | |
| "acc": 0.5667870036101083, | |
| "acc_stderr": 0.029826764082138277 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5153906866614049, | |
| "acc_stderr": 0.01404582678978366 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6493853554249065, | |
| "acc_stderr": 0.011034317290463294 | |
| }, | |
| "boolq": { | |
| "acc": 0.537308868501529, | |
| "acc_stderr": 0.00872067560638845 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.4831649831649832, | |
| "acc_stderr": 0.010253966261288895, | |
| "acc_norm": 0.4734848484848485, | |
| "acc_norm_stderr": 0.010245347015573706 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2363481228668942, | |
| "acc_stderr": 0.012414960524301842, | |
| "acc_norm": 0.2696245733788396, | |
| "acc_norm_stderr": 0.012968040686869148 | |
| }, | |
| "sciq": { | |
| "acc": 0.761, | |
| "acc_stderr": 0.013493000446937587, | |
| "acc_norm": 0.749, | |
| "acc_norm_stderr": 0.01371813351688892 | |
| }, | |
| "piqa": { | |
| "acc": 0.7007616974972797, | |
| "acc_stderr": 0.01068413067313458, | |
| "acc_norm": 0.6931447225244831, | |
| "acc_norm_stderr": 0.010760295070580381 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |