| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.328, | |
| "acc_stderr": 0.014853842487270334 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.321, | |
| "acc_stderr": 0.014770821817934649 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3491666666666667, | |
| "acc_stderr": 0.013767075395077249 | |
| }, | |
| "cb": { | |
| "acc": 0.4642857142857143, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.3013448230839535 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5251941844254132, | |
| "acc_stderr": 0.00498344288867777, | |
| "acc_norm": 0.6989643497311293, | |
| "acc_norm_stderr": 0.0045777070250313644 | |
| }, | |
| "rte": { | |
| "acc": 0.5342960288808665, | |
| "acc_stderr": 0.030025579819366426 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6108918705603789, | |
| "acc_stderr": 0.013702520871485945 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7504008551576697, | |
| "acc_stderr": 0.010008002459430844 | |
| }, | |
| "boolq": { | |
| "acc": 0.6489296636085627, | |
| "acc_stderr": 0.00834811495726361 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6742424242424242, | |
| "acc_stderr": 0.009616642976885964, | |
| "acc_norm": 0.6405723905723906, | |
| "acc_norm_stderr": 0.009845958893373752 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3250853242320819, | |
| "acc_stderr": 0.013688147309729124, | |
| "acc_norm": 0.35665529010238906, | |
| "acc_norm_stderr": 0.013998056902620199 | |
| }, | |
| "sciq": { | |
| "acc": 0.916, | |
| "acc_stderr": 0.008776162089491127, | |
| "acc_norm": 0.897, | |
| "acc_norm_stderr": 0.009616833339695792 | |
| }, | |
| "piqa": { | |
| "acc": 0.7747551686615887, | |
| "acc_stderr": 0.009746643471032145, | |
| "acc_norm": 0.779651795429815, | |
| "acc_norm_stderr": 0.00967053545685313 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |