| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.348, | |
| "acc_stderr": 0.01507060460376841 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.346, | |
| "acc_stderr": 0.015050266127564448 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.355, | |
| "acc_stderr": 0.013819249004047308 | |
| }, | |
| "cb": { | |
| "acc": 0.30357142857142855, | |
| "acc_stderr": 0.06199938655510754, | |
| "f1": 0.24443052074631022 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.04229525846816506 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.43158733320055764, | |
| "acc_stderr": 0.00494285345937155, | |
| "acc_norm": 0.5638319059948218, | |
| "acc_norm_stderr": 0.004948952519517514 | |
| }, | |
| "rte": { | |
| "acc": 0.48736462093862815, | |
| "acc_stderr": 0.030086851767188564 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5469613259668509, | |
| "acc_stderr": 0.013990366632148088 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.692143238909674, | |
| "acc_stderr": 0.010674598158758179 | |
| }, | |
| "boolq": { | |
| "acc": 0.5776758409785933, | |
| "acc_stderr": 0.008638883260317736 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5774410774410774, | |
| "acc_stderr": 0.01013597822298108, | |
| "acc_norm": 0.555976430976431, | |
| "acc_norm_stderr": 0.010195285580783954 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.26535836177474403, | |
| "acc_stderr": 0.012902554762313964, | |
| "acc_norm": 0.2977815699658703, | |
| "acc_norm_stderr": 0.013363080107244487 | |
| }, | |
| "sciq": { | |
| "acc": 0.887, | |
| "acc_stderr": 0.01001655286669686, | |
| "acc_norm": 0.878, | |
| "acc_norm_stderr": 0.010354864712936701 | |
| }, | |
| "piqa": { | |
| "acc": 0.736126224156692, | |
| "acc_stderr": 0.010282996367695562, | |
| "acc_norm": 0.7404787812840044, | |
| "acc_norm_stderr": 0.010227939888173925 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |