| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.347, | |
| "acc_stderr": 0.015060472031706622 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.361, | |
| "acc_stderr": 0.015195720118175127 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.013680495725767787 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.266719222178426 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.47400916152160927, | |
| "acc_stderr": 0.004983035420235718, | |
| "acc_norm": 0.6275642302330213, | |
| "acc_norm_stderr": 0.004824655406075563 | |
| }, | |
| "rte": { | |
| "acc": 0.4981949458483754, | |
| "acc_stderr": 0.030096267148976626 | |
| }, | |
| "winogrande": { | |
| "acc": 0.584846093133386, | |
| "acc_stderr": 0.013848684086658585 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7247461250668092, | |
| "acc_stderr": 0.010328538400500567 | |
| }, | |
| "boolq": { | |
| "acc": 0.5984709480122324, | |
| "acc_stderr": 0.008573784490094754 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6283670033670034, | |
| "acc_stderr": 0.009915897123658788, | |
| "acc_norm": 0.5972222222222222, | |
| "acc_norm_stderr": 0.010063960494989163 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.30119453924914674, | |
| "acc_stderr": 0.01340674176784762, | |
| "acc_norm": 0.3370307167235495, | |
| "acc_norm_stderr": 0.013813476652902269 | |
| }, | |
| "sciq": { | |
| "acc": 0.911, | |
| "acc_stderr": 0.009008893392651518, | |
| "acc_norm": 0.898, | |
| "acc_norm_stderr": 0.009575368801653892 | |
| }, | |
| "piqa": { | |
| "acc": 0.7546245919477693, | |
| "acc_stderr": 0.010039831320422396, | |
| "acc_norm": 0.766050054406964, | |
| "acc_norm_stderr": 0.009877236895137446 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |