| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.344, | |
| "acc_stderr": 0.015029633724408943 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.326, | |
| "acc_stderr": 0.01483050720454104 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.36083333333333334, | |
| "acc_stderr": 0.01386918025244486 | |
| }, | |
| "cb": { | |
| "acc": 0.6428571428571429, | |
| "acc_stderr": 0.06460957383809221, | |
| "f1": 0.4430260047281324 | |
| }, | |
| "copa": { | |
| "acc": 0.76, | |
| "acc_stderr": 0.04292346959909282 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.466938856801434, | |
| "acc_stderr": 0.004978861409119807, | |
| "acc_norm": 0.6233817964548894, | |
| "acc_norm_stderr": 0.0048354759576109425 | |
| }, | |
| "rte": { | |
| "acc": 0.5667870036101083, | |
| "acc_stderr": 0.02982676408213827 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6029992107340174, | |
| "acc_stderr": 0.0137510925198067 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7284874398717264, | |
| "acc_stderr": 0.010284547617192592 | |
| }, | |
| "boolq": { | |
| "acc": 0.617737003058104, | |
| "acc_stderr": 0.008499149690449273 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6334175084175084, | |
| "acc_stderr": 0.009887786585323946, | |
| "acc_norm": 0.617003367003367, | |
| "acc_norm_stderr": 0.009974920384536482 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.30631399317406144, | |
| "acc_stderr": 0.013470584417276511, | |
| "acc_norm": 0.3225255972696246, | |
| "acc_norm_stderr": 0.013659980894277373 | |
| }, | |
| "sciq": { | |
| "acc": 0.909, | |
| "acc_stderr": 0.009099549538400243, | |
| "acc_norm": 0.903, | |
| "acc_norm_stderr": 0.009363689373248125 | |
| }, | |
| "piqa": { | |
| "acc": 0.7448313384113167, | |
| "acc_stderr": 0.010171571592521822, | |
| "acc_norm": 0.7616974972796517, | |
| "acc_norm_stderr": 0.009940334245876219 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |