| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.317, | |
| "acc_stderr": 0.01472167543888022 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.014955087918653605 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33416666666666667, | |
| "acc_stderr": 0.013622434813136783 | |
| }, | |
| "cb": { | |
| "acc": 0.4642857142857143, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.38713450292397655 | |
| }, | |
| "copa": { | |
| "acc": 0.83, | |
| "acc_stderr": 0.03775251680686371 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5299741087432782, | |
| "acc_stderr": 0.004980807231136743, | |
| "acc_norm": 0.7079267078271261, | |
| "acc_norm_stderr": 0.004537865171414025 | |
| }, | |
| "rte": { | |
| "acc": 0.5884476534296029, | |
| "acc_stderr": 0.029621832222417196 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6353591160220995, | |
| "acc_stderr": 0.013527746622429837 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7659005879208979, | |
| "acc_stderr": 0.009791868211495304 | |
| }, | |
| "boolq": { | |
| "acc": 0.65565749235474, | |
| "acc_stderr": 0.008310485054782981 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6813973063973064, | |
| "acc_stderr": 0.009560775507673366, | |
| "acc_norm": 0.6641414141414141, | |
| "acc_norm_stderr": 0.009691180932083508 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3310580204778157, | |
| "acc_stderr": 0.013752062419817836, | |
| "acc_norm": 0.36689419795221845, | |
| "acc_norm_stderr": 0.014084133118104292 | |
| }, | |
| "sciq": { | |
| "acc": 0.939, | |
| "acc_stderr": 0.007572076091557426, | |
| "acc_norm": 0.926, | |
| "acc_norm_stderr": 0.008282064512704156 | |
| }, | |
| "piqa": { | |
| "acc": 0.7845484221980413, | |
| "acc_stderr": 0.009592463115658117, | |
| "acc_norm": 0.7899891186071817, | |
| "acc_norm_stderr": 0.009503353305818578 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |