| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932575 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.328, | |
| "acc_stderr": 0.014853842487270334 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3433333333333333, | |
| "acc_stderr": 0.01371263383046586 | |
| }, | |
| "cb": { | |
| "acc": 0.14285714285714285, | |
| "acc_stderr": 0.047184161362558305, | |
| "f1": 0.13156966490299823 | |
| }, | |
| "copa": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.03684529491774709 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5320653256323441, | |
| "acc_stderr": 0.00497951000177662, | |
| "acc_norm": 0.7050388368850826, | |
| "acc_norm_stderr": 0.004550933142528758 | |
| }, | |
| "rte": { | |
| "acc": 0.49097472924187724, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6148382004735596, | |
| "acc_stderr": 0.013676821287521413 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7536076964190273, | |
| "acc_stderr": 0.009964727533753546 | |
| }, | |
| "boolq": { | |
| "acc": 0.6440366972477064, | |
| "acc_stderr": 0.008374337517726581 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6957070707070707, | |
| "acc_stderr": 0.009441202922359183, | |
| "acc_norm": 0.6717171717171717, | |
| "acc_norm_stderr": 0.00963574950926216 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3378839590443686, | |
| "acc_stderr": 0.013822047922283516, | |
| "acc_norm": 0.3643344709897611, | |
| "acc_norm_stderr": 0.014063260279882412 | |
| }, | |
| "sciq": { | |
| "acc": 0.931, | |
| "acc_stderr": 0.008018934050315155, | |
| "acc_norm": 0.922, | |
| "acc_norm_stderr": 0.008484573530118587 | |
| }, | |
| "piqa": { | |
| "acc": 0.7736670293797606, | |
| "acc_stderr": 0.009763294246879427, | |
| "acc_norm": 0.7845484221980413, | |
| "acc_norm_stderr": 0.009592463115658107 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |