| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.319, | |
| "acc_stderr": 0.014746404865473486 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.01487687202745673 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.32, | |
| "acc_stderr": 0.013471620929769142 | |
| }, | |
| "cb": { | |
| "acc": 0.42857142857142855, | |
| "acc_stderr": 0.06672848092813058, | |
| "f1": 0.3049446887911502 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5259908384783908, | |
| "acc_stderr": 0.004983035420235712, | |
| "acc_norm": 0.7002589125672177, | |
| "acc_norm_stderr": 0.004572081656965643 | |
| }, | |
| "rte": { | |
| "acc": 0.5812274368231047, | |
| "acc_stderr": 0.02969666108123484 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6266771902131019, | |
| "acc_stderr": 0.013594002763035516 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7616247995724211, | |
| "acc_stderr": 0.009853267441685421 | |
| }, | |
| "boolq": { | |
| "acc": 0.6590214067278287, | |
| "acc_stderr": 0.00829097981816109 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6784511784511784, | |
| "acc_stderr": 0.00958409157564062, | |
| "acc_norm": 0.6607744107744108, | |
| "acc_norm_stderr": 0.00971491720776585 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.33276450511945393, | |
| "acc_stderr": 0.01376986304619231, | |
| "acc_norm": 0.3643344709897611, | |
| "acc_norm_stderr": 0.014063260279882415 | |
| }, | |
| "sciq": { | |
| "acc": 0.93, | |
| "acc_stderr": 0.008072494358323499, | |
| "acc_norm": 0.91, | |
| "acc_norm_stderr": 0.009054390204866442 | |
| }, | |
| "piqa": { | |
| "acc": 0.7780195865070729, | |
| "acc_stderr": 0.009696120744662019, | |
| "acc_norm": 0.7872687704026116, | |
| "acc_norm_stderr": 0.009548223123047352 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |