| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.306, | |
| "acc_stderr": 0.01458000605543697 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.323, | |
| "acc_stderr": 0.01479492784334864 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3458333333333333, | |
| "acc_stderr": 0.013736245342311014 | |
| }, | |
| "cb": { | |
| "acc": 0.32142857142857145, | |
| "acc_stderr": 0.06297362289056341, | |
| "f1": 0.2300556586270872 | |
| }, | |
| "copa": { | |
| "acc": 0.84, | |
| "acc_stderr": 0.0368452949177471 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5312686715793666, | |
| "acc_stderr": 0.004980014536539822, | |
| "acc_norm": 0.7075283808006373, | |
| "acc_norm_stderr": 0.004539680764142159 | |
| }, | |
| "rte": { | |
| "acc": 0.5090252707581228, | |
| "acc_stderr": 0.030091559826331334 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6353591160220995, | |
| "acc_stderr": 0.013527746622429842 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7530732228754676, | |
| "acc_stderr": 0.009971991360388979 | |
| }, | |
| "boolq": { | |
| "acc": 0.6605504587155964, | |
| "acc_stderr": 0.008281960446071344 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6856060606060606, | |
| "acc_stderr": 0.009526702423162905, | |
| "acc_norm": 0.6590909090909091, | |
| "acc_norm_stderr": 0.009726579593424019 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3412969283276451, | |
| "acc_stderr": 0.013855831287497717, | |
| "acc_norm": 0.34982935153583616, | |
| "acc_norm_stderr": 0.013936809212158275 | |
| }, | |
| "sciq": { | |
| "acc": 0.944, | |
| "acc_stderr": 0.007274401481697061, | |
| "acc_norm": 0.93, | |
| "acc_norm_stderr": 0.008072494358323497 | |
| }, | |
| "piqa": { | |
| "acc": 0.780195865070729, | |
| "acc_stderr": 0.009661958616651768, | |
| "acc_norm": 0.7889009793253536, | |
| "acc_norm_stderr": 0.009521377378734146 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |