| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.352, | |
| "acc_stderr": 0.015110404505648661 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.014842213153411242 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.013579531277800925 | |
| }, | |
| "cb": { | |
| "acc": 0.5, | |
| "acc_stderr": 0.06741998624632421, | |
| "f1": 0.355846042120552 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.04229525846816506 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.45130452101175067, | |
| "acc_stderr": 0.0049660609953150634, | |
| "acc_norm": 0.5881298546106354, | |
| "acc_norm_stderr": 0.0049116598845061485 | |
| }, | |
| "rte": { | |
| "acc": 0.5306859205776173, | |
| "acc_stderr": 0.030039730592197812 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5493291239147593, | |
| "acc_stderr": 0.01398392886904024 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6819882415820417, | |
| "acc_stderr": 0.010769343495248548 | |
| }, | |
| "boolq": { | |
| "acc": 0.5834862385321101, | |
| "acc_stderr": 0.008622288020674008 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6262626262626263, | |
| "acc_stderr": 0.009927267058259626, | |
| "acc_norm": 0.6123737373737373, | |
| "acc_norm_stderr": 0.009997307914447608 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.28242320819112626, | |
| "acc_stderr": 0.013155456884097222, | |
| "acc_norm": 0.30716723549488056, | |
| "acc_norm_stderr": 0.013481034054980945 | |
| }, | |
| "sciq": { | |
| "acc": 0.903, | |
| "acc_stderr": 0.009363689373248123, | |
| "acc_norm": 0.891, | |
| "acc_norm_stderr": 0.009859828407037186 | |
| }, | |
| "piqa": { | |
| "acc": 0.7377584330794341, | |
| "acc_stderr": 0.010262502565172447, | |
| "acc_norm": 0.7404787812840044, | |
| "acc_norm_stderr": 0.010227939888173929 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |