| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.324, | |
| "acc_stderr": 0.01480686473373886 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.33, | |
| "acc_stderr": 0.014876872027456732 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3258333333333333, | |
| "acc_stderr": 0.013535422043417454 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.0663363415035954, | |
| "f1": 0.27708333333333335 | |
| }, | |
| "copa": { | |
| "acc": 0.82, | |
| "acc_stderr": 0.038612291966536955 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5340569607647879, | |
| "acc_stderr": 0.0049781928934062745, | |
| "acc_norm": 0.716391157140012, | |
| "acc_norm_stderr": 0.0044982802444945074 | |
| }, | |
| "rte": { | |
| "acc": 0.5703971119133574, | |
| "acc_stderr": 0.02979666882912467 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6345698500394633, | |
| "acc_stderr": 0.013533965097638788 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7594869053981828, | |
| "acc_stderr": 0.009883453084862687 | |
| }, | |
| "boolq": { | |
| "acc": 0.6629969418960244, | |
| "acc_stderr": 0.008267329046329363 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6957070707070707, | |
| "acc_stderr": 0.009441202922359185, | |
| "acc_norm": 0.6712962962962963, | |
| "acc_norm_stderr": 0.009638903167022168 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3455631399317406, | |
| "acc_stderr": 0.013896938461145687, | |
| "acc_norm": 0.3643344709897611, | |
| "acc_norm_stderr": 0.014063260279882413 | |
| }, | |
| "sciq": { | |
| "acc": 0.946, | |
| "acc_stderr": 0.007150883521295435, | |
| "acc_norm": 0.942, | |
| "acc_norm_stderr": 0.007395315455792937 | |
| }, | |
| "piqa": { | |
| "acc": 0.7758433079434167, | |
| "acc_stderr": 0.00972989795641006, | |
| "acc_norm": 0.7861806311207835, | |
| "acc_norm_stderr": 0.009565994206915607 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |