| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.334, | |
| "acc_stderr": 0.01492201952373296 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.341, | |
| "acc_stderr": 0.014998131348402697 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3625, | |
| "acc_stderr": 0.013883037874225516 | |
| }, | |
| "cb": { | |
| "acc": 0.3392857142857143, | |
| "acc_stderr": 0.06384226561930824, | |
| "f1": 0.2059178743961352 | |
| }, | |
| "copa": { | |
| "acc": 0.87, | |
| "acc_stderr": 0.03379976689896309 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5325632344154551, | |
| "acc_stderr": 0.004979188195338179, | |
| "acc_norm": 0.7046405098585939, | |
| "acc_norm_stderr": 0.004552718360513099 | |
| }, | |
| "rte": { | |
| "acc": 0.48014440433212996, | |
| "acc_stderr": 0.0300727231673172 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6187845303867403, | |
| "acc_stderr": 0.013650172164160305 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7482629609834314, | |
| "acc_stderr": 0.01003644434459808 | |
| }, | |
| "boolq": { | |
| "acc": 0.6333333333333333, | |
| "acc_stderr": 0.008428386213506826 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6616161616161617, | |
| "acc_stderr": 0.009709034670525096, | |
| "acc_norm": 0.5875420875420876, | |
| "acc_norm_stderr": 0.01010130544786476 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.31399317406143346, | |
| "acc_stderr": 0.013562691224726291, | |
| "acc_norm": 0.32849829351535836, | |
| "acc_norm_stderr": 0.013724978465537364 | |
| }, | |
| "sciq": { | |
| "acc": 0.877, | |
| "acc_stderr": 0.010391293421849877, | |
| "acc_norm": 0.795, | |
| "acc_norm_stderr": 0.01277255409611312 | |
| }, | |
| "piqa": { | |
| "acc": 0.7704026115342764, | |
| "acc_stderr": 0.009812682950815187, | |
| "acc_norm": 0.7850924918389554, | |
| "acc_norm_stderr": 0.009583665082653316 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |