| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.326, | |
| "acc_stderr": 0.014830507204541056 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.321, | |
| "acc_stderr": 0.014770821817934644 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3308333333333333, | |
| "acc_stderr": 0.013588208070708997 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.2692307692307692 | |
| }, | |
| "copa": { | |
| "acc": 0.75, | |
| "acc_stderr": 0.04351941398892446 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4715196176060546, | |
| "acc_stderr": 0.004981680090303695, | |
| "acc_norm": 0.6208922525393348, | |
| "acc_norm_stderr": 0.004841734453506668 | |
| }, | |
| "rte": { | |
| "acc": 0.48375451263537905, | |
| "acc_stderr": 0.030080573208738064 | |
| }, | |
| "winogrande": { | |
| "acc": 0.574585635359116, | |
| "acc_stderr": 0.013895257666646382 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7183324425440941, | |
| "acc_stderr": 0.010401844358587665 | |
| }, | |
| "boolq": { | |
| "acc": 0.6039755351681957, | |
| "acc_stderr": 0.008553881336813413 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6191077441077442, | |
| "acc_stderr": 0.009964428212260372, | |
| "acc_norm": 0.5921717171717171, | |
| "acc_norm_stderr": 0.010083950240041223 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.302901023890785, | |
| "acc_stderr": 0.013428241573185349, | |
| "acc_norm": 0.3174061433447099, | |
| "acc_norm_stderr": 0.01360223908803817 | |
| }, | |
| "sciq": { | |
| "acc": 0.899, | |
| "acc_stderr": 0.009533618929340997, | |
| "acc_norm": 0.877, | |
| "acc_norm_stderr": 0.010391293421849877 | |
| }, | |
| "piqa": { | |
| "acc": 0.7524483133841132, | |
| "acc_stderr": 0.010069703966857104, | |
| "acc_norm": 0.7633297062023939, | |
| "acc_norm_stderr": 0.009916841655042809 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |