| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.34, | |
| "acc_stderr": 0.014987482264363935 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.337, | |
| "acc_stderr": 0.014955087918653603 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.35, | |
| "acc_stderr": 0.013774667009018552 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.06633634150359541, | |
| "f1": 0.29090909090909095 | |
| }, | |
| "copa": { | |
| "acc": 0.8, | |
| "acc_stderr": 0.040201512610368445 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4698267277434774, | |
| "acc_stderr": 0.004980687467486101, | |
| "acc_norm": 0.6165106552479586, | |
| "acc_norm_stderr": 0.004852420856631488 | |
| }, | |
| "rte": { | |
| "acc": 0.5234657039711191, | |
| "acc_stderr": 0.03006330041190266 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5864246250986582, | |
| "acc_stderr": 0.013840971763195303 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7108498129342598, | |
| "acc_stderr": 0.010484068799942061 | |
| }, | |
| "boolq": { | |
| "acc": 0.5975535168195719, | |
| "acc_stderr": 0.008576992126012484 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.593013468013468, | |
| "acc_stderr": 0.010080695355466598, | |
| "acc_norm": 0.5568181818181818, | |
| "acc_norm_stderr": 0.010193324837773497 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.27474402730375425, | |
| "acc_stderr": 0.013044617212771227, | |
| "acc_norm": 0.30802047781569963, | |
| "acc_norm_stderr": 0.01349142951729204 | |
| }, | |
| "sciq": { | |
| "acc": 0.855, | |
| "acc_stderr": 0.011139977517890132, | |
| "acc_norm": 0.814, | |
| "acc_norm_stderr": 0.0123107902084128 | |
| }, | |
| "piqa": { | |
| "acc": 0.7513601741022851, | |
| "acc_stderr": 0.010084511234296857, | |
| "acc_norm": 0.76550598476605, | |
| "acc_norm_stderr": 0.009885203143240536 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |