| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.307, | |
| "acc_stderr": 0.014593284892852634 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928366 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3616666666666667, | |
| "acc_stderr": 0.013876131663123877 | |
| }, | |
| "cb": { | |
| "acc": 0.4107142857142857, | |
| "acc_stderr": 0.0663363415035954, | |
| "f1": 0.28585166486356534 | |
| }, | |
| "copa": { | |
| "acc": 0.83, | |
| "acc_stderr": 0.037752516806863715 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4790878311093408, | |
| "acc_stderr": 0.0049854152506909125, | |
| "acc_norm": 0.6355307707627963, | |
| "acc_norm_stderr": 0.004802974070507194 | |
| }, | |
| "rte": { | |
| "acc": 0.5234657039711191, | |
| "acc_stderr": 0.03006330041190266 | |
| } | |
| }, | |
| "versions": { | |
| "anli_r1": 0, | |
| "anli_r2": 0, | |
| "anli_r3": 0, | |
| "cb": 1, | |
| "copa": 0, | |
| "hellaswag": 0, | |
| "rte": 0 | |
| } | |
| } |