| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.308, | |
| "acc_stderr": 0.014606483127342763 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928373 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.31916666666666665, | |
| "acc_stderr": 0.013462309712005124 | |
| }, | |
| "cb": { | |
| "acc": 0.5714285714285714, | |
| "acc_stderr": 0.06672848092813059, | |
| "f1": 0.40095238095238095 | |
| }, | |
| "copa": { | |
| "acc": 0.81, | |
| "acc_stderr": 0.03942772444036623 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.46634136626170086, | |
| "acc_stderr": 0.004978462690966927, | |
| "acc_norm": 0.6188010356502689, | |
| "acc_norm_stderr": 0.00484688692976345 | |
| }, | |
| "rte": { | |
| "acc": 0.5703971119133574, | |
| "acc_stderr": 0.02979666882912467 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5674822415153907, | |
| "acc_stderr": 0.013923911578623837 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7226082308925709, | |
| "acc_stderr": 0.010353267472010767 | |
| }, | |
| "boolq": { | |
| "acc": 0.6168195718654435, | |
| "acc_stderr": 0.008503021391450788 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6308922558922558, | |
| "acc_stderr": 0.009901987410242747, | |
| "acc_norm": 0.6136363636363636, | |
| "acc_norm_stderr": 0.009991296778159615 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.28668941979522183, | |
| "acc_stderr": 0.01321498632927477, | |
| "acc_norm": 0.310580204778157, | |
| "acc_norm_stderr": 0.013522292098053057 | |
| }, | |
| "sciq": { | |
| "acc": 0.915, | |
| "acc_stderr": 0.008823426366942331, | |
| "acc_norm": 0.903, | |
| "acc_norm_stderr": 0.009363689373248123 | |
| }, | |
| "piqa": { | |
| "acc": 0.7464635473340587, | |
| "acc_stderr": 0.010150090834551788, | |
| "acc_norm": 0.7589771490750816, | |
| "acc_norm_stderr": 0.009979042717267314 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |