| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.35, | |
| "acc_stderr": 0.015090650341444235 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.339, | |
| "acc_stderr": 0.014976758771620339 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3566666666666667, | |
| "acc_stderr": 0.013833742805050713 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.2631578947368421 | |
| }, | |
| "copa": { | |
| "acc": 0.76, | |
| "acc_stderr": 0.04292346959909283 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4500099581756622, | |
| "acc_stderr": 0.004964779805180658, | |
| "acc_norm": 0.5825532762397929, | |
| "acc_norm_stderr": 0.00492130033128556 | |
| }, | |
| "rte": { | |
| "acc": 0.5126353790613718, | |
| "acc_stderr": 0.030086851767188564 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5509076558800315, | |
| "acc_stderr": 0.013979459389140844 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7006948156066275, | |
| "acc_stderr": 0.010590117252248801 | |
| }, | |
| "boolq": { | |
| "acc": 0.5651376146788991, | |
| "acc_stderr": 0.008670528471841557 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5917508417508418, | |
| "acc_stderr": 0.01008556619579125, | |
| "acc_norm": 0.5340909090909091, | |
| "acc_norm_stderr": 0.010235908103438688 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2440273037542662, | |
| "acc_stderr": 0.012551447627856259, | |
| "acc_norm": 0.2883959044368601, | |
| "acc_norm_stderr": 0.013238394422428175 | |
| }, | |
| "sciq": { | |
| "acc": 0.857, | |
| "acc_stderr": 0.01107581480856704, | |
| "acc_norm": 0.769, | |
| "acc_norm_stderr": 0.013334797216936442 | |
| }, | |
| "piqa": { | |
| "acc": 0.7366702937976061, | |
| "acc_stderr": 0.010276185322196764, | |
| "acc_norm": 0.7464635473340587, | |
| "acc_norm_stderr": 0.010150090834551794 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |