| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.339, | |
| "acc_stderr": 0.014976758771620342 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.347, | |
| "acc_stderr": 0.015060472031706615 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3425, | |
| "acc_stderr": 0.013704669762934722 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.2429169746242917 | |
| }, | |
| "copa": { | |
| "acc": 0.7, | |
| "acc_stderr": 0.046056618647183814 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.3819956184027086, | |
| "acc_stderr": 0.004848824710995933, | |
| "acc_norm": 0.483469428400717, | |
| "acc_norm_stderr": 0.0049870536525402735 | |
| }, | |
| "rte": { | |
| "acc": 0.5631768953068592, | |
| "acc_stderr": 0.02985524739031494 | |
| }, | |
| "winogrande": { | |
| "acc": 0.526440410418311, | |
| "acc_stderr": 0.014032823874407225 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.6606092998396579, | |
| "acc_stderr": 0.010949682016358629 | |
| }, | |
| "boolq": { | |
| "acc": 0.5547400611620795, | |
| "acc_stderr": 0.008692488322023063 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5711279461279462, | |
| "acc_stderr": 0.010155440652900154, | |
| "acc_norm": 0.5467171717171717, | |
| "acc_norm_stderr": 0.010214901516731609 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2551194539249147, | |
| "acc_stderr": 0.012739038695202102, | |
| "acc_norm": 0.26621160409556316, | |
| "acc_norm_stderr": 0.012915774781523214 | |
| }, | |
| "sciq": { | |
| "acc": 0.899, | |
| "acc_stderr": 0.009533618929340995, | |
| "acc_norm": 0.903, | |
| "acc_norm_stderr": 0.009363689373248121 | |
| }, | |
| "piqa": { | |
| "acc": 0.6953210010881393, | |
| "acc_stderr": 0.010738889044325161, | |
| "acc_norm": 0.7029379760609358, | |
| "acc_norm_stderr": 0.010661725404814783 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |