| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.293, | |
| "acc_stderr": 0.014399942998441271 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.327, | |
| "acc_stderr": 0.01484221315341124 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3383333333333333, | |
| "acc_stderr": 0.013664144006618275 | |
| }, | |
| "cb": { | |
| "acc": 0.32142857142857145, | |
| "acc_stderr": 0.06297362289056341, | |
| "f1": 0.2706949089557785 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.042295258468165065 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5265883290181239, | |
| "acc_stderr": 0.0049827214724073405, | |
| "acc_norm": 0.7029476199960167, | |
| "acc_norm_stderr": 0.00456025908319738 | |
| }, | |
| "rte": { | |
| "acc": 0.5595667870036101, | |
| "acc_stderr": 0.029882123363118723 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6243093922651933, | |
| "acc_stderr": 0.013611257508380444 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7413148049171566, | |
| "acc_stderr": 0.010126662138021714 | |
| }, | |
| "boolq": { | |
| "acc": 0.6669724770642201, | |
| "acc_stderr": 0.00824302391268888 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.6759259259259259, | |
| "acc_stderr": 0.009603728850095394, | |
| "acc_norm": 0.640993265993266, | |
| "acc_norm_stderr": 0.009843424713072176 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.33447098976109213, | |
| "acc_stderr": 0.013787460322441384, | |
| "acc_norm": 0.3438566552901024, | |
| "acc_norm_stderr": 0.01388064457015621 | |
| }, | |
| "sciq": { | |
| "acc": 0.928, | |
| "acc_stderr": 0.008178195576218681, | |
| "acc_norm": 0.911, | |
| "acc_norm_stderr": 0.009008893392651523 | |
| }, | |
| "piqa": { | |
| "acc": 0.7763873775843307, | |
| "acc_stderr": 0.009721489519176294, | |
| "acc_norm": 0.7883569096844396, | |
| "acc_norm_stderr": 0.009530351270479392 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |