| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.333, | |
| "acc_stderr": 0.014910846164229868 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.326, | |
| "acc_stderr": 0.01483050720454104 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3475, | |
| "acc_stderr": 0.013751753243291852 | |
| }, | |
| "cb": { | |
| "acc": 0.5357142857142857, | |
| "acc_stderr": 0.06724777654937658, | |
| "f1": 0.37227304714989445 | |
| }, | |
| "copa": { | |
| "acc": 0.79, | |
| "acc_stderr": 0.040936018074033256 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.47191794463254333, | |
| "acc_stderr": 0.004981905293878145, | |
| "acc_norm": 0.6139215295757817, | |
| "acc_norm_stderr": 0.004858539527872466 | |
| }, | |
| "rte": { | |
| "acc": 0.5703971119133574, | |
| "acc_stderr": 0.029796668829124674 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5706393054459353, | |
| "acc_stderr": 0.013911537499969163 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7151256012827365, | |
| "acc_stderr": 0.01043751398661172 | |
| }, | |
| "boolq": { | |
| "acc": 0.5669724770642202, | |
| "acc_stderr": 0.00866625130551806 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5913299663299664, | |
| "acc_stderr": 0.010087174498762883, | |
| "acc_norm": 0.5496632996632996, | |
| "acc_norm_stderr": 0.010209047724374145 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.2627986348122867, | |
| "acc_stderr": 0.012862523175351333, | |
| "acc_norm": 0.30716723549488056, | |
| "acc_norm_stderr": 0.013481034054980943 | |
| }, | |
| "sciq": { | |
| "acc": 0.836, | |
| "acc_stderr": 0.011715000693181331, | |
| "acc_norm": 0.781, | |
| "acc_norm_stderr": 0.013084731950262012 | |
| }, | |
| "piqa": { | |
| "acc": 0.7448313384113167, | |
| "acc_stderr": 0.010171571592521822, | |
| "acc_norm": 0.7535364526659413, | |
| "acc_norm_stderr": 0.01005481078967181 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |