| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.322, | |
| "acc_stderr": 0.014782913600996666 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.329, | |
| "acc_stderr": 0.014865395385928362 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.3408333333333333, | |
| "acc_stderr": 0.013688600793296932 | |
| }, | |
| "cb": { | |
| "acc": 0.35714285714285715, | |
| "acc_stderr": 0.0646095738380922, | |
| "f1": 0.22886002886002887 | |
| }, | |
| "copa": { | |
| "acc": 0.77, | |
| "acc_stderr": 0.04229525846816506 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.4501095399322844, | |
| "acc_stderr": 0.00496487956351331, | |
| "acc_norm": 0.5963951404102769, | |
| "acc_norm_stderr": 0.004896173035943317 | |
| }, | |
| "rte": { | |
| "acc": 0.5234657039711191, | |
| "acc_stderr": 0.03006330041190266 | |
| }, | |
| "winogrande": { | |
| "acc": 0.5785319652722968, | |
| "acc_stderr": 0.013878072377497606 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.7087119187600214, | |
| "acc_stderr": 0.010506919924163615 | |
| }, | |
| "boolq": { | |
| "acc": 0.5620795107033639, | |
| "acc_stderr": 0.008677388652709263 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.5904882154882155, | |
| "acc_stderr": 0.01009036816099006, | |
| "acc_norm": 0.5643939393939394, | |
| "acc_norm_stderr": 0.010174341733665226 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.26023890784982934, | |
| "acc_stderr": 0.012821930225112573, | |
| "acc_norm": 0.295221843003413, | |
| "acc_norm_stderr": 0.01332975029338232 | |
| }, | |
| "sciq": { | |
| "acc": 0.877, | |
| "acc_stderr": 0.010391293421849874, | |
| "acc_norm": 0.853, | |
| "acc_norm_stderr": 0.01120341539516033 | |
| }, | |
| "piqa": { | |
| "acc": 0.7546245919477693, | |
| "acc_stderr": 0.010039831320422396, | |
| "acc_norm": 0.7595212187159956, | |
| "acc_norm_stderr": 0.009971345364651062 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |