| { | |
| "results": { | |
| "anli_r1": { | |
| "acc": 0.32, | |
| "acc_stderr": 0.01475865230357487 | |
| }, | |
| "anli_r2": { | |
| "acc": 0.335, | |
| "acc_stderr": 0.014933117490932566 | |
| }, | |
| "anli_r3": { | |
| "acc": 0.33916666666666667, | |
| "acc_stderr": 0.013672343491681812 | |
| }, | |
| "cb": { | |
| "acc": 0.375, | |
| "acc_stderr": 0.06527912098338669, | |
| "f1": 0.28595317725752506 | |
| }, | |
| "copa": { | |
| "acc": 0.89, | |
| "acc_stderr": 0.03144660377352203 | |
| }, | |
| "hellaswag": { | |
| "acc": 0.5285799641505676, | |
| "acc_stderr": 0.004981623292196192, | |
| "acc_norm": 0.7057359091814379, | |
| "acc_norm_stderr": 0.00454779896412668 | |
| }, | |
| "rte": { | |
| "acc": 0.5451263537906137, | |
| "acc_stderr": 0.029973636495415255 | |
| }, | |
| "winogrande": { | |
| "acc": 0.6527229676400947, | |
| "acc_stderr": 0.013380909249751242 | |
| }, | |
| "storycloze_2016": { | |
| "acc": 0.757883484767504, | |
| "acc_stderr": 0.009905870033193868 | |
| }, | |
| "boolq": { | |
| "acc": 0.6464831804281346, | |
| "acc_stderr": 0.008361346005339394 | |
| }, | |
| "arc_easy": { | |
| "acc": 0.672979797979798, | |
| "acc_stderr": 0.009626235849372207, | |
| "acc_norm": 0.6553030303030303, | |
| "acc_norm_stderr": 0.009752321586569784 | |
| }, | |
| "arc_challenge": { | |
| "acc": 0.3430034129692833, | |
| "acc_stderr": 0.013872423223718173, | |
| "acc_norm": 0.34982935153583616, | |
| "acc_norm_stderr": 0.013936809212158284 | |
| }, | |
| "sciq": { | |
| "acc": 0.929, | |
| "acc_stderr": 0.008125578442487923, | |
| "acc_norm": 0.923, | |
| "acc_norm_stderr": 0.008434580140240644 | |
| }, | |
| "piqa": { | |
| "acc": 0.7774755168661589, | |
| "acc_stderr": 0.009704600975718245, | |
| "acc_norm": 0.7861806311207835, | |
| "acc_norm_stderr": 0.009565994206915606 | |
| } | |
| }, | |
| "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 | |
| } | |
| } |