README.md

BERT-large finetuned on MNLI.

The reference finetuned model has an accuracy of 86.05, we get 86.7:

{'eval_loss': 0.3984006643295288, 'eval_accuracy': 0.8667345899133979}