5fa1a76
1
2
We demonstrate that large gains on these tasks can be realized by generative pretraining of a language model on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each specific task.
