The repository for the paper "When Do You Need Billions of Words of Pretraining Data?" - View it on GitHub