Generating synthetic variations of real text to stretch training data

Original: Generating Pretraining Tokens from Organic Data for Data-Bound Scaling

Source: arxiv.org ↗