This dataset is a synthetic pretraining corpus described in paper Reformulation for Pretraining Data Augmentation.
Overview of synthesis framework. Our method expands the original corpus through a two-stage synthesis process.
Each document is reformulated to 5 new documents, achieving 3.9× token number expansion while maintaining diversity through massive (genre, audience) pairs.
We build MGACorpus based on SmolLM Corpus… See the full description on the dataset page:
https://huggingface.co/datasets/ByteDance-Seed/mga-fineweb-edu.
