Skip to content
RAG Repo

FineWeb2

FineWeb2 applies the ideas behind FineWeb to many languages at once. Rather than reusing English-tuned rules, its pipeline adjusts quality filtering and deduplication (removing duplicate and near-duplicate pages) per language, so smaller languages are cleaned on their own terms rather than by English standards.

The data comes from Common Crawl and is published as Parquet files on HuggingFace, streamable with the `datasets` library so you can work language by language without downloading everything. This makes it a practical foundation for non-English and cross-lingual RAG knowledge bases.

web-crawlmultilingualpretrainingfiltereddeduplicatedhuggingface

Related sources