FineWeb-Edu — long-context, high-quality subset (350BT sample)

A filtered slice of FineWeb-Edu keeping only long-context, high-quality English documents.

• Source: HuggingFaceFW/fineweb-edu, sample/350BT (about 946 GiB / 339M rows)
• Filter: int_score >= 4 (top educational tier) AND token_count >= 4000 (long-context) AND language_score >= 0.95 (very-confident English)
• Result: 607,190 documents / 4.81B tokens (selectivity about 0.18%)
• Columns: id, dump, url, score, int_score, token_count, language_score, text

How it was made

Filtered with DuckDB (the heavy out-of-core scan) and re-read with Polars, handing off through this bucket, entirely on Hugging Face Jobs — about 73 minutes, $2.40, no standing infrastructure.

The whole pipeline was built and run by an agent on Jobs. Writeup: danielvanstrien.xyz/posts/2026/buckets-data-backend.

License

Derived from FineWeb-Edu, released under the ODC-By 1.0 license. This subset inherits the same license — please credit FineWeb-Edu (HuggingFaceFW) and Common Crawl.