ROOTS

ROOTS, short for the Responsible Open-Science Open-Collaboration Text Sources corpus, was assembled by BigScience, the year-long open research collaboration that produced the BLOOM multilingual model. It is the corpus BLOOM trained on, and it was built deliberately rather than scraped: much of it comes from sources the community nominated and vetted for quality and provenance, combined with a filtered slice of the OSCAR web corpus. That human-in-the-loop selection is the thing that sets it apart from a pure crawl.

The dataset runs to about 1.6 TB and spans 46 natural languages and 13 programming languages, with a strong presence of languages that most Western web corpora barely cover, including many African, South Asian, and South-East Asian languages. Documents are provided as JSON, and the constituent datasets are published on Hugging Face under the `bigscience-data` organisation, each carrying its own provenance notes.

Access is gated rather than fully open. You browse the component datasets on Hugging Face and request access per subset, so plan for an approval step rather than a single anonymous download. Once you are through, the per-language structure makes it easy to pull just the languages your RAG system needs and skip the rest, which keeps your index lean and your embedding costs down.

ROOTS earns its place when you are building multilingual or low-resource RAG and want vetted text rather than raw crawl. The curation means less boilerplate and spam than a pure web dump, and the language breadth is hard to match. It is cleaned prose rather than pre-chunked passages, so you will still handle splitting and embedding yourself.

The main thing to watch is the licence. ROOTS is released under the BigScience RAIL licence, a responsible-AI licence that permits research and many uses but restricts certain applications, and the exact terms vary by subset. Treat commercial use as restricted and read the licence for the specific subset you plan to use rather than assuming blanket permission. For broader multilingual coverage with more permissive terms, weigh it against OSCAR, mC4, and CulturaX.