Skip to content
RAG Repo

CulturaX gives you broad-coverage web text in 167 languages, ready for pretraining or for building non-English and multilingual retrieval systems. It merges two well-known web corpora, mC4 and OSCAR, then runs the combined data through a pipeline of language identification, URL-based filtering, metric-based cleaning, document refinement, and deduplication (removing duplicate and near-duplicate text so you are not storing the same passage twice).

The dataset is published on Hugging Face, so you can stream it or download it in parts rather than pulling all 6.3 T tokens at once. The licence is ODC-By, which permits commercial use with attribution, but the data still inherits the underlying mC4 and OSCAR terms, so check those if your use case is sensitive.

multilingualpretrainingcleaneddeduplicatedmc4oscar

Related sources