Skip to content
RAG Repo

OSCAR

OSCAR runs Common Crawl snapshots through a language classifier and splits the text by language, giving you clean per-language slices rather than one mixed pile. This makes it one of the more practical starting points when you need web text in a specific language, especially languages that larger English-first datasets barely cover.

It spans more than 150 languages and is refreshed as new Common Crawl snapshots and pipeline versions land. Later releases add document-level annotations, such as quality and content flags, so you can filter before you index or train.

Access is through the OSCAR project site and HuggingFace, with some larger or more sensitive language sets gated behind a quick access request. The metadata is released under CC0, while the underlying page content stays subject to its original terms.

web-crawlmultilingualpretraininglanguage-classification

Related sources