Credits

The wrong-layout fix tells real words from mistyped ones with small language models. This is what they are derived from.

Tatoeba

Language models are derived from sentences of the Tatoeba project (tatoeba.org), contributed by its users, licensed CC BY 2.0 FR.

FineWeb, FineWeb-2 and C4

Language models are derived from the FineWeb, FineWeb-2 (Hugging Face) and C4 (Allen Institute for AI) datasets, licensed ODC-By 1.0, built from Common Crawl.

What ships in Langouste

Only aggregate letter-sequence statistics, a few hundred kilobytes in total. No words and no sentences from these sources are in Langouste.