French & European
Sourced directly from our EU inventory — French and other European languages at native quality.
Multilingual corpus
Non-English book data, at the source.
A multilingual physical catalog and a pipeline that processes any language or script — one supplier instead of stitching several together.
The gap
Why non-English is scarce
Web crawls skew heavily English. High-quality non-English text — especially long-form books — is comparatively rare, and often lower quality where it exists.
Native book corpora are the strongest non-English pretraining source and the least digitized. Sourcing them physically, in-market, is the reliable way to get them — which is exactly what an EU bookseller with a multilingual catalog is positioned to do.
Languages
Coverage
Chinese, native-reviewed
Chinese titles processed with in-house native-speaker review through our Chinese subsidiary.
Any language
The pipeline OCRs and structures whatever language or script you need — not English-first.
Quality
Per-language quality
- Language detected per page, not assumed per book
- Every page quality-scored; the score ships with the text
- Native review on flagged non-Latin pages before delivery
See also book data for pretraining and provenance & licensing.
FAQ
Common questions
Why is non-English book data scarce?
Web crawls skew heavily English, so high-quality long-form non-English text is comparatively rare and often lower quality. Native book corpora are the strongest non-English source and the least digitized.
Which languages do you cover?
French and other European languages from our EU inventory, native-reviewed Chinese, and any language or script the pipeline OCRs and structures — not English-first.
Is language detected per book or per page?
Per page. Language is detected per page, not assumed per book, and every page is quality-scored.
How is non-Latin quality ensured?
Flagged non-Latin pages get native review before delivery.
Non-English tokens no one else has
Send a sample size and a target language; we'll return EANs to dedup against your corpus.
Run the overlap check