book.json
Metadata and parsed structure: chapters, sections, figures, footnotes.
Book data for pretraining
Book data for LLM pretraining.
Paper-only books, OCR'd and structured into pretraining-ready bundles. Net-new tokens — text that has never entered a digital corpus, not the data every model already trained on.
The case
Why book data
The open web is picked over and deduplicated to exhaustion. The marginal token that still moves a model is one it has never seen.
Long-form, edited, professionally written prose is the densest such source — and the part still trapped on paper is the part no crawler has reached. We supply exactly that: books we own physically, turned into clean text and structure, verifiable title by title against our inventory.
The bundle
What you get
Every book is a self-contained bundle — machine-readable end to end, consistent title to title.
pages.jsonl
Per-page text, quality-scored, with language detected per page.
Cover images
Front, back, spine, and inside flaps where present.
Throughput
Scale & format
- 1,000+ books per day at target throughput
- JSON + JSONL, the same shape every title — drops into a pretraining pipeline without bespoke parsing
- Any language or script; per-page quality scores ship with the text
More on lineage and licensing in provenance & licensing, and on non-English coverage in the multilingual corpus.
FAQ
Common questions
What makes this data “net-new”?
It's OCR'd from paper-only books that have never entered a digital corpus — not in Common Crawl, archive.org, Google Books, or LibGen. That makes it text no model has trained on, verifiable title by title.
What format do you deliver?
Each title is a self-contained bundle: book.json (metadata and parsed structure) and pages.jsonl (per-page, quality-scored text), plus cover images — the same shape every title, so it drops into a pretraining pipeline without bespoke parsing.
How is text quality handled?
Every page is quality-scored and the score ships with the text, so you can filter on ingestion.
Can we check novelty before committing?
Yes. We send a sample of EANs and you dedup against your own corpus locally — nothing of yours ever leaves your side.
See what's net-new to your corpus
Send a sample size and we'll return a list of EANs to dedup against your own corpus.
Run the overlap check