Our long-term goal is for RSI data to become the primary dataset for pre-training as well as post-training. We are developing a synthetic pre-training corpus of textbook-style, long-form reasoning text, built by the same engine that produces our reasoning data.
What makes it worth training on.
A pre-training corpus with clear rights, instead of a web crawl with open questions.
Similarity-deduplicated against sources and public corpora →Written at the depth the next generation needs, not the average of the web.
Graduate & competition math: 1M+ problems →Designed to scale with compute, which pre-training volumes require.
Undergraduate STEM: 40M+ today, 100M target →Samples on request.
This line is in development. Labs interested in early evaluation can request access.
Request a sample →How it is sliced.
Form
Textbook-style, long-form expository and reasoning text.
Status
In active development. Scale is not yet published; we will share figures once they are firm.
Why it matters
Today's open pre-training corpora are mostly web crawls: unclear rights, uneven quality, low difficulty. That data trained the last generation of models.
Early access
Labs interested in co-developing or evaluating early slices can contact us.
Record format.
| Field | Type | Description |
|---|---|---|
| text | string | Long-form document in Markdown with LaTeX. |
| domain | string | Subject area. |
| level | string | Target depth. |
Built for.
- Pre-trainingA copyright-clear, high-difficulty foundation corpus.
- Mid-trainingDomain and reasoning enrichment before post-training.
Other linesfrom the same engine.
Interested in early access?
Tell us the pre-training corpus slice you need. We will send a pilot sample cut to your specification for your own evaluation.