Before investing in training at scale, we tested the simplest possible hypothesis: does a tokenizer trained on specialized medical text actually segment clinical vocabulary better than a generalist one? Measured result: -32% fewer tokens needed than the current Mistral tokenizer, -21% fewer than GPT-5.6. Every step is published with numbers, not just promises.
Weights initialized at random, never borrowed from an existing model. A real proprietary model, not a layer bolted onto someone else's weights.
No training data generated by a proprietary LLM. The rule → case → outcome format comes mechanically from real regulatory text, never from synthetic generation.
BDPM, PubMed FR, NCBI Bookshelf, open research. Methodology and code reproducible end to end, without ever touching real patient data.
Every published number comes from a real run on real data — never a projection, never a marketing figure.
Before investing in training at scale, we validated the core hypothesis: does a tokenizer trained on a specialized medical corpus actually segment medical vocabulary better than a generalist tokenizer?
fertility = tokens / word, aggregated over 3,254 real medical
documents (full RCP drug labels, PubMed FR literature, frontier research). Methodology
and source code reproducible end to end — no patient data used.
Three pillars, each documented and reproducible.
BDPM, PubMed FR, NCBI Bookshelf, open research — zero patient data. Vocabulary empirically optimized (32k→50k), not guessed.
Phase 1 — doneCell Ontology, UBERON, MeSH — 21,287 real concepts used to tier the corpus, from foundational to frontier research.
Phase 1.5 — doneEvery drug label mechanically split along its own regulatory structure. No data generated by a proprietary LLM.
Ready for Phase 2Beyond the technical performance measured above, timing works in our favor.
French HDS regulation requires health data to be hosted by an HDS-certified provider. OpenAI and Anthropic are not certified, which puts their models out of reach for this data — a regulatory block, not a technical one.
EHR, revenue-cycle, and medical-coding software vendors: AI infrastructure to license, not another consumer product.
Two tiers. The first validates the approach at low cost; the second is the real deliverable.
Every step, dated. Not announcements — the real record of work in progress.
The pipeline is ready to run. As soon as GPU access clears, Tier 1 moves from CPU to a real run.
Confirms the end-to-end pipeline (data → tokenization → training → eval) is stable over hundreds of real steps.
Applications filed with startup programs (OVHcloud, Scaleway, AWS) to fund training at scale.
Mechanical split of regulatory documents — zero data generated by a proprietary LLM.
Cell Ontology, UBERON, MeSH — 21,287 real concepts tiering the corpus from foundational to frontier research.
First falsifiable technical proof: French clinical vocabulary is segmented more efficiently than with a generalist tokenizer.
The tokenizer and data pipeline are real and working today — not slides. The trained model (Tier 1) will be available for demo once real training is complete.