Distileo
A distillery for hyper-specialised models.
Discuss a projectThe problem
A generalist model does not know your documents: it paraphrases its general culture where your work demands your rules, your references, your entries. And handing a sensitive corpus to a foreign API is often out of the question. You need a model faithful to your documents — grounded — and deployable wherever your constraints require.
Who it's for
- Organisations whose value sits in a French business corpus: firms, finance departments, regulated sectors
- Teams that need a model deployable in France (French cloud) or on their own machines
- Pilot domain: accounting and financial analysis (PCG, ANC standards, CGI)
What it does
A complete 14-step pipeline
PDF, Word, Markdown or text ingestion → local structured extraction → constrained Q&A generation → a per-project JSONL dataset.
Dual quality judging
Every pair is scored on faithfulness to the source excerpt AND normative accuracy; below thresholds, it is rejected.
Reasoning distillation
A sovereign teacher model reasons over your questions; the student learns from those reasonings through local QLoRA fine-tuning.
An A/B gate before delivery
The fine-tuned model is compared to its base by dual judging — the grounding gain must be demonstrated, or nothing ships.
A usable deliverable
Merged, quantised, exported to Ollama, with deployed-model identity verification (SHA-256 fingerprint).
Full traceability
Data, costs and quality gates tracked per project, with content-fingerprint deduplication.
How it works
Distileo is not a model: it is the factory that produces them. Extraction of your documents runs entirely locally; the teacher's reasoning runs on the Mistral stack hosted on French cloud; the student's fine-tuning (Ministral family) runs locally; deployment goes to a French host or your own machines. The central, measured lesson: a judge-filtered dataset increases the model's grounding; unfiltered data degrades it below its base. The judging quality is the product.
- Sovereignty mapped step by step: local extraction, teacher and judge on French cloud, local training
- One full cycle already completed end-to-end on the pilot domain, with a grounding gain measured by dual judging
- A web control room: an eight-station production line, sovereignty badges, cost tracking
Status & access
FAQ
Is Distileo a model I can download?
No. Distileo is the factory: it produces one model per project, distilled from your documents. The deliverable is yours and deploys to a French host or your machines.
Do my documents leave my machines?
Extraction and fine-tuning run locally. Only the teacher's reasoning step runs on a French cloud (Mistral stack); a fully on-premise run can be discussed per project.
What guarantees the produced model's quality?
Two automatic judges score every dataset pair (faithfulness to your documents, normative accuracy), and a final A/B gate compares the fine-tuned model to its base: no demonstrated grounding gain, no delivery.
What document types?
PDF, Word, Markdown and text, in French. The pilot domain is accounting and financial analysis; other regulated domains fit equally well.
What does it cost?
Each project is quoted individually, based on the corpus and the target deployment level. Let's talk.
Your corpus deserves its model
Describe your documents, deployment constraints and intended use — we will frame the distillation project.