Distileo
TCKC's distillery for hyper-specialised models.
See the applicationsThe problem
A generalist model does not know a domain corpus: it paraphrases its general culture where a regulated domain demands its rules, its references, its entries. And sending a sensitive corpus through a foreign API is often out of the question. Truly grounded assistants need models faithful to the documents — and deployable wherever the constraints require.
Why at TCKC
- TCKC's applications rely on official French corpora: chart of accounts, ANC standards, tax and labour texts
- Distileo explores how to derive specialised models from them, deployable in France (French cloud) or locally
- 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-cycle 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 the questions; the student learns from those reasonings through local QLoRA fine-tuning.
A mandatory A/B gate
The fine-tuned model is compared to its base by dual judging — the grounding gain must be demonstrated, or the model is discarded.
A usable model
Merged, quantised, exported to Ollama, with deployed-model identity verification (SHA-256 fingerprint).
Full traceability
Data, costs and quality gates tracked cycle by cycle, with content-fingerprint deduplication.
How it works
Distileo is not a model: it is the factory that produces them. Document extraction 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 stays local. 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
FAQ
Is Distileo a downloadable model?
No. Distileo is the factory: it produces specialised models, distilled from a domain corpus, for the R&D behind TCKC's applications.
Do the documents leave the machine?
Extraction and fine-tuning run locally. Only the teacher's reasoning step runs on a French cloud (Mistral stack).
What guarantees the produced model's quality?
Two automatic judges score every dataset pair (faithfulness to the documents, normative accuracy), and a final A/B gate compares the fine-tuned model to its base: no demonstrated grounding gain, and the model is discarded.
What document types?
PDF, Word, Markdown and text, in French. The pilot domain is accounting and financial analysis; other regulated domains fit equally well.
How does it relate to TCKC's applications?
The pilot domain — accounting and financial analysis — is that of Parles au PCG and FEC.NEWS: Distileo explores how specialised models can strengthen the grounding of this kind of application.
TCKC's applications
Distileo belongs to TCKC's R&D; the applications in production are in the portfolio.