DATASET
Your examples, curated
We curate examples from your real cases — input and expected output, reviewed with your team.
// SERVICE · FINE-TUNING
A model that speaks like your company.
We tune open-source models on your real examples: your vocabulary, your formats, your criteria. For repetitive tasks where precision matters.
THE USUAL QUESTION
RAG or fine-tuning?
RAG
For knowledge that changes
Prices, stock, manuals, regulations: the model searches your sources at answer time. You update the document and you are done — citation included.
FINE-TUNING
For behavior that repeats
Tone, format and judgment learned from your examples: classify, extract, write the same way every time — faster and with less context.
In practice, we combine both: RAG to know, fine-tuning to do.
See local AI + RAG →WHEN IT PAYS OFF
Tasks where fine-tuning pays
- CLASSIFICATION
- With your vocabulary
- Tickets, complaints and invoices tagged with the categories your company uses — not generic ones.
- WRITING
- With a fixed format
- Quotes, technical reports, sales replies: always the same structure, in your tone.
- EXTRACTION
- From your documents
- Delivery notes, purchase orders, contracts: the fields you care about, straight into the system.
- ASSISTANTS
- Single-task
- Data entry, checking and validation: a small, tuned model beats a giant, generic one.
HOW WE DO IT
From example to model in production
TRAINING
On open-weight models
LoRA / QLoRA on open-weights models, on your infrastructure or ours — never with your data in external APIs.
EVALUATION
Against your quality bar
Against a test set with your quality criteria. The numbers are shared — no black box.
DEPLOY
Local, versioned, reversible
Versioned and local, with rollback. Scheduled retraining as you gather more examples.
DELIVERABLES
What the service includes
- ✓ Curated, versioned dataset, owned by your company
- ✓ Documented training and evaluation, with metrics per version
- ✓ Model deployed on your infrastructure, integrated into your system
- ✓ Scheduled retraining and quality monitoring in production