Fine-tuning means continuing to train a language model on examples specific to your company, so that it adopts your vocabulary, your formats and your everyday tasks.
Key points
- Fine-tuning changes the way the model answers: vocabulary, tone, format. Not its factual knowledge.
- Facts come from your documents, retrieved for each question (RAG).
- An adapted model only goes into production if it passes the evaluation defined with you.
What it changes, and what it doesn’t
Fine-tuning acts on the way the model answers: vocabulary, tone, answer format. Facts come from your documents, retrieved for each question and cited as sources. The two approaches complement each other.
Instructions, RAG or fine-tuning: which one?
There are three levers to adapt a model. Instructions, also called the prompt, steer a given answer. Document retrieval (RAG) brings the facts from your documents to each question. Fine-tuning durably changes the way the model answers.
You almost always start with the first two. Fine-tuning comes in when the style or format has to stay consistent across a large number of answers.
Instructions (prompt)
What changesOne given answer
When to use itOne-off task
Document retrieval (RAG)
What changesThe facts available
When to use itAnswers from your documents
Fine-tuning
What changesThe way it answers
When to use itConsistent style or format
Try it
Which lever for your need?
Answers must rely on your internal documents?
Do these documents change often?
Do you want an answer style, vocabulary or format that is always the same?
Is the task one-off and simple to describe?
Our recommendation
Answer the questions to see the recommendation.
What you provide
We prepare these examples with you: sorting, formatting, anonymisation if needed.
- Examples of questions and good answers
- Reference documents
- Templates of documents you often produce
How it is evaluated
The adapted model is compared with the starting model on cases defined with you, against criteria written down during scoping. If it fails the evaluation, it does not go into production.
Choosing the starting model
No model is imposed. The choice is made per project between Mistral, LLaMA, Qwen and Gemma, depending on the language, the available hardware and the licence.
Frequently asked questions
Further reading
- RAG: an AI that answers from your documentsHow an AI finds the right passages in your documents, answers and cites its sources. Diagram, step-by-step demo and limits, in plain English.
- Private LLM for business: the guideWhat a private LLM is, its architecture, the GPU memory it needs (calculator), what stays on your side and its current limits.
In 30 minutes, we look together at whether a pilot makes sense for you.