AI model training and LLM fine-tuning in Corsica
AI model training means teaching a model a specific task from your own examples, so that it performs that task better or more cheaply than a general-purpose model. From Bastia, we adapt, evaluate, deploy and maintain custom models, and we tell you frankly when a simpler approach will do.
AI model training and fine-tuning: the basics
An AI model is a program that has learnt to perform a task from examples. Large language models (LLMs), which understand and write text, are trained by their developers on enormous volumes of data, with resources far beyond the reach of an SME. Building one from scratch is not on the table.
Fine-tuning starts from an existing model and continues its training on a set of examples specific to your business: requests labelled by category, product sheets written in your house style, annotated documents. The model keeps its general knowledge and learns your way of doing things.
A custom AI model can also be more modest: a classification or forecasting model trained solely on your data, for a well-defined task. Being smaller, it can then run on your own servers at lower cost.
LLM fine-tuning, RAG or prompting: when is training worth it?
This is the most important question, and the answer is often no. There are two simpler approaches, and we try them first.
- 01
Well-written prompts
Prompting means describing precisely to the model what you expect, backed by a few examples. It is simple to get right, inexpensive and enough for a large share of uses.
- 02
RAG
RAG (retrieval-augmented generation) gives the model access to your documents at the moment a question is asked. It is the right approach for information that changes: prices, procedures, case files. Fine-tuning, by contrast, is not designed for memorising facts.
- 03
Fine-tuning
It becomes worthwhile when you need consistent behaviour that prompting cannot achieve: a very strict output format, a house style, specialist vocabulary, a classification task repeated at high volume, or a smaller model that is cheaper to run, particularly on your own hardware.
Custom AI model examples by sector
A custom model makes sense when a task comes up often and follows rules specific to your organisation, which can apply as much to an SME in Haute-Corse as to a public body or a startup.
- 01
Retail
Products classified automatically in the catalogue, descriptions written in the brand's voice, customer messages sorted by reason for contact.
- 02
Tourism and hospitality
Reviews analysed against your own themes, such as welcome, cleanliness or food, and replies written in the establishment's style.
- 03
Professional practices
Recognising document types and extracting fields specific to your area of practice, with a model that can run locally to preserve confidentiality.
- 04
Trades and industry
Recognising parts or defects in photos, classifying call-outs, help with costing based on past jobs.
- 05
Software publishers
A specialised model built into your product, more predictable and cheaper to run than a general-purpose model.
From data to a model in production: our method
- 01
Understand
We define the task precisely, what a good answer looks like, the data available, your budget and timeline. We first measure what prompting and RAG can achieve.
- 02
Design
If training is justified, we choose the base model, prepare the examples and build a separate evaluation set that will act as referee. The plan and success criteria are approved by you before development.
- 03
Build
Data preparation, training, then a comparison of the fine-tuned model with the simpler approaches on the same evaluation set. You follow progress and see the measured results.
- 04
Launch and support
Deployment online or on a local server, quality monitoring in real conditions, retraining when your data or needs change.
Evaluating a model before trusting it
A model cannot be judged on the impression left by a few tries. We build with you a set of real cases, each with its expected answer, that the model never saw during training. It lets us measure accuracy objectively, compare several options and check that a new version does no worse than the last.
We also review errors one by one: a model that is rarely wrong, but badly wrong when it is, may be unacceptable for some tasks. It is the same insistence on systematic checking that runs through our own projects, such as La Forge, where every agent's output is tested and reviewed before it is accepted.
Training data, licences and maintenance: what to check
- 01
Representative examples
A model reproduces what it sees. Incomplete, unbalanced or incorrect examples produce a model that makes the same mistakes.
- 02
Personal data
Training on personal data requires a lawful basis under GDPR and often prior anonymisation. A model can sometimes reproduce fragments of its examples, so nothing goes in that must not come out.
- 03
Licences
The base model has its own licence, which governs commercial use and distribution of the fine-tuned model.
- 04
Portability
A model fine-tuned with an online provider generally stays with that provider; a fine-tuned open model can run on your own server. This choice is made at the design stage.
- 05
Maintenance
Needs change and better base models appear: a custom model has to be monitored and retrained over time.
What determines the cost of a custom AI model
- 01
Data preparation
Often the biggest item: gathering, cleaning and labelling reliable examples.
- 02
Model size
Fine-tuning a small open model does not take the same computing resources as fine-tuning a large one.
- 03
Evaluation requirements
Number of cases to check, severity of possible errors, comparisons to run.
- 04
Deployment and follow-up
A model served by an online provider, on dedicated hosting or on an AI server on your premises, then monitoring and retraining.
Technologies and methods
- Fine-tuning
- Open models
- Evaluation sets
- Machine learning
- GPU server
- Python
Frequently asked questions
What is the difference between fine-tuning and RAG?
RAG gives the model access to your documents when it answers: ideal for information that changes. Fine-tuning alters the model itself to teach it a behaviour, such as a style, a format or a classification task. The two can be combined.
How many examples are needed to train a model?
It depends on the task and the base model: a simple, well-defined task needs fewer than a nuanced one. The quality and consistency of the examples matter more than their number. We assess this with you from the start.
Can we train a model on our data without sharing it?
Yes, by fine-tuning an open model on a server you control, on your premises or with a host of your choice. Both the training data and the resulting model then stay with you.
Can a trained model become outdated?
Yes. If your business, categories or documents change, its performance can drop. That is why we monitor its quality and retrain it when needed.
Do I need a custom model to get AI suited to my business?
Rarely as a first step. Well-written prompts and access to your documents through RAG are enough in many cases. We reserve training for situations where it brings a measurable gain.