Define the expected behaviour.
We can begin by defining instructions, examples and output format. This provides a useful baseline for measuring what further adaptation adds.
AI trainingYour examples.
Your examples.
A model better suited to the task.
The studio helps prepare data, adapt a model and evaluate its behaviour on a specific task. Each step should inform a concrete decision.
Our proposed supportLearn a task.
Learn a task.
Verify what changes.
Classifying incoming requests, recognising an image category or producing a response in a business format: training starts by describing the expected result and the errors that matter.
We can assess adapting an existing model, a network dedicated to a focused task and integration into your tools. Based in Pont-Audemer, the studio supports organisations across France and holds European ambitions for its AI projects.
A measurable loop.Prepare · Train · Evaluate
- Starting point
- A defined task
- Raw material
- Checkable examples
- Decision
- Shared criteria
Scope, deliverables and target environment are specified during project scoping.
Neural networksLayers of computation.
Layers of computation.
Training with a defined framework.
A network transforms data through layers of computation. Training adjusts its parameters using examples and an objective. Evaluation then observes behaviour on data held apart from training.
- 01InputsData to analyse
- 02Computational layersThe learned transformation
- 03OutputsThe answer to evaluate
To understand layers and parameters: the PyTorch neural networks guide.
Choose an approach
The right approach for your needs.
Scoping compares several approaches using the same examples. System complexity should remain proportionate to the task and its operation.
Find the right information.
An assistant can retrieve relevant passages from your documents and provide them to the model with the request. This approach addresses sources, updates and answer traceability.
Explore the document assistantAdapt through examples.
Fine-tuning continues a pretrained model’s learning for a task. Methods such as LoRA train adaptation matrices while keeping the original weights frozen. Their relevance is checked against the selected use cases.
The principle is described in the Hugging Face PEFT documentation.
Assess a focused task.
For classification, image analysis or prediction on structured data, we can assess a dedicated network. Available examples, the baseline method and runtime constraints guide experimentation.
From dataset to business evaluation.
- 01 · Define
The task and baseline
We define inputs, expected answers, success criteria and acceptable errors. Current behaviour provides a comparison for deciding whether experimentation adds value.
- 02 · Prepare
Examples and evaluation sets
Work covers usage rights, cleaning, annotation and data separation. Held-out test cases should evaluate the model’s ability to handle new examples.
- 03 · Experiment
Reproducible adaptation
We prepare tests, track configurations and retain useful versions. Observed errors guide iterations, with attention to rare cases and requests outside the intended scope.
- 04 · Integrate
A system ready for operation
The selected version is tested in its target environment. Depending on the project, we prepare the API, human approval rules, technical logs, documentation and update conditions.
Deliverables your team can take over.
- A definition of the task, scope and acceptance criteria.
- An inventory of data used, transformations and annotation rules.
- Test configurations and, according to the chosen solution, deliverable weights or adapters.
- An evaluation report presenting observed results and test limitations.
- Runtime, integration and monitoring documentation suited to your team.
Prepare your project.
Must a model be trained to use our documents?
The starting point can be a document assistant that retrieves passages from your sources and provides them to the model. Training becomes an option when behaviour, classification or a format needs to be learned from examples. The two approaches can also complement each other.
How much data is needed?
That depends on the task, its variety and the starting model. Scoping first examines a sample: consistency of expected answers, difficult cases, duplicates and the ability to reserve examples for testing. Useful volume is determined from that assessment.
Can the model then be used with Ollama?
This option must be checked against the architecture, weight format and produced adapters. Ollama documents importing compatible models and adapters. We plan an export and runtime test in the target environment before selecting this deployment route.
Let’s start with the task.
Your need, examples and success criteria give experimentation a framework. The studio helps turn those elements into a next step.
Talk with the studio
AI integration · Data architecture · The studio’s AI expertise