Fine-tuning
Définition
L'affinage poursuit l'entraînement d'un modèle pré-entraîné sur des données spécifiques à une tâche ou un domaine. Full fine-tuning updates all parameters; parameter-efficient methods (par ex. LoRA, adapters) update a small subset to reduce cost.
Utilisez-le quand you need stable, task-specific behavior or style (par ex. domain language, output format) and have enough labeled data. For frequently updated knowledge or one-off questions, RAG or prompt engineering are often better. See LLMs for the full training pipeline.
Comment ça fonctionne
You start from a base model (par ex. a pretrained LLM) and a dataset of task examples. You define a loss (par ex. cross-entropy for classification, next-token for generation) and run optimization (par ex. Adam) on your data. Le résultat est un fine-tuned model whose weights are updated (fully or only adapters/LoRA). Instruction tuning uses (instruction, response) pairs so the model learns to follow prompts; domain fine-tuning uses in-domain text or labeled tasks. Validation and early stopping prevent overfitting; often only 1–5% of parameters are updated with LoRA to save compute.
Cas d'utilisation
Fine-tuning is the right tool when you need a model to follow a specific style, domain, or task better than prompting alone.
- Adapting a base model to a specific domain (par ex. legal, medical)
- Teaching a consistent output format or style (par ex. JSON, tone)
- Improving performance on a narrow task with limited labeled data