MORAIDICTIONNAIRE IA
Entraînement

Fine-tuning (Fine-tuning)

Retraining a pre-trained model on a focused dataset to specialize it for your task.

Picture a doctor already trained for years: to become a cardiologist, they don't relearn biology from scratch — they specialize. Fine-tuning applies this idea to AI: you start from a model pre-trained on vast general corpora, then continue training it on a smaller, targeted dataset to adapt it to a specific task, domain, or style.

Why fine-tune instead of retraining everything

Training a large model from scratch costs millions and demands mountains of data. Fine-tuning leverages transfer learning: general knowledge (grammar, reasoning, world structure) is already there. We merely adjust the network's weights with a supervised signal, minimizing a loss over the new examples:

$$\theta^{*} = \arg\min_{\theta}\ \mathbb{E}{(x,y)\sim \mathcal{D}(x),\, y)\big]$$}}}\big[\mathcal{L}(f_{\theta

Starting from pre-trained weights $\theta_0$, we nudge them gently toward the target task — hence a low learning rate, so the original knowledge isn't wiped out (catastrophic forgetting).

Full vs. efficient fine-tuning

Approach Parameters updated Cost Typical use
Full fine-tuning All weights High Deep adaptation, large budget
LoRA / PEFT A few % Low Fast specialization, many variants
Instruction tuning All or partial Medium Making a model "obedient"

PEFT methods (Parameter-Efficient Fine-Tuning), such as LoRA, train only small added matrices, making the operation feasible on modest hardware.

Fine-tuning isn't reinventing intelligence — it's steering it toward your need.

Explore the full AI dictionary →