Grands modèles de langage (LLM)
Définition
Les grands modèles de langage sont des modèles basés sur les transformers entraînés sur des données textuelles massives (et parfois multimodales). They exhibit emergent abilities: few-shot learning, raisonnement, and tool use when scaled and aligned (par ex. via RLHF).
Un modèle mental utile : le pré-entraînement apprend la prédiction du prochain token sur d'énormes corpus et donne au modèle une large connaissance and language ability. Instruction tuning (and similar) trains the model to follow user instructions and formats. Alignment (par ex. RLHF, DPO) shapes behavior to be helpful, honest, and safe. At inference time you can use the model zero-shot, few-shot, or augment it with récupération (RAG) or tools (agents).
Comment ça fonctionne
Pretraining learns prédiction du prochain token on large corpora and produces a base model. Optional fine-tuning (par ex. fine-tuning) adapts it to tasks or instruction formats; alignment (par ex. RLHF, DPO) optimizes human preference and safety. The deployed model is then used at inference time. You can call it zero-shot (no examples), few-shot (with prompt engineering), or augment it with RAG (récupération as context) or agents (tools and loops). The diagram summarizes the training pipeline and the two main inference augmentations.
Cas d'utilisation
LLMs are used wherever you need flexible language understanding or generation, from chat to code to analysis.
- Chat, summarization, and translation
- Code assistance and generation
- Question answering and research assistance (often with RAG or tools)
Avantages et inconvénients
| Pros | Cons |
|---|---|
| Flexible, one model for many tasks | Cost and latency |
| Strong few-shot performance | Hallucination, bias |
| Enables agents and tool use | Requires careful evaluation |
Documentation externe
- OpenAI – Models overview — GPT and capabilities
- Google AI for Developers — Gemini and APIs
- Anthropic – Models — Claude and documentation
- Hugging Face – NLP course — From transformers to LLMs