Tu es un expert en fine-tuning de LLMs. Je veux entraîner un modèle sur [DONNÉES: DOCUMENTS TECHNIQUES, SUPPORT CLIENT, CODE SPÉCIFIQUE...]. Pipeline complet de fine-tuning: 1. **Dataset Preparation** : Tokenization, formatting, deduplication, quality filtering 2. **Model Selection** : Base model choice (Llama, Mistral, Falcon) vs training from scratch 3. **Training Setup** : LoRA vs QLoRA vs full fine-tuning, hyperparameter optimization 4. **Hardware Requirements** : GPU memory calculation, batch size optimization, multi-GPU training 5. **Training Process** : HuggingFace Trainer, wandb logging, checkpoint management 6. **Evaluation Metrics** : Perplexity, BLEU, ROUGE, human evaluation 7. **Inference Optimization** : Quantization (INT8/FP16), ONNX export, GPU vs CPU deployment 8. **Deployment** : Ollama, vLLM, FastAPI serving, load balancing 9. **Safety & Alignment** : Content filtering, prompt engineering, bias detection 10. **Continuous Improvement** : Active learning, feedback loops, model versioning Génère les scripts Python complets, les configurations de training, et le pipeline d'inférence.