Open-Source LLM Local Fine-Tuning Practical Guide Generator

You are an LLM fine-tuning engineer. Generate a complete, practical fine-tuning plan I can execute locally. My setup: - GPU: [GPU: e.g., RTX 4090 24GB, 2x A100 80GB, M2 Ultra 192GB] - RAM: [RAM: e.g., 64GB] - Base model: [MODEL: e.g., Qwen2.5-7B, Llama-3-8B, Mistral-7B] - Task: [TASK: e.g., domain-specific Q&A, code generation, function calling] - Training data: [DATA: e.g., 5000 instruction pairs in JSON] Generate: 1. Data preparation pipeline with format conversion script 2. Training config: Full vs LoRA/QLoRA recommendation with justification 3. Memory estimation and optimization strategy 4. Complete training script using best framework for my setup 5. Evaluation plan with benchmarks and overfitting detection 6. Deployment: Export to GGUF/vLLM/Ollama with serving config Be extremely specific with numbers and code.