A hands-on guide to fine-tuning LLMs for real-world applications. The author shares a personal example of fine-tuning a 7B parameter model for filling out synoptic reporting templates for breast cancer, achieving ~98% accuracy compared to ~35% with prompt engineering and RAG. The task involved complex input formats and strict field ordering, where even one error invalidated the entire output. Fine-tuning with QLoRA eliminated API costs (estimated at $320,000) and significantly improved performance.

The article covers when to fine-tune versus using RAG, the mathematical intuition behind LoRA/QLoRA, technical implementation details, and evaluation methods. Key scenarios for fine-tuning include rigid formatting requirements, cost constraints with large system prompts, and complex combinatorial instructions that exceed in-context learning limits.

RAG is better suited for frequently changing knowledge bases or when only minor behavior augmentation is needed. The mathematical foundation involves supervised fine-tuning with cross-entropy loss on curated (prompt, completion) pairs, while parameter-efficient methods like LoRA avoid the computational costs and knowledge destruction risks of full fine-tuning.

Readers will learn when to fine-tune, why it works, and how to implement it practically using techniques like QLoRA for efficient low-rank adaptation.