Putting it together

You've got the decision tree, the machinery, and the guardrails. So when should you actually fine-tune?

What you've got now

In six short lessons you've built a working method for customization — not a reflex to fine-tune everything, but a set of decisions you can run on any model that isn't good enough yet.

• —Decide first — knowledge gap → RAG, behavior gap → maybe fine-tune, else prompt.
• —Know the machinery — SFT shifts behavior toward your examples; LoRA does it cheaply.
• —Data is the spec — a few hundred clean, consistent examples beat thousands of noisy ones.
• —Distill — use a big teacher to train a small, fast, cheap student.
• —Evaluate honestly — held-out data, compared to the base model, every change.

The one rule to keep

If you remember one thing: reach for the cheapest tool that closes the gap. Prompt, then RAG, then fine-tune — and only fine-tune when you have a clear behavior gap and the clean data to teach it. Most 'we need to fine-tune' problems are a prompt or a retrieval problem in disguise.