Fine-Tuning and LoRA
Decide when training a model beats prompting or retrieval, and see how LoRA adapts a large model by training a few million extra parameters.
Should you fine-tune at all
Fine-tuning means continuing to train a pretrained model on your own examples so its weights change. It is the most expensive way to change behaviour, so try the cheaper ones first.
| Need | Try first | Why |
|---|---|---|
| Follow instructions better | Clearer prompt, examples in the prompt | Free to iterate |
| Use private or changing facts | RAG (retrieval) | Facts live in an index you can update |
| A fixed output format or house style | Prompt, then fine-tune | Training makes it consistent |
| Cheaper or faster at one narrow task | Fine-tune a small model | A small tuned model can match a large prompted one |
| New behaviour prompts cannot reach | Fine-tune | Last resort |
You need an eval first (see the evals lesson). Without one you cannot tell whether the tuned model is better or just different.
1 of 6