Fine-tuning is additional training of a pre-trained foundation model on a smaller, domain-specific dataset to adapt it for specific tasks, voices, or formats. The result is a customized model that performs better on the target task than the base model alone. Costs include training compute, dataset preparation, and potential overfitting to the fine-tuning data. It's one of three main ways to specialize foundation models for specific applications (alongside prompting and RAG).
The three customization approaches:
| Approach | What it does | When to use |
|---|---|---|
| Prompting | Guide model behavior through prompts | Quick experimentation; flexible needs |
| RAG (retrieval-augmented generation) | Inject relevant data at inference time | Knowledge needed b... |