What is LLM fine-tuning?
Fine-tuning is the process of taking a pretrained language model and continuing its training on a smaller, domain-specific dataset to specialize the model for a task or style. It bakes knowledge into the model weights, in contrast to RAG which keeps the knowledge in an external database.
When to fine-tune
Fine-tune when you need consistent style or format that prompting cannot reliably produce, or when you have a task-specific dataset that improves performance over what the base model offers. Fine-tune sparingly: the operational overhead is significant and updates require a fresh training run.
When not to fine-tune
If the goal is to make the model answer using your knowledge base, do not fine-tune. Use RAG. Fine-tuned knowledge is impossible to audit, expensive to update, and prone to drift. RAG keeps facts in a database where you can edit them in real time.
Cost shape
Full fine-tuning of a modern open-weight model can require a GPU cluster for hours. LoRA and QLoRA reduce this to a single GPU for a few hours by training only a small adapter on top of frozen weights. For most customer-support use cases, RAG plus prompt engineering reaches similar quality at much lower cost.
See it working on message.com.
Unified inbox, grounded AI, flat pricing. 14 days free, no credit card.
Start free trial →