What is fine-tuning?
Fine-tuning is the process of taking an existing AI model and training it further on a smaller, specialized set of examples. The goal is to make the model better at a specific task, style, or subject than it is by default.
Last reviewed .
How fine-tuning works
You start with a pre-trained foundation model and provide examples of the inputs and outputs you want. The model's internal settings are adjusted so it produces similar results on new inputs.
The quality of those examples largely decides the quality of the result.
Fine-tuning, prompting, and RAG
Prompt engineering changes the instructions you give a model, and retrieval-augmented generation supplies it with relevant documents at the time of the request. Both are usually faster and cheaper to try than fine-tuning.
Fine-tuning is more often worth considering for consistent tone, format, or specialized behavior that prompts alone can't deliver. It's less suited to teaching a model facts that change often.
What to consider first
Fine-tuning needs good training data, technical effort, and ongoing maintenance. Try better prompts and retrieval first, and fine-tune only if they fall short.
Common questions
Is fine-tuning the same as training a model?
Not quite. Training from scratch builds a model from raw data at great cost. Fine-tuning starts with a model that's already trained and adjusts it for a narrower purpose.
Should I fine-tune a model or use RAG?
Use RAG when the model needs access to specific or changing information. Consider fine-tuning when you need consistent style or behavior. Some projects use both.
How much data does fine-tuning need?
It depends on the model and task. Quality matters more than volume, so start with a small, clean, representative set of examples.
Does fine-tuning stop hallucinations?
No. It can improve behavior on a task, but the model can still produce incorrect statements, so outputs still need review.
Want help putting this to work?
Tell us where your marketing stands today, and we'll show you what to do first.