Techniques

What Is Fine-Tuning?

Fine-tuning is the process of further training a pre-trained AI model on a smaller, specialised dataset to improve its performance on specific tasks, domains, or output styles.

The Plain-English Explanation

A pre-trained model like GPT-4 is a generalist — it knows a lot about everything but may not excel at your specific use case. Fine-tuning takes this general model and trains it further on your data — your customer support transcripts, your legal documents, your medical records — so it becomes an expert in your domain.

Think of it like hiring a university graduate (pre-trained model) and then training them on your company's specific processes, terminology, and standards (fine-tuning). They arrive with broad knowledge; you add the specialised expertise.

Why It Matters

Fine-tuning bridges the gap between general AI capability and domain-specific performance. For organisations that need AI to use specialised terminology, follow specific formats, or perform at expert level in a narrow domain, fine-tuning can be the difference between a useful tool and an unreliable one.

How It Works

You prepare a dataset of examples showing the inputs and outputs you want. The pre-trained model's weights are then adjusted through additional training on this data, teaching it to prefer the patterns in your examples. The result is a model that retains its general capabilities while performing significantly better on your specific tasks.

Examples in Practice

Common Misconceptions

Myth: Fine-tuning is always better than prompting.

Reality: For many tasks, well-crafted prompts or RAG systems achieve comparable results at a fraction of the cost and effort. Fine-tuning is best when you need consistent behaviour that prompting alone can't achieve.

Myth: You need thousands of examples to fine-tune.

Reality: Modern fine-tuning techniques can work with as few as 50–100 high-quality examples. Quality matters more than quantity — well-curated examples outperform large, noisy datasets.

Myth: Fine-tuning is permanent and can't be adjusted.

Reality: You can fine-tune again with new data, adjust the training parameters, or revert to the base model. It's an iterative process, not a one-time commitment.

Related Terms

Further Reading

Learn Fine-Tuning in Depth

Module 3 of AI Fundamentals covers fine-tuning alongside other model customisation techniques — helping you understand when and how to tailor AI to your specific needs.

Explore AI Fundamentals

Frequently Asked Questions

When should I fine-tune versus use RAG?
Use RAG when you need the AI to access and cite specific documents. Use fine-tuning when you need the AI to change its behaviour, style, or domain expertise consistently across all interactions. Many organisations use both.
How much does fine-tuning cost?
Costs vary by model and dataset size. Fine-tuning GPT-3.5 might cost $10–50 for a small dataset. GPT-4 fine-tuning is significantly more expensive. Open-source models can be fine-tuned for the cost of computing time.
Can anyone fine-tune an AI model?
OpenAI, Google, and others offer fine-tuning through their APIs with no machine learning expertise required. You prepare the data in a specific format and submit it. However, designing good training data still requires careful thought.
Back to AI Glossary