What is it?
A pre-trained model has learned general language ability from huge data. Fine-tuning takes that model and trains it further on a much smaller, targeted dataset so it specialises. In the most common form, supervised fine-tuning, you provide many examples of input → ideal output (for instance a support email and the reply your team would send), and the model's parameters are nudged to produce outputs like those.
Full fine-tuning updates all parameters and is expensive. Parameter-efficient methods such as LoRA (low-rank adaptation) train small add-on matrices while the base model stays frozen, drastically reducing the compute and storage required.
Why does it matter?
Prompting alone can leave gaps: output format that drifts, a house style that is hard to describe, a narrow classification task, or latency and cost issues because a long prompt is repeated on every call. Fine-tuning can bake the behaviour into the model so shorter prompts suffice, and can let a smaller, cheaper model match a larger one on a narrow task.
How does it work?
- Define the task and success criteria — and first try careful prompting; it is far cheaper.
- Build a dataset: hundreds to thousands of high-quality, representative examples, consistently labelled. Quality matters more than quantity.
- Split into training and held-out evaluation data.
- Train (via a provider's fine-tuning service or your own tooling on an open-weights model).
- Evaluate against the held-out set and against the un-tuned baseline (How to Evaluate AI Systems and Outputs).
- Deploy and monitor, and retrain when requirements change.
Important limits: fine-tuning is poor at reliably injecting new factual knowledge compared with retrieval, can make the model worse at things outside the training distribution, and any facts it does absorb are frozen at training time and hard to update or delete.
Example
A company wants every product description in a fixed JSON schema and a distinctive brand voice. Prompting gets 85% of outputs right but needs a long prompt with many examples. They collect 1,500 approved descriptions, fine-tune a smaller model, and measure schema-valid rate and reviewer preference against the prompted baseline. (Illustrative figures — you must measure your own.) If the descriptions must also reflect this week's price list, that part comes from retrieval or tools, not from the weights.
When should I use it?
Consider fine-tuning when you have: a stable, well-defined task; plenty of clean examples; a behaviour prompting cannot achieve consistently; or a volume where shorter prompts or a smaller model save meaningful cost.
Do not start with it when the issue is missing or changing knowledge (What is RAG (Retrieval-Augmented Generation)?), when you have few examples, or when requirements are still shifting. See RAG vs Fine-Tuning: Which Should You Use?.
Common mistakes
- Fine-tuning to teach facts, then being surprised by errors and staleness.
- Low-quality or inconsistent training data — the model learns the inconsistency.
- No baseline comparison, so you cannot show the tuned model is better than a good prompt.
- Training on sensitive data without considering that it may be memorised and reproduced (AI Privacy and Security: What to Know Before You Paste Data).
- Forgetting maintenance: tuned models need re-evaluation when the base model or the task changes.