On this page
Definition
Fine-tuning is additional training applied to a pretrained model on curated examples so the model specialises in a task, tone, format or domain.
Simple explanation
A fine-tuned model is like a general professional who has spent six months working only on your kind of problem. Ask the same question and you get a more specific, more consistent answer.
Why it matters
For high-volume, narrow tasks, a fine-tuned smaller model can beat a larger general model on quality, cost and speed.
How it works
- 1CurateAssemble a dataset of ideal input/output pairs.
- 2TrainRun additional training on top of a base model.
- 3EvaluateCompare against the base model on your task.
- 4DeployUse the fine-tuned model in place of the base.
Real examples
Products named for illustration only. Inclusion is not an endorsement.
- OpenAI fine-tuningGPT-4o mini and 4.1 available for fine-tuning via the API.
- Open-weight modelsLlama and Mistral variants routinely fine-tuned in-house.
Advantages
- Consistent style and format.
- Shorter prompts (behaviour baked in).
- Can outperform larger models on narrow tasks.
Limitations
- Needs quality data — bad data means bad model.
- Doesn't add new facts reliably (use RAG for that).
- Locked to the base model version.
- Costs money to train and retrain.
Common misunderstandings
- ClaimFine-tuning teaches the model new facts.RealityIt shifts behaviour and style; grounding facts is still RAG's job.
- ClaimFine-tuning is always the answer.RealityFor most business problems, a well-designed RAG pipeline is faster and cheaper.
Frequently asked questions
When should I fine-tune?
When you need consistent style, when prompt engineering has plateaued, or when a smaller model must match a larger one.
How much data do I need?
Hundreds to a few thousand high-quality examples for style/format work.
Is fine-tuning expensive?
Training is one-off; inference is often cheaper than the general model over time.
The Tool Money Lab perspective
Fine-tuning is a specialist tool. In our experience, 80% of teams that plan to fine-tune should first try better prompts, better retrieval and a smaller general model.
Conclusion
Fine-tuning specialises a model. It is powerful, but it is the last mile — not the starting point — of an AI project.