What Is Fine-Tuning? – Definition and How It Works
Learn what fine-tuning does, how full and parameter-efficient methods differ, what it is good and bad at, and how to prepare a dataset.
Fine-tuning continues training a pre-trained model on a smaller, targeted dataset so it adapts to a style, format, or task. It changes behaviour reliably; it changes knowledge poorly.
1. What It Changes
The model's weights are updated, so the change is baked in rather than supplied at request time.
- Good at: output format, tone, domain vocabulary, task-specific behaviour, and reducing prompt length.
- Poor at: teaching new facts, keeping information current, or making a model cite sources.
That split is the whole decision framework, and it is why retrieval and fine-tuning are complements rather than alternatives.
2. Full Versus Parameter-Efficient
| Approach | What it updates | Cost | Practical note |
|---|---|---|---|
| Full fine-tuning | All weights | Highest compute and storage | One full model copy per task |
| LoRA and adapters | Small added matrices | Much lower | Adapters swap per task at serving time |
| Quantised variants | Adapters over a compressed base | Lowest | Runs on modest hardware |
Parameter-efficient methods reach comparable quality for most adaptation tasks at a fraction of the cost, which is why they dominate practice.
3. Preparing the Dataset
Quality dominates quantity. A few hundred consistent, correct examples usually beat tens of thousands of noisy ones.
- Consistency. Every example should demonstrate the behaviour you want, in the same format.
- Coverage. Include the variations the model will meet, including edge cases.
- Correct labels. Errors are learned as readily as correct patterns.
- A held-out set. Evaluation on training data measures memorisation.
- No secrets. Anything in the data can surface in outputs.
4. When to Fine-Tune
- Prompting has been tried and the gap is consistency, not knowledge.
- The output format is strict and repeated at volume.
- Prompt length is a material cost, and examples can move into the weights.
- Latency matters and a smaller fine-tuned model can replace a larger prompted one.
- The domain has vocabulary and conventions a general model handles awkwardly.
5. When Not To
- The information changes. Retrieval handles freshness; retraining does not.
- Facts are the problem. Fine-tuning is a poor way to install knowledge and encourages confident errors.
- Prompting has not been exhausted. Most perceived needs disappear with clearer instructions and examples.
- Data is thin. Below a threshold, fine-tuning teaches noise.
6. Evaluating the Result
Compare against the prompted baseline on the same held-out set, using task metrics rather than impressions. Watch for regression on general capability — a narrowly tuned model can lose breadth, which only a broader evaluation reveals.
7. Operational Considerations
Fine-tuned models are versioned artefacts: they need storage, deployment, rollback, and re-evaluation when the base model updates. Adapters reduce that burden considerably, since the base stays shared.
Budget for periodic re-tuning; a model tuned on last year's conventions drifts as the domain moves.
8. Summary
Fine-tuning adapts behaviour — format, tone, task shape — by updating weights, and parameter-efficient methods make it cheap enough to be routine.
It is the wrong tool for facts and freshness, which belong in retrieval. Exhaust prompting first, prepare a small consistent dataset carefully, and evaluate against a prompted baseline rather than against expectations.