Fine-tuning: definition, use and boundaries
Why the Term Matters
Many teams reach for fine-tuning too early because it sounds like the most thorough fix. Often a better prompt or a link to your own documents is enough. Fine-tuning is like an employee who drills a task until it sits. It pays off when a pattern always runs the same way and is hard to put into words. Whoever knows the difference saves effort and picks the right lever.
When It Is Relevant
- A format or style should sit reliably, without being re-explained on every request.
- There are enough clean examples of the desired input and output.
- A smaller, cheaper model should handle a narrow task as well as a large one.
- Prompt engineering hits a wall even though the instruction is carefully worked out.
When Not
When it is about current or frequently changing facts, fine-tuning is the wrong route, because imprinted knowledge ages fast. Then RAG fits. And when only a few or messy examples exist, retraining leads to unreliable behaviour. Only once a pattern is clear and enough evidence for it exists does fine-tuning become a sensible investment.
Industries That Use It
- Financial services: bring recurring reports and filings into a fixed, checked format.
- Media and publishing: produce text reliably in a defined house voice across many pieces.
- Industrial manufacturing: turn free text from inspection records into structured fields, every time.
- Customer service: classify requests into fixed categories, where a narrow model works faster and cheaper.
Typical Use Cases
- A model that sorts support tickets reliably into a fixed category scheme.
- A text generator that writes product descriptions consistently in the company tone.
- An extractor that pulls the same fields from documents in the same shape every time.
- A small model that takes over a narrow routine task more cheaply than a large one.
- A reshaping of technical jargon into a standardised, checkable output.
Decision Guide
| Signal | Fine-tuning fits | Alternative |
|---|---|---|
| Style or format should sit firmly | fine-tuning with clean examples | try prompt engineering first |
| Current facts needed | rather not | RAG with a knowledge base |
| Few examples available | gather data first | prompt engineering bridges the gap |
| Cost per request too high | fine-tune a small model | check a larger model |
Example
A financial services firm produces many similar filings and wants them in an exact format. In the workshop the team checks whether a careful prompt already suffices. Because the format is very strict and many checked past cases exist, a small dataset for fine-tuning is built. Testing runs against real examples with tricky edge cases. Only then does the model go into operation. This describes the method, not an asserted client metric.
Service Connection
Whether fine-tuning or a lighter route fits is settled in the AI workshop. The build itself belongs to AI development, operation to AI implementation.
Sources
This framing describes common practice when adapting models. Effort, data volume and legal questions depend on the model and the project and should be checked there.