Term: Fine-tuning
~6 min read
Estimated time: ~6 min read — for the in-app brief plus opening the primary source.
What this is
Fine-tuning further trains a general model on specialized examples so it behaves more like your domain — style, formats, or task patterns.
Everyday example
A company wants every customer email draft in a strict brand voice. Fine-tuning (or strong prompt libraries) can teach that style. It still will not know yesterday’s unpublished price list unless you also ground it with RAG or tools.
Fine-tuning further trains a model on your examples so it leans toward your style or labels — it is not the same as giving it your files at ask-time.
- Good for a house voice or a narrow classifier.
- Bad as the first answer to “make it know our 400 policies” — that is RAG.
- You inherit data-rights and eval duties for the extra training.
- Behavior changes can be hard to unwind.
Next action: If a vendor proposes fine-tuning, ask why RAG or a system prompt is not enough.
What changes in how you lead
How decision rights, process, and ownership should change.
- Purchasing treats fine-tune as a change to the model, with data and exit terms.
- Do not fine-tune on personal or secret data without a written path.
Compare related ideas
Fine-tuning vs RAG
Choose RAG when answers must cite current company documents. Choose fine-tuning when you need consistent behavior/style on a task and have good example sets — many programs use both.
Open RAGDeep dive
Not a substitute for clean process design or live data access.
Ask vendors what was fine-tuned, on whose data, and how regression is tested after updates.
Compare options: prompt library (fastest), RAG (your documents), fine-tuning (behavior/style), full custom training (rare and expensive).