Fine-tuning takes an already-trained model and further trains it on your own curated dataset — adjusting its weights to change behavior, tone, format consistency, or a specialized skill, rather than teaching it brand-new general knowledge.
Where Fine-Tuning Fits in the Bigger Picture
Base model (pretrained)
↓ instruction tuning + alignment (done by the model provider)
Instruction-tuned, aligned model — what you get via an API
↓ YOUR fine-tuning (optional, on your own data)
A model further adapted to your specific behavior/style/taskFine-tuning is almost always applied on top of an already capable, instruction-tuned model — not a from-scratch training process. See LLM Training and Instruction Tuning for the training stages a model goes through before it ever reaches you.
What Fine-Tuning Is Good At
- Consistent output format — a very specific structured format prompting alone struggles to hit reliably at scale
- Tone/style consistency — a distinctive brand voice applied consistently across thousands of varied inputs
- Specialized task performance — a narrow, well-defined task where many labeled examples exist
- Reducing prompt length/cost — behavior baked into the model's weights instead of lengthy few-shot examples repeated in every prompt
What Fine-Tuning Is NOT Good At
Teaching a model new, specific factual knowledge (a product catalog, current events, internal documents) is generally better solved with RAG — fine-tuning doesn't reliably make a model "memorize and recall" facts the way retrieval does. See RAG vs Fine-Tuning for the full comparison.
What This Section Covers
| Note | Focus |
|---|---|
| Fine-Tuning vs Prompting | When fine-tuning is worth the added cost/complexity vs prompting |
| Supervised Fine-Tuning | The standard example-based training approach |
| LoRA, QLoRA, Parameter-Efficient Fine-Tuning | Making fine-tuning cheaper and more accessible |
| Fine-Tuning Dataset | What good training data actually looks like |
| Fine-Tuning Evaluation | Verifying the fine-tuned model actually improved |
| Best Practices | Practical guidance for a fine-tuning project |
Common Mistakes
- Reaching for fine-tuning as the first solution to a problem prompting or RAG could solve more simply and cheaply
- Fine-tuning to add factual knowledge instead of using RAG — a very common, avoidable misapplication
Interview Relevance
Q: "When would you choose fine-tuning over prompting or RAG?" — behavior/format/style consistency at scale, where prompting alone proves unreliable, is the expected core answer — not "for adding new knowledge."
Practice Question
A team wants their model to always output responses in a very specific 5-field JSON structure with 99%+ reliability, and prompting alone gets it right only 85% of the time. Discuss whether fine-tuning is a reasonable next step.
