Supervised fine-tuning (SFT) is the standard fine-tuning approach — training on labeled (input, ideal output) example pairs, the same core technique used in instruction tuning, just applied to your own specific dataset instead of a general-purpose one.
The Core Idea
Training examples (input, ideal output pairs):
Input: "Customer: My package arrived damaged, what do I do?"
Output: "I'm sorry to hear that! Please reply with your order
number and a photo of the damage, and we'll process
a replacement within 24 hours."
Input: "Customer: Can I change my shipping address?"
Output: "If your order hasn't shipped yet, reply with your
order number and the new address, and we'll update it."
... (potentially hundreds to thousands of examples)The model adjusts its weights during training to better match this pattern of input → ideal output, generalizing the demonstrated style, tone, and format to new, unseen inputs.
The Training Process, Conceptually
1. Collect a dataset of (input, ideal output) pairs representing
the exact behavior you want.
2. Split into training and validation sets.
3. Run the fine-tuning job (via your provider's fine-tuning API
or your own training infrastructure).
4. Evaluate the resulting model against the validation set AND
real-world examples it wasn't trained on.
5. Compare against the un-fine-tuned baseline — did it actually
improve on the metric that matters?How Much Data Is Needed?
There's no universal number — it depends on task complexity, how different the desired behavior is from the base model's default behavior, and data quality. Providers offering fine-tuning APIs typically publish minimum/recommended dataset sizes for their specific offering — quality and consistency of examples generally matters more than raw quantity, and a smaller set of carefully curated examples often outperforms a larger set of inconsistent ones.
Practical Use Case
A customer support team with a large history of well-handled tickets (a natural source of high-quality input/output pairs) is in a strong position to fine-tune a support-response model — the real, historical data doubles as training data with minimal extra collection effort.
Common Mistakes
- Using inconsistent examples (different agents/writers with very different styles) without curating for consistency, causing the model to learn an unclear, muddled pattern
- Not holding out a validation set separate from training data, making it impossible to honestly measure whether the model actually generalized well
- Skipping comparison against the un-fine-tuned baseline, so you can't actually confirm fine-tuning helped
Interview Relevance
"What does a supervised fine-tuning dataset actually look like?" — labeled (input, ideal output) pairs demonstrating the exact desired behavior, at the scale and consistency needed for the model to generalize the pattern.
Practice Question
Write 2 example (input, ideal output) training pairs for fine-tuning a model to write concise, empathetic responses to negative product reviews.