Supervised fine-tuning data
Supervised fine-tuning data is a collection of prompt and ideal-response pairs written or approved by humans, used to teach a pretrained model to follow instructions.
Last reviewed
What supervised fine-tuning data contains
- Instruction or prompt
- Human-written or human-edited response
- Optional system prompt or conversation history
- Quality labels
Who buys it and how it is priced
| Attribute | Value |
|---|---|
| Belongs to | post-training data |
| Main buyers | AI labs, enterprises fine-tuning open models |
| Pricing | Priced per example, scaled by length, domain difficulty and reviewer seniority. |
Quality criteria and risks
Quality criteria
- Factual accuracy
- Style consistency
- Coverage of task types
- Deduplication against public sets
Risks
- Contamination with benchmark answers
- Responses copied from models with restrictive terms
- Narrow task coverage
Providers that list supervised fine-tuning data
Auction and pricing mechanisms that fit
- Posted-price sale: Fits catalogue datasets that can be sold many times, such as cloud data exchange listings.
- Reverse auction: Fits a buyer that knows what it needs and wants several vendors to compete, such as sourcing expert data from labs.
Key terms
- SFT
- SFT, or supervised fine-tuning, is a post-training method in which a pretrained model learns from prompt and ideal-response pairs.
- fine-tuning
- Fine-tuning is the further training of a pretrained model on a smaller dataset to adapt it to a task or style.
- post-training
- Post-training is the set of steps after pretraining, such as supervised fine-tuning and reinforcement learning, that shape a model's behavior.
- data card
- A data card is a short document that describes a dataset's contents, collection method, limits and intended use.
Frequently asked questions
- How many SFT examples does a model need?
- It depends on the task. Small, high-quality sets of thousands of examples can shift style and format, while broad capability gains need far more.