SFT
Supervised fine-tuning. A model is trained further on examples of the behavior you want: for each prompt, one good answer.
AuraOne / Human Data / SFT data
SFT data from qualified experts.
SFT (supervised fine-tuning) data is a set of prompts paired with the answers you want a model to give. Qualified experts write and correct those answers for AI teams that fine-tune language models on their own tasks.
Every dataset starts from an explicit task and rubric. Uncertain work goes to review. Quality, rights, and delivery records leave with it.
- Input
- Source examples, task instructions, your rubric, and the expertise the task needs
- Output
- Accepted examples in your schema, with the review history and the delivery manifest
The terms
What SFT data is
Four terms, in plain words.
Instruction data
SFT examples where the prompt is an instruction or a question and the answer follows it.
Demonstrations
Answers written or corrected by a person to show the model what a good answer looks like.
SFT versus preference data
SFT shows the model one good answer. Preference data compares several answers and says which is better. See RLHF and preference data.
What you receive
What an SFT-data program delivers
Your contract sets the exact scope. Model Improvement also lists SFT, DPO, and reward-model dataset exports.
Written and corrected examples
Qualified specialists create or assess difficult examples. Where the model is wrong, the expert says what right looks like, and why.
- Examples created or corrected against your task and rubric
- Expert corrections and written rationales
Review history
Uncertain work goes to review. Each example keeps the record of who checked it and how disagreements were settled.
- The review and adjudication history for each example
- Quality, rights, and delivery records
A dataset you can load
Accepted examples arrive in the schema and file format you ask for, with a manifest that names the rights and the acceptance criteria.
- Accepted examples in the requested schema and file format
- The rights record and the delivery manifest
How an SFT program runs
Four steps, each with its own record.
01
Define and calibrate
Confirm the instructions and the rubric. Confirm reviewers are ready before production.
Task version and calibration result
02
Write and correct the examples
Experts write or correct each answer against the approved task.
Item-level action and reviewer identity
03
Review and adjudicate
Check quality, resolve disagreements, and retain exception decisions.
Quality sample and adjudication log
04
Approve delivery
Verify every file against its checksum. Record the rights and acceptance criteria in the manifest before handoff.
Delivery approval and a signed manifest checksum
Choosing
SFT data, preference data, or both?
Teams usually decide by reading the model's current answers.
Start with SFT data
When the model does not yet answer in the form, style or depth your task needs. Examples show it what good looks like.
Add preference data
When the model can already produce several plausible answers and you need it to favor the better ones.
Measure either way
Keep a held-out evaluation set so you can tell whether the new data helped. Model Improvement builds task-specific evaluation sets and compares a candidate with its baseline.
Scope and limits
What this program does not do
Data, not training
AuraOne prepares the examples and the record. Training, fine-tuning, and model-performance gains are separate scopes.
Rights come first
Use terms are agreed per program, and the rights record travels with the data.
Check model-written drafts
Ask any provider how model-written drafts are reviewed before they become training data. Our guide to synthetic data and model judges explains what to validate.
Questions
Questions about SFT data
What is SFT data?
SFT data is the training data for supervised fine-tuning: a set of prompts, each paired with the answer you want the model to give. It is sometimes called instruction data or demonstration data.
What is the difference between SFT data and RLHF data?
SFT data shows the model one good answer for each prompt. RLHF data, also called preference data, compares several answers and says which one is better. Both come from expert judgment, and teams often use both.
Who writes the examples?
Qualified specialists create or assess difficult examples against your task and rubric. On expert programs, a second reviewer checks the work, and an adjudicator settles disagreements against criteria you control.
What arrives with the data?
Accepted examples in the schema and file format you ask for, the review and adjudication history for each example, the rights record, and a delivery manifest. Every file is verified against its checksum before handoff, and your contract sets the exact scope.
Does AuraOne fine-tune the model?
An SFT-data program delivers the examples and the record, not a fine-tuned model. Training, fine-tuning and model-performance gains are separate scopes, agreed separately.
How many examples do I need?
It depends on the task, the model and the quality bar, so volume is scoped per program. A sensible start is a small, carefully reviewed set that you can measure on your own evaluation before you scale it. A program starts as a pilot with an agreed scope, a defined deliverable and an acceptance check.
Guides
Related reading
Synthetic Data and Model Judges: What to ValidateSynthetic data can expand coverage, but quality depends on task design, calibrated judges and measured human review. Learn what to validate before using it.How to Hire Domain Specialists for AI Data WorkA practical workflow for qualifying, calibrating and reviewing domain specialists for AI data, with evidence checks for buyers and prospective reviewers.
Tell us what your model needs to learn.
We scope the specialists, the review, and the rights with you. Related pages cover RLHF and preference data, data labeling and annotation and the Human Data overview.