Human-authored SFT dataset creation service
We write supervised fine-tuning (SFT) datasets by hand: instructions, step-by-step reasoning, expert answers and conversations by vetted professionals. No LLM output, double-reviewed.
Free samples made to your spec, reply within 48 hours.
In short
An SFT (supervised fine-tuning) dataset is a set of prompt–response pairs that shows a language model exactly how to answer. We create them with people, not with another model: domain experts write the prompts and responses, a second expert reviews every record, and we screen for copied and LLM-generated text before delivering JSONL with a datasheet and commercial license.
Why human-authored SFT data
SFT data sets the ceiling for how your model writes, reasons and refuses. Distilling responses from another LLM is fast, but it copies that model's mistakes and style, it can conflict with the provider's terms of use, and it makes your model sound like everyone else's.
Human-authored data carries real expertise: how a Vietnamese teacher explains a proof, how an accountant reasons about a tax edge case, how a support agent calms an angry customer.
What we can write
- Instruction–response pairs for general or domain assistants
- Step-by-step reasoning and chain-of-thought solutions (math, code, professional decisions)
- Expert Q&A in law, accounting and tax, agriculture and health
- Multi-turn conversations for chat and support assistants
- Preference pairs (chosen vs rejected) for DPO and RLHF
- Held-out evaluation sets with rubrics that never appear online
How a project runs
1. You send the task, target domain, language, format and volume. 2. Within 48 hours we send example records written to your spec and a fixed quote. 3. After you approve, we recruit and brief the writers, deliver in batches and fold your feedback into each batch. 4. The final delivery includes the data, schema, datasheet and license.
Quality controls on every record
- Writers vetted for the domain (teachers, engineers, lawyers, accountants, clinicians)
- Second-expert review of every record; a lead reviewer resolves disagreements
- Screening for plagiarism and LLM-generated writing
- Style guide and schema agreed with you before production starts
- Per-record metadata: author role, reviewer, difficulty, topic
SFT data sources compared
| Distilled from an LLM | Crowdsourced | Expert-authored (ours) | |
|---|---|---|---|
| Domain accuracy | Inherits model errors | Variable | Checked by a second expert |
| Style diversity | Low | Medium | High |
| Licensing risk | Provider terms may apply | Low | Low, documented rights |
| Cost per record | Lowest | Low | Higher, fewer records needed |
Matching datasets
Start from a catalog dataset with free samples, or ask us to adapt one to your spec.
- Vietnamese Math ReasoningReasoningSecondary-school and university-entrance math problems in Vietnamese, each solved step by step by a qualified teacher and checked by a second one.On requestFrom $490
- Professional Reasoning TracesReasoningHow experienced accountants, supply-chain planners, marketers and engineers think through real work decisions, written out step by step by the professionals themselves.On requestFrom $790
- Code Debugging ReasoningReasoningRealistic bugs in Python, TypeScript, Java and SQL with the step-by-step reasoning a senior engineer used to find, explain and fix them, plus a test that proves the fix.On requestFrom $690
- Vietnamese Legal Q&AExpert Q&AEveryday questions on labour, family, land, business and traffic law in Vietnam, answered by licensed lawyers with citations to the legal documents in force.On requestFrom $990
Tell us what you are building
Describe the data you need in two minutes. Within 48 hours we send samples made to your spec, a fixed price and a delivery date. You pay only after you approve the samples.
Request samplesFrequently asked questions
How much does a custom SFT dataset cost?
It depends on domain, length and review depth. Catalog datasets show package prices on each page; custom work is quoted after we send sample records, usually within 48 hours.
How many SFT examples do I need?
For a focused skill or domain, a few hundred to a few thousand high-quality examples often beat much larger synthetic sets. We usually suggest a pilot of 200–500 records first.
Can you follow our existing schema and guidelines?
Yes. Share your schema, style guide and examples; we write to them and can sign an NDA first.
Can the dataset be exclusive to us?
Yes. With an Exclusive license the records are made for you only and never licensed to anyone else.
Further reading
Last updated .