Supervised fine-tuning data. Examples a practitioner would sign.
In fine-tuning, the model becomes what it is shown. eeko commissions demonstrations, reasoning traces, and multi-turn dialogues from credential-checked specialists, writes them to a versioned style guide, and has a second expert check every item before it counts. Pilot set first, fixed price.
What is supervised fine-tuning data?
Supervised fine-tuning data is a set of prompts paired with the responses you want the model to give. After pre-training, it is the first thing that shapes how a model behaves: what it knows how to do, how it structures an answer, when it hedges, when it asks a question, and when it declines. The model copies the examples. If the examples are subtly wrong, the model is confidently wrong.
That puts the weight on authorship. A few thousand demonstrations written by people whose colleagues would act on that answer usually move a model further than a hundred thousand scraped or model-generated ones. It also puts weight on consistency: thirty experts writing in thirty voices teach a model to be thirty assistants.
eeko fine-tunes the assistants we ship inside enterprise knowledge systems and agent deployments, so our SFT data is built the way we need it ourselves: specialist authors, one versioned style guide, a second specialist on every item, and an honest record of whether a human wrote it.
SFT datasets, from coverage plan to receipt.
Six pieces of work, scoped against one written specification. Most engagements use four or five of them.
- Clinical, legal, financial, scientific, technical
- Short answers to long-form documents
- Notation and assumption rules set in the guide
- Step-level review
- Clarify, correct, refuse, recover
- Tool-use transcripts to your schemas
- Calibration set before production
- Version stamped on every item
- Return-with-comments loop
- Reviewer ID on the record
- Task and difficulty taxonomy
- Volume directed by eval results
Expert demonstrations for LLMs, one item at a time.
Volume comes from running this loop many times in parallel, not from loosening it.
From brief to accepted item
- Brief — the writer receives a prompt from the coverage plan, the current guide version, and the acceptance criteria for that task type.
- Draft — a specialist writes the response, trace, or dialogue the way they would for a colleague who will rely on it.
- Attest — the writer records how the item was produced: human-made, AI-assisted, or AI-generated.
- Check — intake fingerprints the text, scans for generator signatures, and compares against registered and known material.
- Review — a second expert checks correctness, completeness, and guide adherence, and returns it with comments if it falls short.
- Accept — the item, its origin class, its guide version, and its review record join the hash-chained trail for the order.
Returned items are revised by the original writer and reviewed again. An item that cannot be made to pass is dropped and replaced, never delivered with a note.
A reasoning trace that did not pass first time.
An illustrative item from a commissioned set of worked financial-analysis problems, followed from draft to acceptance.
From coverage plan to a signed set.
Fixed scope, fixed price, pilot before volume. Standing orders repeat the loop as your evals find new gaps.
Instruction tuning data by industry and discipline.
Who writes, who reviews, and what the guide has to say about caveats and citations all depend on the field.
Supervised fine-tuning data questions.
What is supervised fine-tuning data?
Supervised fine-tuning data is a set of prompts paired with the responses you want the model to produce, used after pre-training to teach behavior by example. Each example shows the model what a good answer looks like in your domain: its content, its structure, its tone, and how it handles uncertainty. The model imitates what it sees, which is why who writes each example matters more than how many there are.
How many SFT examples do we need?
Fewer than most teams expect, if they are the right ones. The number turns on what behavior you are teaching and which base model you start from. We usually start with a focused pilot set across your highest-priority task types, you measure the effect on your own evals, and then we scale only the categories that moved. The coverage plan sends volume to where it changes the model instead of fixing a count in advance.
Do you write reasoning traces as well as final answers?
Yes. For math, science, code, and analysis we commission full worked solutions with every intermediate step visible, and the second reviewer checks the steps, not only the conclusion. How much working to show, where to state assumptions, and what notation to use are all set in the style guide, so traces from different writers read as one consistent assistant.
How do you stop writers from using AI to draft demonstrations?
Each writer attests to how an item was produced, and the record stores it as human-made, AI-assisted, or AI-generated. Intake scans for generator signatures in metadata and runs text fingerprints against registered and known material. The second-expert review catches the familiar shape of unedited model output. If your specification requires fully human-written items, anything marked assisted or generated is left out.
How much does SFT data cost?
Per item, against the agreed specification. The price moves with the expertise needed, how long and hard each demonstration is, whether a reasoning trace is included, how many turns a dialogue runs, and how deep the review goes. A long clinical or legal answer with a second specialist's review costs more than a brief general-purpose answer. Volume pricing applies to larger and standing orders, and the pilot set is priced on its own.
Can we use our own style guide?
Yes, and most teams with a deployed assistant should. Writers learn your guide and pass a calibration set before touching production. If you do not have one, we write one together during scoping. In both cases the guide carries a version number, each item stores the version it was written under, and mid-project changes are tracked so older items can be filtered or revised.
Can SFT data include tool use and agent transcripts?
Yes. Writers can produce transcripts where the assistant calls tools, reads results, recovers from errors, and asks for clarification, following the tool schemas you provide. These items need writers who have worked with the workflow in question, which we account for in selection. It is the same kind of demonstration data we use to shape the agents eeko builds for its own enterprise clients.
Related data services.
Show the model what good looks like.
Five minutes to send a dataset request. Within one business day a data lead comes back with questions, a draft coverage plan, and a per-item price. No deck. A ranked list of what to write first.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: