What a fine-tuning statement of work should include

A statement of work (SOW) turns a fine-tuning proposal into something you can hold a vendor to. Most disputes on these projects are not about whether training ran. They are about whether the result is good enough, who owns what, and what happens to the data. A good SOW answers those before work starts.

This is a buyer’s checklist, not legal advice. Have your own counsel read the final document alongside the master agreement it sits under.

Scope: the task in one paragraph

Start with what the model must do, stated plainly enough that someone outside the project could judge it.

• The task (for example, “extract twelve fields from supplier invoices into JSON”).
• The base model and size, or the shortlist the vendor will choose from and how.
• The tuning method: supervised, preference, reinforcement, LoRA or full.
• What is out of scope: retrieval, an application around the model, ongoing retraining.

If the scope says “improve model performance” and nothing more, there is no scope.

Data: who supplies, cleans and labels it

Item	Write down
Source data	What the buyer supplies, in what format, by what date
Cleaning	Who removes duplicates, personal details and stray test items
Labeling	Who writes or reviews examples, against which guideline
Volume	Target number of training and held-out examples
Access	Where the data lives during the project and who can reach it

Tie the timeline to data delivery. If your data arrives three weeks late, the schedule should move with it, and the SOW should say so rather than leave it to goodwill.

Evaluation and acceptance: what counts as done

This is the section to spend time on.

1. The evaluation set. Who builds it, how many items, and a rule that it is never used for training.
2. The metric. Exact match, a rubric score, a pass rate, latency, cost per call. Pick the one that reflects the business goal.
3. The baseline. The score of the current approach (often the untuned model with a good prompt), measured the same way.
4. The acceptance threshold. For example, “at least 10 points above baseline on the agreed set, at or below the agreed cost per thousand calls”.
5. What happens if it misses. A further iteration at no charge, a reduced fee, or the right to stop.

Without an agreed set and threshold, acceptance becomes a matter of opinion, and opinions get expensive.

Ownership: weights, adapters, data and tools

• Tuned weights or adapters. Who owns them, and can you export them? If the base model is open-weight, its license still applies to the result.
• The evaluation set and grader. These should belong to you. Every future retrain is judged against them.
• Training scripts and configuration. Ask for them as a deliverable so the run can be reproduced.
• The vendor’s own tools. Reasonable for the vendor to keep, with a license for you to use outputs.

Data handling and deletion

Copy the vendor’s data commitments into the SOW, not just a link to a policy page that can change.

• Training data is not used to train or improve the vendor’s other models.
• Data is stored in a named region, with named sub-processors.
• On completion (or on request), the vendor deletes your data and confirms it in writing within a stated number of days.
• Access is limited to named roles.

If a provider offers a lower price in exchange for using your data, make the choice explicit in the SOW either way.

Hosting and handover

Training is a one-off. Serving is the bill that continues. State:

• whether hosting is included, for how long and at what rate after that;
• the endpoint type (shared per-token, dedicated per hour, or your own infrastructure);
• the published price basis the vendor is using, with the date it was read;
• what the handover includes: weights, documentation, a runbook, a walkthrough.

Timeline, milestones and payment

Link payments to milestones you can check: data accepted, baseline reported, first tuned model evaluated, acceptance met. Avoid paying the bulk on kickoff.

Change control

Fine-tuning projects change shape. A new data source turns up, the base model is deprecated, a grader needs rewriting. Agree how changes are requested, estimated and approved, and at what day rate.

Security and compliance

List any review the vendor must pass (a SOC 2 report, your questionnaire, a data processing agreement) and by when. Name who is responsible for removing personal details before data is shared.

One-page summary

Clause	The question it answers
Scope	What exactly is being built?
Data duties	Who supplies, cleans and labels?
Evaluation and acceptance	How do we know it is done?
Ownership	Who keeps weights, test set and scripts?
Data deletion	What happens to our data afterward?
Hosting	What does it cost to run, and for how long?
Change control	How are surprises priced?

When we compare the best LLM fine-tuning companies, published answers to the ownership, data and hosting questions carry real weight, because they are the clauses hardest to negotiate later. If you want vendors whose published terms already match the SOW you have in mind, request a shortlist.

Sign the SOW once the acceptance clause makes sense to someone who was not in the sales call.