How to use this template#
A fine-tuning SOW differs from a software SOW in one respect that drives everything else: the outcome is a measured error rate, not a feature list. The template therefore anchors acceptance to an evaluation set that both parties approve before training starts. Sections in square brackets are to be completed. It is a starting point for discussion with counsel, not legal advice.
1. Scope and objective#
1.1 Task. The Provider will develop a model that [describe the task, e.g. "extracts 40
fields from commercial property claim submissions into the Client's JSON schema v3"].
1.2 Out of scope. [e.g. OCR of handwritten documents; languages other than English;
integration with the Client's claims system beyond a documented REST endpoint.]
1.3 Base model. Selected by the Provider through a documented bake-off of at least
[3] candidate models with licences permitting the Client's intended commercial use.
1.4 Environment. All training and evaluation will run in [Client AWS account / on-prem
cluster]. No Client data will leave that environment.2. Success metric and acceptance#
2.1 Evaluation set. Within [10] business days of kickoff, the parties will agree an
evaluation set of at least [1,500] items, sampled from [source] and stratified by
[segments]. Labels will be adjudicated by [two] Client experts. The set is frozen on
approval and will not be used for training or model selection.
2.2 Primary metric. [Field-level exact match], computed by the evaluation harness
delivered under 4.3.
2.3 Acceptance threshold. The model is accepted if, on the frozen set:
(a) the primary metric is at least [93.0%];
(b) the lower bound of the paired 95% bootstrap interval of the difference versus
the Baseline (defined in 2.4) is above zero;
(c) no segment listed in Annex A scores below [85.0%];
(d) all regression assertions in Annex B pass;
(e) p95 latency is below [500 ms] on [hardware].
2.4 Baseline. The Client's current production system as measured on the frozen set in
week [1].Clause 2.3(b) ties acceptance to a statistically meaningful improvement rather than a point estimate; size the evaluation set so it can detect the improvement you care about (sample-size guidance). If any metric is produced by an LLM judge, add a clause requiring judge calibration against human labels (why).
3. Data responsibilities#
| Item | Responsible party | Due |
|---|---|---|
| Representative examples (≥ [500]) for scoping | Client | Kickoff |
| Access to historical records in the training environment | Client | Week 1 |
| Domain experts for guideline writing and adjudication ([3] hours/week) | Client | Weeks 1–4 |
| Annotation guidelines | Provider, approved by Client | Week 2 |
| Additional expert labelling (up to [10,000] items) | Provider | Weeks 2–4 |
| PII/PHI de-identification before data reaches training | [Client / Provider under BAA] | Week 1 |
Delays in Client-supplied items extend the timeline day for day. That single clause prevents most schedule disputes.
4. Deliverables#
4.1 Trained model weights or adapters, and merged/quantized serving builds.
4.2 Training code, configuration files and data-processing pipeline.
4.3 Evaluation harness, frozen evaluation set, and regression suite.
4.4 Data card, model card and evaluation report with confidence intervals.
4.5 Deployment in [environment] with runbook and monitoring dashboards.
4.6 Knowledge-transfer sessions ([2] x 2 hours) for the Client's engineering team.5. Intellectual property and data#
5.1 Ownership. All deliverables, including trained weights, are owned by the Client on
payment. The Provider retains ownership of pre-existing general-purpose tooling and
grants the Client a perpetual, royalty-free licence to use it with the deliverables.
5.2 Data use. Client data, labels and derived models are used solely for this SOW and
are never used to train models for any other party.
5.3 Deletion. Within [10] business days of acceptance the Provider deletes all Client
data and intermediate checkpoints in its custody and certifies deletion in writing.
5.4 Base-model licence. The Provider will document the licence of the selected base
model and confirm it permits the Client's intended use.6. Change control and termination#
6.1 Go/no-go. At the end of week [2] the Provider delivers a feasibility report. If the
report recommends not proceeding, either party may terminate and the Client pays only
for work to that point; the evaluation set and harness are delivered.
6.2 Changes to the task, metric, threshold or evaluation set require written agreement
and may change price and timeline.The go/no-go clause makes the first two weeks a cheap option on the whole project, which is how our feasibility sprint is structured.