> For the complete documentation index, see [llms.txt](https://docs.unitlab.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.unitlab.ai/solutions/architecture-and-adoption-guides/evaluation-and-rollout.md).

# Evaluation and rollout

Run a small pilot on your own data, check the result, then scale to production.

Start small, prove the result on your own data, then scale.

## 1. Pilot (1 to 2 weeks)

* Upload 50 to 200 items that include your hard cases: occlusion, low light, rare classes.
* Build the ontology and Instructions, and label the pilot with 1 to 3 annotators.
* Add a Review stage and, if you have a model, a Model stage for pre-labels.

## 2. Check

* Export a release in the format your training code reads and load it end to end.
* Look at **QA**: rejection rate, quality-check findings, and consensus if you used it.
* Measure time per item with and without pre-labels.

<figure><img src="https://2337177515-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fi6jXAxtoN6UXTLAJFexq%2Fuploads%2FNfuVL4ZdsvleLZAbCCBC%2Fsol-releases.webp?alt=media" alt="Releases list with version v1, COCO format, split and download button" width="800"><figcaption><p>The pilot is done when a release loads in your pipeline without fixes.</p></figcaption></figure>

## 3. Scale

* Freeze the ontology version and Instructions; change them through a new version, not in place.
* Add annotators and reviewers with the right roles.
* Connect cloud storage for large data and automate uploads and releases with the SDK.
* Review QA results every week and feed failures back with [Model-in-the-loop iteration](https://docs.unitlab.ai/solutions/program-design-guides/model-in-the-loop-iteration.md).
