# Dataset and experiment releases

Canonical: https://savrn.com/cloud/docs/lab-datasets-releases

SAVRN Cloud · Coming soon. Public documentation and a browser simulation are available; connected services are not yet available.

The dataset used by a run needs a stable revision, permitted purpose and a clear relationship to its training or evaluation split. This SAVRN Cloud guide follows sample collections into experiments and explains how provenance and release decisions support later review of a candidate.

**SAVRN Cloud · Coming soon · Public preview · Browser-local simulation · No live compute or payments**

## Review a local dataset

Open [Datasets](https://savrn.com/cloud/console/#/datasets) and inspect a sample collection before using it in an environment or training run. Review its purpose, description and version. **Add dataset** accepts a name, description, purpose, rights note and 2–50 sample lines, with a 12,000-character limit. Use fictional lines. Purpose controls the actions: **Frozen evaluation set** enables **Evaluate**; **Training examples** enables **Train adapter**. **Export dataset** retains the labeled collection.

Do not enter actual research records. Browser-local persistence is not an institutional repository, rights-management system or retention service. A local approval or version label demonstrates the intended workflow only.

## Connect data to experiments

Follow a dataset through [Environments](https://savrn.com/cloud/console/#/environments), [Evaluations](https://savrn.com/cloud/console/#/evaluations) and [Training](https://savrn.com/cloud/console/#/training). A useful experiment record identifies the exact dataset revision and its split, not merely a mutable title. The held-out evaluation set should answer whether the change generalizes beyond the training examples.

A fictional extraction dataset can include straightforward, missing-field and ambiguous cases. Record how each category is represented.

## Before connected services launch

Establish provenance, permitted purpose, rights, consent where required, project access, retention and deletion obligations. Freeze release hashes and split assignments. Check duplication, contamination and sensitive content before use. Distinguish source corrections from a new dataset release. Preserve enough lineage to reproduce a checkpoint while respecting deletion requirements. Approval for inference does not automatically authorize training, and approval for one institution does not permit sharing its examples across customers.