Dataset versions

A dataset version is a fixed, read-only copy of a dataset at a point in time. Create one when your data is ready, then keep working on the live dataset: the version never changes, so exports and training runs made from it are reproducible.

What a version captures

A version copies every ready asset in the dataset, with its metadata:

• Image datasets: the images, labels, annotations, and the review status of each image.
• Document datasets: the source files, the extracted text of every processed page (including your corrections), figures, and the review status of each document.

Assets that are still uploading or being validated are not included. Later edits to the live dataset do not affect existing versions.

Create a version

1. Open the dataset and select Versions.
2. Select Create version.
3. In Create dataset version, select Create version.

The version is created in the background, so you can leave the page. Its row shows the status, when it was captured, how many assets, labels, annotations, or pages it holds, and the archive size.

Status	Meaning
Queued	The version is waiting to be created.
Creating	Annota is copying the data.
Ready	The version can be used for exports and training.
Failed	The version could not be created. Create a new one to retry.

Only one version can be created at a time per dataset. If creation fails, the person who requested the version gets an email. Version numbers are never reused, also after you delete a version.

Version limit

Each dataset keeps a limited number of versions. The counter next to Create version shows how many you have saved and the limit. Ready versions and versions being created count toward it; failed versions do not. When you reach the limit, delete a version you no longer need.

Use a version

• Export: choose the version as the Source when you create an export.
• Train a model: on a ready version of an image dataset, select Train new version to train a model on it.

Delete a version

1. In Versions, select the delete icon on the version’s row.
2. Confirm in Delete dataset version?.

Deleting a version removes only the copy. The live dataset is not affected. You cannot delete a version while it is being created or exported, or while a model trained on it exists. Such a version shows Retained by model training: delete the models trained on it to release it.

Note

Creating versions uses no credits, but version archives count toward your organization’s storage. Creating and deleting versions needs edit access to the dataset. People with view access can see the list. See Sharing and access.

Next steps

• Export datasets
• Train a model
• Review assets