Dataset versions
Publish a reproducible set of image references and annotation revisions.
Before you start
At least one eligible image must have Route Training and a saved revision. The image bank must be available.
Quick start
- Check Training selection and choose a parent version when extending an earlier dataset.
- Set New image split: Automatic distributes new images; explicit Train, Val or Test keeps the chosen role.
- Choose Publish next version, review its name and version, then publish.
- Select the version and inspect its split counts before training. Both Train and Val are required for training.
Functions and choices
Publish changes
Publishing needs eligible changes. Correct annotation problems in Annotator. Route-read errors appear above the list; saved versions remain available.
Import and export
Import checks files and class mapping. Export checks format and destination.
Check older data
Inspect Dataset checks references. Historical migration starts with an inventory and a dry run.
Use the menu
The menu includes inspection, Model Journal, Training Queue and Guard Set. Command–Shift–P opens publication.
Result
A versioned manifest with explicit Train, Val and optional Test membership.
Next step
Choose the published version in Training. If Val is empty, prepare validation images and publish a new version.