Manage splits, collaboration, and advanced dataset actions

Review the current counts and whether the dataset already has split membership.

Enter Training, Validation, and Test ratios whose total does not exceed 100 percent.

Use Random for general assignment or Stratified to preserve label or class balance when the dataset supports it.

Use an integer seed when the same random assignment must be reproducible.

Review the resulting counts and policy, then create the splits and inspect representative samples in each tab.

Confirm the existing ratios, strategy, seed, and lock state before changing anything.

Use the update operation to place unassigned additions without rewriting established Training, Validation, or Test membership.

Confirm any intended growth in Validation or Test and record the updated counts.

Choose the irreversible lock only after the evaluation membership is approved. Once locked, new samples cannot be assigned to Test and test controls are disabled in later add workflows.

Compare pre- and post-update membership and confirm that prior assignments stayed unchanged.

Clear search and structured filters for a full copy. For a filtered clone, verify the selected split, search term, and filter chips; selecting a split by itself does not activate filtered cloning.

Treat the source count shown by the dialog as a preview of the active query, then enter a clear destination name and description. The source is evaluated again when you submit.

Enable split preservation to retain matching Training, Validation, and Test membership together with available lock, seed, and strategy settings.

Wait for cloning to complete, then use the completed destination total as the authoritative clone count and compare its samples, labels, annotations, and split counts with the submitted source scope.

Use Private for restricted access, Organization for organization-wide visibility, or Global only when broad sharing is intended and allowed.

Open collaborator controls, add or remove the intended people, and verify their effective access without broadening visibility unnecessarily.

Organization datasets are visible to organization users, while editing Global content depends on workspace permissions even when it can be viewed.

Archive datasets that should leave active workflows and restore them only when they should return to use.

Treat deletion as permanent, read the confirmation carefully, and verify the exact dataset name and ownership before proceeding.

Open Run Inference for compatible model predictions or Extract Slices for an eligible DICOM dataset.

Choose the deployment and samples for inference, or choose the offered slice-selection controls for DICOM extraction.

Follow the queued and running state until the operation completes or fails.

Review inference results or extracted slices on representative samples and record any failed items before retrying.