Label Sets - BETA

• “Which exact labels did we train model v3 on?”
• “How many labels did the team add between last month’s delivery and this one?”
• “Can I re-export the dataset we delivered to the customer in March?”

Key Concepts

Who Can Use Label Sets

Create a Label Set

1. Open the Project.
2. Click Versioning in the Project header. The Versioning drawer opens on the Label sets tab. If the Project has no label sets yet, the drawer shows No versioning configured and the Versioning button shows a warning dot.
3. Click Create label set. If the Project already has a label set, click New label set instead.
4. In the New label set dialog, specify the following:
   • Name: A meaningful name, for example Vehicle detection or Customer delivery.
   • Which labels to capture: Select one of:
     • Every label on the main branch: Captures all labels in the Project, regardless of where their tasks are in the Workflow.
     • Only labels in a workflow stage: Captures only the labels on tasks that are in the Workflow stage you select. For example, select your Complete stage to capture only finished, approved labels.
5. Click Create label set.

Create a Version

1. Click Versioning in the Project header.
2. Find the label set and click Create version. The dialog summarizes what the version will capture, for example Captures labels in Complete. Those labels are frozen into the new version.
3. Optionally, enter a Version name, for example Training run 2026-11. If you leave the name blank, the version is named after its number, for example v4.
4. Click Create version.

Read the Version History

• The version number (v1, v2, …) and its name, if you gave it one.
• The number of labels the version holds, or its status: Capturing… while the version is being built, or Capture failed if it could not be built.
• The date the version was created.
• A Manual tag for versions you created manually, and a Latest tag on the newest version that is ready to export.

Export a Version

1. Click Versioning in the Project header.
2. Find the version in the label set’s version history and click Export.
3. In the Export dialog, specify the following:
   • Format: JSON or COCO.
   • Generate signed URLs: Specify whether to generate signed URLs for the data units. Enabling this option can significantly increase export times.
   • Data group label format: New or Legacy. See Data Group Label Format Options.
4. Click Export.

{
  "label_set": {
    "id": "<label-set-id>",
    "name": "Vehicle detection",
    "description": "",
    "origin": "workflow"
  },
  "version": {
    "id": "<version-id>",
    "version_number": 4,
    "name": "Training run 2026-11",
    "cut_at": "2026-11-03T10:15:00+00:00",
    "cut_by": "<user>",
    "label_count": 12840
  },
  "sources": [
    {
      "captured_source_id": "<id>",
      "source_type": "project_branch",
      "project_hash": "<project-id>"
    }
  ]
}

Edit a Label Set

1. Click Versioning in the Project header.
2. Click Edit configuration on the label set.
3. Change the Name or the Which labels to capture selection.
4. Click Save.

Limitations

• One Project, main branch only. A label set captures labels from the main branch of a single Project.
• Manual versions only. Scheduled versions (for example, every 24 hours or every 1,000 completed tasks) are not yet available. A label set’s trigger shows Manual only.
• Data groups. Versions cannot yet be created for Projects that contain Data Groups. Encord displays the error This project contains data groups, which versions cannot capture yet.
• SDK. Label sets cannot yet be created, versioned, or exported using the SDK. To snapshot labels programmatically, copy labels to a branch.

Best Practices

• Version what you deliver. Point a label set at your final Workflow stage (usually Complete) and create a version every time you hand labels to a model team or a customer.
• Name versions after the event, not the date. The date is already recorded. A name like Customer delivery 3 or Model v3 training is easier to find later.
• Keep the manifest. The label_set_version.json file in every export is your audit trail.
• Use separate label sets for separate purposes rather than changing the selection on one set. A set whose selection keeps changing produces versions that are hard to compare.