Skip to content

Datasets – Version Management

The Datasets page lets you manage versioned snapshots of your annotation data for a project. Access it from the sidebar (Datasets) while a project is open, or navigate to /projects/:projectId/datasets.

Dataset versioning separates live staging data (actively being annotated) from frozen snapshots (locked versions ready for training or export).

Accessing Dataset Versions

  1. Open the Annotations module and go to Projects.
  2. Click a project to open the Project Dashboard.
  3. Click Datasets in the sidebar, or navigate via the project context.

If no project is selected, clicking Datasets in the sidebar prompts you to select a project first.

Page Layout

The page is divided into two panels:

Left Panel – Version History

A timeline of all dataset versions for the project:

Version Description
v0 – Live Staging Data Real-time sync of current project images and annotations. Status: dynamic (updates as you annotate).
v1, v2, … Frozen snapshots created from the live staging pool. Status: locked (immutable).

Each version card shows:

  • Version tag (e.g., v0, v1)
  • Version name
  • Created date (or "Real-time Sync" for v0)
  • Total image count

Click any version to view its details in the right panel.

Live Staging Data (v0)

The v0 version represents your project's current, in-progress dataset:

  • Status: Dynamic — counts update in real time as images and annotations are added
  • Volume Metrics:
  • Total Images – All images in the project
  • Instances – Total annotation instances across all images
  • Unannotated Images – Images without any annotations
  • Pipeline Split: Unassigned — splits are configured when you freeze a version
  • Class Distribution: Live bar chart showing instance counts per label

Freeze Version

When your staging data is ready, click Freeze Version to create an immutable snapshot:

  1. A Configure Pipeline Splits modal opens.
  2. Set the train/validation/test split percentages (must total 100%):
  3. Training Set (%) – Default: 70%
  4. Validation (%) – Default: 20%
  5. Test Set (%) – Default: 10%
  6. Click Confirm & Freeze.

The system creates a new frozen version (e.g., v1 – Frozen Snapshot v1) with the configured splits and locks the current annotation state.

Frozen Versions (v1+)

Frozen versions are read-only snapshots with finalized metrics:

Volume Metrics

  • Total Images
  • Instances (annotation count)
  • Unannotated Images

Pipeline Split

A visual bar showing the train/validation/test distribution set during freeze:

  • Train (blue)
  • Val (purple)
  • Test (pink)

Class Distribution

A scrollable list of all labels with instance counts and relative bar charts.

Export

Click Export on a frozen version to download annotations in your chosen format. See Export formats for supported formats.

Role-Based Access

Your role badge is displayed in the page header:

Role Access
Admin Full access — freeze versions, export data
Member View version history and metrics
Annotator View version history and metrics

Team management requires a Pro subscription. See Upgrade Plan.

  • Project Name (uppercase link in header) – Returns to the Project Dashboard
  • Dataset Versions title – Current page context

Workflow Summary

  1. Annotate images in job batches within your project.
  2. Monitor live metrics on the v0 Live Staging Data version.
  3. When ready, Freeze Version with your desired train/val/test splits.
  4. Export the frozen snapshot for model training or external use.
  5. Continue annotating — v0 keeps updating while frozen versions remain locked.