Core concepts
Before you install tools or call APIs, learn the main ideas Labelty uses everywhere.
Organization and project
Section titled “Organization and project”- An organization is your company or team account (billing, wallet, members).
- A project is a workspace for one labeling / ML problem (one asset type and label schema).
Almost everything you do (keys, assets, training, notebooks, workflows) is scoped to a project.
Assets
Section titled “Assets”An asset is one piece of data in a project — typically an image file stored by Labelty.
Assets move through statuses (for example uploaded → annotated → reviewed → approved). Training only exports approved annotations with a label_id.
Screenshot: Project asset list showing status filters
Label schema
Section titled “Label schema”A label schema defines the classes you annotate with (name, color, options).
The SDK returns both:
- Raw schema rows (
schemas) - A normalized
definition.classeslist for convenience
You need a label_id (UUID) when sending predictions or mapping model outputs to classes.
Annotations and predictions
Section titled “Annotations and predictions”- An annotation is a labeled region on an asset (points are usually normalized to
0–1). - A prediction from the SDK is an annotation created by a model via
predictions.create_batch. - Asset status and assignee are not changed when you ingest predictions through the SDK.
Model registry
Section titled “Model registry”A project has a model registry. Each registered model can have:
- Metadata (name, version, type, framework)
- Weight artifacts
- An inference package (
inference.pyand related files) used at runtime
New models are often inactive (is_active: false) until you explicitly activate them. Predictions and many inference paths require an active model.
Training run
Section titled “Training run”A training run trains (or fine-tunes) a model on approved labeled assets using managed compute (SageMaker).
Typical lifecycle:
QUEUED → RUNNING → SUCCEEDED | FAILED | STOPPED_EARLY
After success, you still need a separate build step to package inference and activate the output model for auto-label / deployment.
Auto-labeling vs inference workflow
Section titled “Auto-labeling vs inference workflow”| Auto-labeling | Inference workflow | |
|---|---|---|
| Purpose | Label assets inside a project | Expose a production inference API |
| Trigger | Run from the Labelty UI / API on existing assets | External client calls POST …/workflows/{id}/run |
| Code | Model inference.py |
Plus a workflow.py orchestrator |
| Output | Annotations on project assets | JSON predictions to the caller (optional route into Raw Data) |
Notebook
Section titled “Notebook”A notebook is a per-project SageMaker Studio JupyterLab environment. Labelty injects your project API URL and key so you can use the SDK without pasting secrets into every cell.
Wallet and plans
Section titled “Wallet and plans”Many features (training GPUs, notebooks, workflow runs) debit a prepaid wallet and may require an active subscription / plan feature. If a call fails with a billing or entitlement error, top up the wallet or check your plan.
Next step
Section titled “Next step”Create and store a project API key: Authentication.