Skip to main content
A project is the workspace shared by Data Labeling and Train. It belongs to one team and holds the resources for one body of work: datasets, pipelines and their runs, project files and secrets, Training drafts and Training Runs, and Models with their Model Versions. The same project can connect the products. You can label examples in a project dataset, prepare those rows for Training, and train from that dataset without creating a separate account or balance.

​
Ownership and access

Every project belongs to one team. The team’s access level applies to the project and its resources:
  • Full access members can create and change project resources.
  • Read only members can open projects, datasets, pipelines, runs, and results but cannot change them.
All projects in the team use the team’s single balance. See Billing for the different ways pipeline runs and Training Runs use it.

​
What a project contains

​
Create or select a project

Choose the surface that matches your work:
  • Web app for Data Labeling. From the team’s project list, click + Project, describe the labeling task, and attach data if needed. The project opens with the agent building a pipeline from that description.
  • CLI for either product. Run toloka project create <alias> --team-id <team-id> in a new checkout. Use toloka project add <alias> <project-id> to select an existing project instead.
  • HTTP API for either product. Create or list projects through the beta API, then use the project ID with Data Labeling or Train resources.
The CLI records one selected project in toloka.json for a checkout. A second project needs a separate checkout or an explicit project ID with its state files kept separate. See CLI project selection for the precedence rules.

​
Empty draft projects

Choosing a product or template on the home screen creates a draft project. The platform automatically deletes a draft project that owns nothing once it is more than a day old, measured from when the draft was created. A draft owns nothing when it has no dataset, pipeline, secret, Training, Model, or other project resource. Deletion also removes the draft’s agent chat thread and any project files uploaded to it. If you return to an old browser tab for an empty draft, the project is gone. Start a new draft from the home screen instead. Drafts that own a dataset or any other resource are kept.

​
Use a project for Data Labeling

A Data Labeling project combines the agent chat, datasets, and pipelines. Start with Get your first labeled dataset to create a project in the web app and run its first pipeline. For source-controlled authoring, the TypeScript SDK defines pipeline source and the CLI imports, checks, plans, and applies it.

​
Use a project for Train

Train uses the same project and dataset resources. A Training draft names a model source, a method, a frozen slice of a project dataset with its split column, and settings. A billed Training Run copies that configuration and publishes a Model Version when it completes. Start with Train your first model. If your examples come from labeling results, follow Prepare training data from labels.