How the platform works

Toloka Platform has two product workflows: Data Labeling produces labeled data, and Train produces a Model Version you can download. They share projects, datasets, team access, and a team balance.

​

Choose your starting point

You do not need to train a model to label data. You do not need a labeling pipeline to train from an existing dataset.

​

Shared resources

A team owns projects and their balance. Access follows your team membership. Use Teams, access, and balance to bring in a colleague or add funds. A project holds the datasets and work for a task. Its pipelines belong to Data Labeling; its Trainings, Training Runs, Models, and Model Versions belong to Train. See Projects. A dataset holds items and fields. Data Labeling reads and writes those fields. Train reads selected rows as conversation examples. The same project can use both workflows, but a labeling export is not automatically a training file: prepare the labeled results in the required conversation format.

​

Data Labeling: describe, review, run, export

Describe your task to the project agent. It builds a pipeline, a connected set of nodes that labels, transforms, or routes your items. You review the configuration, resolve its checks, and approve the cost before deployment. A pipeline run processes the selected items. Labeling nodes use experts or, where supported, an LLM. Quality-control settings determine how labels are reviewed. You can inspect results and export them from the run. A deployed pipeline is locked. To process another batch, clone it and deploy the copy. Read How data labeling works for the division of work between you, the agent, and the pipeline.

​

Train: configure, estimate, run, download

A Training is an editable draft: choose a method, a model source, dataset selections, and settings. A free cost estimate validates the inputs and gives you a maximum charge. A Training Run is one execution of that draft. Starting it places a hold for the maximum charge on your team’s balance. A completed Training Run publishes a Model Version inside a Model. You download the files and serve them in your own environment; the platform does not host inference. Use the CLI or HTTP API for project Training. The first-model tutorial walks through a CLI run. How training works explains its lifecycle and states.

​

Costs are product-specific

Both products use the team balance, but their charge rules differ. Data Labeling has a cost forecast and a frozen per-run contract. Train estimates a maximum charge and settles GPU usage against a hold. Read Billing, then the relevant product’s pipeline pricing or Training cost before launching paid work.

​

Work from code

The CLI covers Data Labeling and Train. The TypeScript SDK authors and tests labeling pipelines. The HTTP API exposes both product workflows. When you need both products, follow Label, train, and evaluate. For an unfamiliar term, use the glossary.