How datasets and episodes work

An episode represents a unit of data for training, often a single example or task. A dataset collects episodes into versioned and immutable groups and provides tooling to review data.

Episode model

An episode references one or more recordings and a time range. This allows for curation without duplicating the underlying data.

A recording can appear in any number of episodes. Deleting a recording leaves the episodes and their history intact, but its data is no longer available for visualization or download.

Each episode contains every topic for the set of recordings. A subset of topics can be selected when training or downloading to avoid reading extra data.

Episode versions

Changing an episode's time range, recordings, or properties creates a new version. Episode versions are immutable and record who made the change.

If a change gives the episode a different meaning, save it as a new episode instead.

Properties

Custom properties for episodes can track organization-specific info. It's common to use properties to track pipeline status and to link annotation files kept in your own storage. See Label with an external tool.

Organization admins define properties, each with a key, display label, and value type.

Filter by properties in the app, in Search, or through the API. Training code can filter on properties.

Drafts and dataset versions

When episodes are added, removed or modified in a dataset, they exist in a draft state for the next dataset version. Review and commit the version to create an immutable record of these changes. Further updates to episodes will not impact existing committed dataset versions.

Committing a new dataset version for each iteration of your model allows for easily reproducible training data. See Manage datasets for committing, comparing, and restoring versions.