An episode is one task attempt: a time range of a recording with a task, an outcome, and optional subtasks. A dataset version is a frozen selection of episodes.
Where episodes come from
- The project workflow proposes episodes after processing. AI-proposed episodes are marked "AI-proposed, not reviewed" until someone saves or accepts them.
- On a recording's page you can draw a new episode on the timeline, edit an episode's task text, and see its revision history.
- If recordings are not segmented yet, the Episodes page offers Segment recordings.
Work with episodes
Open Episodes in the project sidebar. Filter by outcome, label state, task, duration, frame rate, camera count, idle share, subtasks, or a failing check. Select episodes and:
- Set the task or the outcome (success, failure, or partial) in bulk. Bulk actions are disabled above 50,000 matching episodes.
- Create a labeling job to hand them to annotators. See Annotation.
- Create a dataset version from the selection. See Dataset versions.
Which label fields a project uses (task, outcome, subtasks, relevance, attributes) is set in the project settings under Annotation schema.