Six checks, all explained, all disputable. This build checks metadata. Video decode and per episode signal checks come next, and the rubric version will bump when they do.
MetadataIs meta/info.json present, readable, and complete? Without it, tools cannot load the dataset.
Format versionIs it standard LeRobot v2.0, v2.1, or v3.0? Custom formats break the shared tooling.
Episode volumeFewer than 5 episodes is a test upload, not training data. It gets a junk flag.
Task labelsAre the task strings real descriptions, or placeholders like "task desc"? Placeholder labels poison language conditioned training. v3.0 label parsing is coming.
Camera streamsDoes the schema declare at least one RGB stream, and what are the streams named?
LicenseIs there a real license, and does it allow commercial use? "None" and "other" both fail.
Disputes
Think a check got your dataset wrong? Email dispute@datumrobotics.ai with the dataset id and the check. Disputes and their outcomes will be public. The rubric is versioned, and when a dispute changes it, every affected dataset gets rechecked.
The neutrality pledge
Datum never collects data, never trains models, and never sells robots. We only check. That is the whole point: every other player in robot data is conflicted on at least one of those.
For pipelines and agents
All facet data is public JSON at data/facets.json. Point your training pipeline or agent at it. An API with filters is coming.