Who checks the quality of robot datasets?

Robot training data has marketplaces, collection companies, and hosting, but until recently nobody whose whole job is checking quality. That is what Datum is.

What does Datum actually do?

Every build, Datum scans the most downloaded LeRobot datasets on the Hugging Face Hub and runs six checks on each: metadata, format version, episode volume, task labels, camera streams, and license. Results are public with their evidence, dataset by dataset. Current scan: 1000 datasets checked, 230 flagged as junk, 582 passing every check.

Why does the referee have to be independent?

Everyone else in robot data has a conflict. Hosting platforms also sell robots and train models. Data companies grade their own product. Model labs compete with other labs. Datum's pledge is structural: no data collection, no model training, no robot sales, ever. Checking is the entire business, which is the only reason the checks can be trusted.

Can I disagree with a check?

Yes, and publicly. The rubric is open and versioned. Disputes go to dispute@datumrobotics.ai, outcomes are published, and when a dispute changes the rubric, every affected dataset is rechecked. A trust layer earns trust by being wrong in public and correcting fast.

What should you do with it?

Training: use the finder, junk is filtered out by default. Publishing: run datum-lint before you upload. Buying data: check the scorecard before you pay for anything.