How to tell if a robot dataset is good

A robot dataset can look fine on its Hub page and still waste your training run. 230 of the 1000 most downloaded LeRobot datasets (23%) fail basic checks. Here is what to look at before you download 40GB.

Does the metadata load?

Open meta/info.json first. If it is missing or incomplete, tools cannot load the dataset at all. In our latest scan, 147 of 1000 datasets had no readable metadata file. That is the fastest reject there is.

Are the task labels real sentences?

Task labels are the sentences that tell a language conditioned model what each episode shows. Placeholders like "task desc" teach the model that words mean nothing. The team behind SmolVLA had to re-annotate noisy community labels by hand before pretraining, which is documented in their paper. If the labels are placeholders, budget time to rewrite them or pick another dataset.

How many episodes does it have?

Fewer than 5 episodes is a test upload, not training data. For fine-tuning a single task, practitioners commonly record tens of episodes. More matters less than clean.

Are the cameras named consistently?

A wrist camera called wrist in some episodes and Wrist_Cam in others will break your data loading or silently confuse training. Consistent lowercase names with underscores are the safe convention.

Can you legally use it?

172 of 1000 checked datasets have no usable license. No license means you cannot safely build a product on it, no matter how good the data is.

The shortcut

Datum runs all six checks on every public LeRobot dataset, every build, and publishes the evidence. Use the finder to filter junk out, or run pip install datum-lint on your own dataset before you upload.