State of Robot Data · Issue 0 · 2026-08-23
We checked the 1000 most downloaded LeRobot datasets on the Hugging Face Hub. 230 are junk. Here is what that means and how to check before you train.
230 of 1000 datasets (23%) fail the junk gate: no readable metadata, fewer than 5 episodes, or every task label a placeholder. These are the most downloaded datasets, the ones people actually pull. Downloads do not mean quality.
1. 172 datasets have no usable license. If you train a commercial model on them, that is your legal risk.
2. 147 have no readable info.json at all. Tools cannot even load them.
3. Placeholder labels like "task desc" are still being uploaded daily. They poison language conditioned training silently.
Before you train: filter with the Datum finder. Junk is hidden by default. Before you upload: run pip install datum-lint and fix what it flags. Both are free.