How to share a robot dataset people can actually use

Your episodes upload automatically when you stop recording. Making them useful to strangers takes ten more minutes, and those ten minutes are the difference between a dataset that gets used and one that gets junk-flagged.

Step 1: lint before your final push

pip install datum-lint
datum-lint ~/.cache/huggingface/lerobot/YOURNAME/first-pick-place

It checks your metadata, labels, cameras, and episode files locally, with no network calls, and tells you exactly what to fix. Placeholder labels and mixed camera naming are the two most common catches. Fixing them before upload beats strangers finding them after.

Step 2: declare a license

On your dataset's Hub page, set the license field. Apache-2.0 or CC-BY-4.0 if you want your data used, including inside open models. In our latest scan, 172 of 1000 popular datasets skipped this and are legally unusable. Thirty seconds, do not skip it.

Step 3: write the card like a label

Three lines on the dataset page beat three paragraphs: what robot, what task, what cameras. People filter before they read.

Step 4: watch your scorecard appear

Datum re-checks the most downloaded public datasets every night. As your dataset gets found and downloaded, it enters the index and gets its own public scorecard showing every check and its evidence. That page is yours to link in your README, your posts, anywhere. Clean data with real labels and a real license stands out precisely because a third of popular uploads get flagged.

Why share at all?

Community data is real training material now: SmolVLA pretrained on 481 community datasets. Sharing puts your work inside the next generation of open robot models, with your name on it.