Export makes a downloadable pack from one dataset version. You start an export from the version page, not from a separate screen. A version can have several exports, and each one keeps its own format and options.
Who can do this: Owner, Editor (project roles).
Before you start
- A dataset version exists in the project.
Open Export on the version
Open the version from Versions, then click Export at the top right.
Choose the format
Pick a Format. Robot projects offer every format. Human and video projects do not offer RLDS or robomimic HDF5. For RLDS and robomimic HDF5, a hint under the field says what to do first.
| Format | Use it for |
|---|---|
| LeRobot v3 | For LeRobot v3. This is the default. |
| LeRobot v2.1 | For LeRobot v2.1. |
| Original files | The original uploaded files. |
| MP4 and JSON | MP4 video plus JSON metadata. |
| RLDS | For OpenVLA and Octo. Register the dataset in their loader config first. |
| robomimic HDF5 | For robomimic. |
Set what the pack contains
Set Cut by, Splits, Include labels, Video codec and Task language. LeRobot v3, LeRobot v2.1 and MP4 and JSON exports include the camera videos. For robomimic HDF5 you can also turn on Include camera images. Images make the file much larger, and the estimate shows the size.
| Field | What to enter | Example |
|---|---|---|
| ① Cut by | Episode writes one entry per episode. Subtask writes one per subtask. | Episode |
| ② Splits | The splits to include: Train share, Validation share and Test share. A split with no episodes cannot be chosen. | Train share |
| ③ Include labels | Adds the labels of each episode to the pack. | On |
| ④ Video codec | Keep original, H.264 or AV1 for camera videos. | H.264 |
| ⑤ Task language | The language of the task text in the pack: English or Chinese. | English |
Check what the export contains
The top of the dialog, under What this export contains, shows the episodes and recordings in the pack, any AI-proposed episodes that no one has reviewed, and any privacy findings. Read it before you start. A recording with privacy findings is listed there and never shows as passed.
- No consent document is recorded yet: add a consent file on the recordings page.
- Anonymization has not passed: open the recording, run the anonymize step and finish the privacy checks.
- Recordings with audio that has no approved redaction yet.
Start the export
Click Start. The export appears under Exports on the version page, with its status.
Download the pack
When the status is Succeeded, click Download on the export row to get the pack. The Use this version section on the card shows commands that load the same version with the Hugging Face tools. The command-line tool is described in the CLI reference.
Export status
| Status | Meaning |
|---|---|
| Queued | The export is waiting to start. |
| Running | The export is being built. |
| Succeeded | The pack is ready. Download appears, unless the load check failed. |
| Failed | The export did not finish. Open Details on the row to see why. |
Load check
A load check opens the pack with its loader. Only Passed is a pass. The other results are not passes.
| Result | Meaning |
|---|---|
| Passed | The loader opened the pack. |
| Structure checked, loader not run | The files are in place, but the loader did not run. This is not a pass. |
| Not evaluated | The check has not run yet. |
| Not applicable | The check does not apply to this format. |
| Failed | The pack did not load. Download is hidden. |
Troubleshooting
My format is missing
Human and video projects do not offer RLDS or robomimic HDF5. Use a robot project for those formats.
Start is disabled
Start stays disabled when no split is selected, or when the preview refuses the export for privacy. Read the preview to see what is missing.
The export is larger than expected
Camera images make a robomimic HDF5 export much larger. The estimate shows the size. If the export may go past the storage cap, the dialog says so.