Datasets and Versions
A dataset is the product of curation and the only input to training. A version is the contract between the two: every export creates a new version number, and once frozen, a version can't be changed.
Where: Export in the data workbench header. Exported versions are listed under Data Curation → Dataset Management.
Export Formats
| Format | Description | Use with |
|---|---|---|
| HDF5 | One .hdf5 file per sample, with images, robot state, and actions aligned by timestamp | The platform's built-in training templates and in-house training frameworks |
| LeRobot | Organized to the LeRobot dataset specification, with Parquet tables and MP4 video | Popular open-source training frameworks |
See Data Formats for file layouts.
Exporting a Dataset
- In the data workbench, confirm the project is fully curated: labels confirmed, ranges defined, and every item reviewed.
- Click Export, enter a dataset name and notes, and choose HDF5 or LeRobot.
- Click Start Export. The platform generates the target data from the project's instructions:
- items marked for exclusion are skipped;
- time spans are trimmed according to keep and excluded ranges;
- an export manifest is written.
- When the export finishes, the dataset status changes to Exported and the version is frozen.
The export manifest always includes labels
The export manifest always records each sample's labels, notes, and ranges. There is no option to drop quality labels.
Export Blockers
An export is blocked in the following cases, and the page shows the specific reason:
| Blocker | Resolution |
|---|---|
| Pending stitching | Some cross-archive samples haven't been stitched yet; wait for stitching or resolve it manually |
| Pending confirmation | Some candidate labels or feedback samples haven't been confirmed |
| Failed quality check | Some items haven't been reviewed or didn't pass quality checks |
| Incomplete channels | Some items are missing channels; remove them from the project or re-upload them, then export |
| Pipeline running | Wait for the pipeline to finish |
Export Results
Export results are shown as "succeeded (failed)", for example 99 (1). The two numbers add up to the number of items included in the export. Click Details to view the log, which lists the data ID and reason for every failed sample.
Versioning Rules
| Rule | Description |
|---|---|
| Frozen on export | A frozen version can't be changed. Later changes to the project produce new versions only. |
| Version numbers | Incremented within the project; unique and never reused |
| Parent version | Versions created by merging or adding new data record a reference to their parent |
| Sample provenance | Records each sample's source (collection / feedback / pipeline), the time range covered, and the number of new samples |
| Training binding | Only Exported versions can be used for training; training jobs can't be created from a version that is still exporting |
Adding New Data
There are two ways to handle newly collected data:
- Export separately: Create a new curation project for the new data, export it as a new dataset, and train on it separately.
- Merge into a new version: Merge it with a curated dataset in the same format to create a new version. Samples are deduplicated by sample ID, and the parent version is recorded.
Format requirements for merging
Merging requires the same format and compatible modalities. Converting between HDF5 and LeRobot must be validated separately; formats are never mixed silently.
Dataset States
| State | Meaning |
|---|---|
| Pending Export | Submitted and waiting in the queue |
| Exporting | Files are being generated |
| Exported | Complete and frozen; available for training |
| Export Failed | The export failed. Check the log and export again. |
Dataset Preview
Click Preview in Dataset Management to see the sample list, each sample's video and robot-state curves, and the labels and ranges in the export manifest.
Training While Collecting
You can feed data in batches: after each batch is collected, curate it, export a new version, and create a training job from that version. Every training job is bound to one fixed version.

