Datasets and exports - Capibo

Each mission gets one dataset as soon as its first video is accepted. The dataset grows as more videos pass review. Its version increases when the mission’s tasks change and becomes v2.0 when the mission completes. A dataset only contains accepted videos. To audit what was rejected, list videos with status=rejected.

Two ways to get the data

List and download

Page through GET /datasets/{id}/videos and download each video from its signed download_url. Best for incremental syncs and custom pipelines.

Export

POST /datasets/{id}/exports packages everything in a training format and delivers it to your storage. Best for full snapshots.

Export formats

`format`	Output
`lerobot`	A Hugging Face LeRobot dataset: videos, episode metadata and task labels.
`rlds`	RLDS episodes as TFRecord shards, for TensorFlow and Open X-Embodiment tooling.
`webdataset`	`.tar` shards that pair each MP4 with its JSON annotations, for streaming data loaders.
`mp4-json`	Plain MP4 files plus one JSON annotation file per video.

Destinations

`destination`	Delivered to
`s3`	The S3 bucket connected in Settings → Integrations.
`gcs`	The connected Google Cloud Storage bucket.
`hf`	A private repository in your connected Hugging Face organization.
`link`	A signed download link, valid for 24 hours.

Exports blur any faces or screens detected in frame by default (blur_faces: true). Pass created_after to export only videos accepted since your last export.