Readable reference
Give teams a memorable owner/dataset reference for discovery. Its alias may advance while pinned revisions remain fixed.
Dataset registry · product preview
Give every training dataset a permanent address.
Avala's managed operations deliver governed Physical AI datasets today. The registry preview shows how each release will bind a readable reference to an immutable revision, manifest, origins, and declared rights.
- Available now
- Managed dataset operations
- In development
- Self-serve registry + public resolver
Readable for teams. Digest-pinned for machines. Reproducible after handoff.
MCAP + ROS
LiDAR + point clouds
Multi-camera video
Telemetry + trajectories
More than dataset management
Turn a folder into an addressable training artifact.
The registry contract in development separates the name teams use for discovery from the immutable identity a training run records. Each release is designed to bind what it contains, where it came from, and the terms declared for its use.
Canonical identity
Dataset and organization names may be renamed; the UID-and-digest canonical reference remains the machine identity.
Content manifest
Seal logical paths, roles, byte sizes, source origins, and SHA-256 hashes into the revision manifest.
Rights declaration
Snapshot declared terms and attestation hashes with the revision. Avala does not infer legal rights from a filename or source.
Revision lineage
Keep mutable aliases separate from immutable inputs.
The planned registry follows explicit parent relationships across revisions while aliases stay convenient for discovery and canonical digests preserve the exact training input.
- Explicit parent relationships connect each revision to the release it followed.
- A readable alias such as main may advance; pinned revision identities do not.
- A revision's content identity cannot be rewritten after creation.
Release evidence
Know exactly what the model can resolve.
The preview release record shows the contract in development: a sealed manifest, origin identities, declared rights, and an immutable revision digest.
- Record exact object count, total bytes, logical paths, roles, and SHA-256 hashes.
- Expose safe public origin identities while private locators and credentials stay excluded.
- Pin declared open-download terms and rights-attestation hashes to the revision.
- Availability checks fail closed before an eligible public revision can resolve.
Dataset release record
Preview- Dataset
- avala/warehouse-pick
- Revision digest
- decafbad…
- Manifest digest
- 55b9c1e2…
- Objects
- 83,440
- Total size
- 1.8 TB
- Origins
- 2 declared
- Rights
- Open download · declared
- Lifecycle
- Published
avala://datasets/6eaf2db0-466f-4b41-9a5e-3f9e6f0531d7@decafbaddecafbaddecafbaddecafbaddecafbaddecafbaddecafbaddecafbadPublic resolver preview
Resolve once. Pin what training consumed.
Eligible public revisions with declared open-download rights can load without an API key when public resolver access is enabled.
When resolution is enabled, the Python SDK fetches objects lazily and verifies declared size and SHA-256 before returning each object. Persist the canonical reference—not a movable alias—in training config.
See the Python SDKStart with one dataset release
Define the dataset identity your model pipeline can depend on.
Bring a representative sensor dataset, storage boundary, and current handoffs. We'll map a managed delivery available today and the digest-pinned release contract for your model pipeline.