Versioning
A training run should always be attributable to an exact dataset state. Releases are immutable and identified by a version you can pin.
Version numbers
- MAJOR — the schema or the directory layout changed in a way that breaks an existing loader.
- MINOR — episodes were added. Existing episode IDs and annotations are unchanged.
- PATCH — annotations or metadata were corrected. No episodes added or removed.
Immutability
A published version never changes. If an annotation is wrong, the correction ships as a new patch version with a changelog entry — the original stays retrievable, so an old training run stays reproducible.
The one exception is withdrawal: if a contributor or site owner withdraws consent, the affected episodes are dropped from future versions and access to affected files in existing versions is revoked. That event is recorded in the changelog.
Pin a training run
Record the stable identifier next to your run. It resolves to one immutable file set, permanently.
dataset: curatrix:electrical-assembly@1.3.0
schema: 2.1
sha256: 9f2c1d0b4a... # from checksums.txt, verified at load timeChangelogs
Every release carries CHANGELOG.md: episodes added, annotations corrected, schema changes, and any withdrawal that removed material. It is written for the person deciding whether to re-run training.