v0.13.1

Added

• Joint overlay datasets: Add backend-agnostic JointDataset support for overlay training across independently sampled primary and secondary datasets, plus dedicated joint samplers and loader validation so tuple-based pairing is only used with joint datasets.

Changed

• Dataset documentation and coverage: Expand dataset-layer docstrings, clarify aligned (MixedDataset) versus unaligned (JointDataset) composition semantics, and add focused tests for joint dataset construction, pairing, and loader/sampler validation.

Fixed

• On-demand driver usability: Restore the ability to omit both iterations and epochs when using the driver and I/O manager in on-demand mode, while still rejecting the ambiguous case where both are specified.

Full Changelog: v0.13.0...v0.13.1