Current public v3 datasets, with native image geometry, separate annotation tasks and documented provenance.