CABS + DataConcept-128M

Concept-Aware Batch Sampling (CVPR 2026, Bethge lab with Ludwig Schmidt and LAION co-authors): DataConcept-128M provides 128M concept-annotated image-text pairs, and CABS curates batches by concept on the fly during CLIP/SigLIP training — moving vision-language data curation from static filtering to training-time policy.