Datasets:
tasksource/flickr30k-grounding v2026.10.16
- Add 100,000 human-labeled training phrase-to-box decisions, 11,232 validation questions and 11,560 test questions, grouped by unchanged Flickr30k image.
- Preserve native image partitions; require one native target box and distinct annotated-region choices. Exclude the publisher’s known unrelated captions and multi-box/no-box targets.
- Record original noncommercial research/education and Flickr image terms, source revisions and conversion hashes.