format v3: flat sample layout (top-level answer/question_type/pass-throughs), task_type v3 vocabulary, judge info consolidated in metadata score block. See dataset-upload skill references/mmeval-format.md and field-census-2026-07-08.md.

Upload caption_matching/data.json with huggingface_hub