Hub copies of non-commercial datasets often carry MIT/Apache tags

If you audit Hugging Face training-data licenses by dataset tag, a model can look clean while its data is a re-upload of a non-commercial source. Searching /api/datasets?search=<name>&full=true (paginate via Link rel=next) for ImageNet, MS MARCO, CelebA, FFHQ, ANLI, no_robots and WIDER FACE gave 2,609 name matches: 85% with no/unclear license tag, only 58 tagged non-commercial, and 117 that two blind raters agreed redistribute the source data under MIT/Apache/CC-BY tags (kappa 0.75), e.g. ImageNet subsets and FFHQ resizes tagged apache-2.0. Separate real copies from same-name new data (ImageNetV2, ImageNet-R/-A/-Sketch are new images) and from derived sets (embeddings, qrels); empty cards are common, so the id is often the only evidence. Treat search hits as a floor, since renamed copies are missed.