25% of 'commercial-OK' HF models list non-commercial training data

Measured on the Hugging Face top-20k models by downloads (2026-11-03): of 1,964 non-re-upload models with a commercial-use license (Apache/MIT/CC-BY/OpenRAIL/Llama...) that list datasets: in their card, 25% list a dataset whose own terms say non-commercial/research-only, 32% list only commercial-OK data, 40% list data with no/unclear license or an id that doesn't exist on the Hub (13.6% of all model->dataset links are dead ids). License tags alone show only 3.6%: the restrictions mostly live in card prose and gate prompts (extra_gated_prompt in cardData via /api/datasets/<id>), so scan those, not just license: tags. ImageNet is 3/4 of flags (mostly timm, whose GitHub README itself says to assume the dataset license applies to the weights); the #1 downloaded model all-MiniLM-L6-v2 (Apache-2.0) trains on 9.1M MS MARCO triplets, which Microsoft says are for non-commercial research only. Watch for boilerplate: Salesforce cards say 'for research purposes only' under an Apache-2.0 tag; count tag/text conflicts separately.