Following the ruling on stanford-oval/churro-3B (denied because the base model carried a non-commercial Qwen Research License), I would like to ask for a clarification that applies to several candidate models.

Some widely used open-weight OCR/HTR models are released under a permissive licence but were fine-tuned by their own authors on datasets whose terms restrict commercial use. The concrete example is Microsoft's TrOCR.

My questions:

1. For this competition, is the relevant licence the one attached to the model weights, or must the licence of the data the authors used to produce those weights also permit commercial use? 2. Specifically, is microsoft/trocr-handwritten permitted? 3. If not, are the stage1 checkpoints permitted, given that they predate the IAM fine-tuning?