R.O.A.D. Barbados Historic Handwriting Challenge
'/%3e%3c/svg%3e)
Hello, I have a quick question about the "pretrained models must be openly available to everyone" rule.
On behalf of my team on this very interesting task, I write to ask about the usability of Stanford OVAL's CHURRO-3B (https://huggingface.co/stanford-oval/churro-3B), a 3B-parameter VLM fine-tuned from Qwen2.5-VL-3B-Instruct for historical document/handwriting transcription.
The model is freely downloadable by anyone on Hugging Face — no application, no gatekeeping. However, its license field is listed as qwen-research, which is Alibaba's non-commercial, research-only license (this differs from the base Qwen2.5-VL-3B-Instruct model itself, which is Apache-2.0 — the restriction appears to come from licensing terms on some of the historical datasets CHURRO was fine-tuned on).
Given the rule states pretrained models are allowed "as long as they are openly available to everyone," could you clarify whether:
- "Openly available" refers to accessibility (anyone can download it, which CHURRO satisfies), or
- It also requires a license that permits use in a competition with cash prizes (which CHURRO's research-only license may not satisfy)?
We'd rather confirm this now than build further on it and find out later it isn't eligible. Thanks in advance!
MICADEE
Thanks for raising this, and for sharing the model and licence details.
After reviewing the licence, stanford-oval/churro-3B would unfortunately not be permitted for this challenge under its current Qwen Research License. The licence restricts use to non-commercial purposes, while solutions developed through this challenge may need to be used or deployed by the challenge partner beyond research or evaluation.
To avoid creating downstream licensing restrictions for the final solution, please use a model whose licence permits commercial use. If a separate commercial licence for this model is available and obtained, we’d be happy to review that separately.
Thanks again for checking before proceeding.