Thanks to everyone who has raised questions around external data and pretrained models. For this challenge, we want to keep the final solutions reproducible and ensure that winning solutions can be used by the challenge host without downstream licensing restrictions. The following rules therefore apply:

• Competition data only may be used for training or fine-tuning. Additional external datasets may not be used to train, fine-tune, pseudo-label or otherwise adapt your model.
• Publicly available pretrained base models are allowed, provided their licence permits the challenge host to use, modify, reproduce and deploy the resulting solution, including for commercial purposes. Models or weights restricted to research-only, academic-only or non-commercial use are therefore not permitted.
• Pretrained model weights are permitted where the weights themselves are publicly available under a licence that allows this downstream use. Participants are not required to independently verify the licences of every dataset used upstream by the original model author, unless the model licence or documentation explicitly carries those restrictions through to downstream use.
• All model weights and inference code used to produce the final submission must be provided as part of the final solution and must be reproducible.

In short: you may start from an approved/open pretrained model, but its licence must allow the challenge host to use the resulting solution without non-commercial or research-only restrictions, and any further training or adaptation must use only the data provided as part of this competition.

We’re adding this clarification to make the rules as clear and consistent as possible for everyone. All the best!