Our team has started developing ML pipelines for the competition and would appreciate clarification on a couple of rule interpretations:
1. Pre-trained models and “external data”: Are pre-trained models (e.g., ImageNet-trained CNNs, CLIP, pre-trained transformers) considered “external data” under the competition rules? These are standard practices in modern ML, but are technically trained on external datasets not provided by the competition.
2. Open source model restrictions: Are there any restrictions on using large open source multimodal models (like Llama, Qwen, etc.)? While these meet the open source license requirement, they could require substantial compute resources and might create advantages based on hardware access rather than methodology.
Would appreciate any guidance on these points to ensure we’re all interpreting the rules consistently!
Per the competition rules, external data is not allowed in this competition. However, participants can use pre-trained computer vision models and language models as long as they were 1) available freely and openly in that form at the start of the competition and 2) were trained in such a way as to be compliant with the data use agreement for the challenge. Please ensure when leveraging these models that you are also complying with the data use agreement for this challenge.
The rules also state that “any software required to generate the solution must also be available under an open source license that does not prohibit free commercial use”. The Llama license appears to be open source and commercially unrestricted with the exception of extremely large organizations and is permissible for this competition, but certain Qwen models appear to restrict commercial use more significantly. All licenses listed as OSI-approved open source licenses are considered free and open source for this competition - if you are unclear whether the license for a specific model you wish to use is free and open source, please follow up on those licenses specifically.