Track 1 is focused on improving the multimodal Apertus v1.5 8B and 70B models. There are two different paths. Path A focuses on red-teaming Apertus. Submissions can focus either on the 8B or 70B model.

The submission must comply with the HackApertus T&C.

🎯 The Challenge

This challenge accepts submissions that find and document where Apertus underperforms, behaves unexpectedly, or produces outputs that do not meet the expectations of a sovereign, multilingual, trustworthy model. Participants can submit up to 5 issues. Issues can be submitted under any topic, in any modality (text, image and speech). The following topics are of particular interest in this hackathon:

• Misrepresentation of culture and values
• Stereotyping and bias
• Disclosure of private and personal information
• Generation of IP and copyrighted materials
• Misrepresentation of history and factual incorrectness

🔧 Resources, Tools & Support

Check our resources & tools page for detailed information: https://hackapertus.notion.site/resources-tools

📑 You can learn more about AI Red Teaming in this playbook and course from Humane Intelligence.

💻 Technical Preferences

📦 Submission Requirements & Deliverables

❗️ Submissions are not handled on Devpost but via this URL only:

http://hackapertus.ch/online-hack/submissions

Github repo template: https://github.com/HackApertus/project-template

The submission must: