Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 20·5 min read
Position: Current Model Cards Are Insufficient for Downstream Governance of Open-Weight Foundation Models
TL;DR
Current model cards for open-weight foundation models (OWFMs) are inadequate for ensuring safety and governance. A multi-layered approach integrating model cards, acceptable use policies, and licenses is essential for effective oversight.
✦ Why It Matters
Review and enhance your model governance strategies by integrating model cards, AUPs, and licenses today.
Key Takeaways
- Analyze existing model cards for safety gaps.
- Integrate acceptable use policies with model cards.
- Reassess the suitability of open-source licenses for OWFMs.
- Develop a comprehensive governance framework for AI models.
Full Summary
Open-weight foundation models (OWFMs) are increasingly prevalent, prompting a need for improved governance strategies. An analysis of 500 model cards from Hugging Face reveals significant safety gaps in existing frameworks, which often overlook critical aspects like model heritage and alignment provenance.
The position paper advocates for a comprehensive governance approach that combines model cards, acceptable use policies (AUPs), and licenses to address these shortcomings. It highlights that standard open-source licenses may undermine the enforceability of AUPs, further complicating governance.
By evolving these components into integrated safety artifacts, the paper aims to create a more robust framework that encompasses informational, normative, and legal dimensions, ultimately enhancing the safety and accountability of OWFMs.