AI Profile
summary
Defines an SPDX AI profile that extends a software bill of materials to record model properties, training process and provenance, and to enable automated compliance and vulnerability checks. Captured fields include energy consumption, hyperparameters, metrics and decision thresholds, safety risk and explainability, plus dataset lineage and noise. Worked examples include Llama 2 metadata and an LF Open Dataly collection showing 37,000 datasets described. Remaining gaps are heavy textual fields, no integrated SBOM workflow and limited tooling for automatic analysis.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:30AI software bill of materials need
- 2:55Limitations of AI docs standards
- 4:43AI profile structure core
- 6:34AI profile properties and relationships
- 7:58Why model cards fall short
- 9:01AI profile properties and fields
- 10:48AI profile example: Llama 2
- 11:53AI profile challenges and next steps
- 12:54Data set profile intro
- 14:39Why data sheets are insufficient
- 16:03Data set profile properties
- 17:07Data set profile in open projects
- 17:50Data set profile challenges
related talks
23mAI Factsheets as Industry Standard
47mUsing AI To Bridge the Gap Between Safety Standards and Open Source Development
40mAI-Powered! Or Is It Just Hype?!
33mAI in the Security Stack
17mBuilding Trust in the Age of AI: From Data Openness to Responsible Innovation
16mNavigating the AI Jungle
15mAI's Impact on Development, Post-Production & Support Journeys
15mCyber Risk Assessment, IT Asset Management and… AI obviously