Model cards

Standardized documentation: Model cards provide benchmarked evaluation and performance characteristics for trained machine learning models, ensuring transparency.

Model Cards are short, structured documents accompanying trained machine learning (ML) models, designed to increase transparency and support responsible AI. They function like a "nutrition label" for AI, detailing key facts: intended use, performance evaluation procedures, and known limitations. Crucially, a card provides benchmarked evaluation across diverse demographic and phenotypic groups (e.g., race, sex, geographic location) to highlight potential biases or differential performance. This standardized reporting (proposed in a 2018 paper) helps developers, policymakers, and end-users understand a model’s specific strengths and weaknesses, enabling safer, more informed deployment.

https://arxiv.org/abs/1810.03993