MLCommons Candela: AI risk benchmarking-as-a-service for developers, deployers (mlcommons.org)

MLCommons has introduced Candela, a benchmarking-as-a-service platform focused on AI risk and reliability, addressing the pressing need for developers and deployers to understand and measure AI model behavior and safety. With participation from over 125 organizations, including major tech companies and academic institutions, Candela offers consensus-built AILuminate benchmarks that assess an AI system's responses across various hazard categories. This initiative is significant as it enhances transparency and trust in AI deployments, enabling businesses to evaluate normal risks, such as unsafe customer interactions or data leaks, with precision. Candela employs a private and stable scoring system through custom classifiers trained for safety evaluations, rather than relying on varying LLM outputs. This method uses item response theory for standardized scoring across different models and configurations, ensuring consistent comparison over time. By providing three types of results—practice tests, annotated results, and official scores—Candela facilitates various internal assessments, helping organizations make informed safety decisions as they tailor AI solutions. As AI usage rapidly expands, the introduction of standardized evaluations will be crucial for fostering responsible AI governance and meeting future regulatory requirements.