A 20-billion-parameter safety classifier from OpenAI designed to detect harmful content and enforce usage policies in open-source deployments.