Amazon introduced Reverse Direct Preference Optimization (rDPO), a new technique for selective unlearning in AI models. This method powers Amazon Nova's Customizable Content Moderation Settings to reduce over-deflection while maintaining model quality. The blog also offers guidance for users to apply these techniques in their own AI experiments.
Source preview
Server-side extracted preview paragraphs from the original source.
In this post, we introduce Reverse Direct Preference Optimization (rDPO), the novel unlearning technique behind Amazon Nova Customizable Content Moderation Settings (CCMS), and show how it reduces over-deflection while preserving model quality.
Organizations deploying foundation models (FMs) often encounter a common challenge: model safeguards designed for content moderation can also prevent legitimate, business-critical use cases.