Amazon Nova's Selective Forgetting: rDPO Reduces Over-Refusal
Decision Brief
Amazon Nova's Customizable Content Moderation Settings (CCMS) now include Reverse Direct Preference Optimization (rDPO), a new selective forgetting technique. rDPO enables the model to 'forget' specific content types while retaining normal processing for others. By reversing preference optimization, it reduces over-refusal—instances where content is incorrectly blocked—without compromising quality. For developers and enterprises, rDPO offers flexible content governance. Users can define custom moderation rules for scenarios like social media or e-commerce, precisely filtering unwanted content while minimizing false positives. AWS provides guidance on applying these techniques for similar results in user experiments.
Sources
- AWS:Machine Learning Blog
Applied ML, infra, and deployment guidance useful for AI builders on AWS.
- AWS:Machine Learning Blog