How uniopen customized Amazon Nova to their retail moderation policies for production deployment
aws ·
Uniopen, a digital platform by Taiwan’s Uni-President Enterprises Group, customized Amazon Nova 2 Lite to meet specific retail moderation policies involving nine behavior categories and three subject types. The team utilized supervised fine-tuning with Low-Rank Adaptation in Amazon SageMaker AI to bridge the gap between general-purpose model capabilities and their unique business taxonomy. This customization significantly improved performance, raising the Per Behavior Macro F1 score from 0.5852 to 0.8364 and the Subject Type Macro F1 from 0.4162 to 0.8302. Following fine-tuning, the team applied prompt-level optimization to simplify output formats, which further increased both metrics above their respective production targets of 0.8500 and 0.8200. The deployment architecture separates primary moderation requests handled by Amazon Nova 2 Lite from correction workflows supported by Amazon Nova 2 Pro. Human reviewers must verify all candidate corrections before they enter the training set, ensuring that generated labels are not treated as ground truth. Automated hard and soft gates within an Argo Workflows pipeline prevent promotion if quality declines or warnings are detected. This approach allows routine moderation to be automated while reserving human review for ambiguous cases, maintaining responsible AI controls throughout the process.
aws