The uniopen platform team adapted the Amazon Nova 2 Lite model to its own moderation rules

According to an article by AWS, the uniopen platform team adapted the Amazon Nova 2 Lite model to its own moderation rules using supervised fine-tuning with the Low-Rank Adaptation (LoRA) method in the Amazon SageMaker AI service, followed by output optimization via prompting. Moderation of customer interactions distinguishes nine behavior categories and three subject groups: brand, other, and forbidden. For a correct decision, the model must correctly determine both components. The same rules apply to web, tablets, and mobile devices.

The training set contained 3 391 segments of customer conversations. The team evaluated the base model, the adapted model, and the adapted model with an optimized prompt on the same separate test set of 737 segments. According to AWS, the base model, before adjustments, achieved a Per Behavior Macro F1 of 0.5852 and a Subject Type Macro F1 of 0.4162, which was not sufficient for production use under the uniopen platform's rules.

The Amazon Nova 2 Pro model proposes fixes for reported errors, but each fix must be verified by a human before being included in the training data. Mandatory regression tests halt deployment in the event of failure; warning signals, such as low confidence or degradation in a particular category, require administrator approval. A candidate without these warnings can proceed automatically after passing the tests. Availability of the models varies by AWS service region. See the source article for details.