Uniopen's customized Nova model clears retail moderation targets

Uniopen customized Amazon Nova 2 Lite to exceed both of its production moderation thresholds, AWS reported. On 737 held-out conversation windows, fine-tuning and a subsequent prompt change raised behavior-classification Macro F1 from 0.5852 to 0.8550 and subject-classification Macro F1 from 0.4162 to 0.8491. The targets were 0.8500 and 0.8200, respectively; Macro F1 weights each category equally. The team trained on 3,391 conversation windows, then simplified outputs from JSON to a line-based format without further training. Clearing both gates allowed routine moderation to use the customized path. Human review remains mandatory for ambiguous cases and corrections reused as training data.