# Uniopen's customized Nova model clears retail moderation targets

_Published Thursday, October 1, 2026 at 11:56 AM EDT · AI · Latest · Tier 2 — Notable_

![Uniopen's customized Nova model clears retail moderation targets — Primary](https://d2908q01vomqb2.cloudfront.net/f1f836cb4ea6efb2a0b1b99f41ad8b103eff4b59/2026/09/14/ML-21411-featured-image.png)

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.

## Sources

- [Artificial Intelligence](https://aws.amazon.com/blogs/machine-learning/how-uniopen-customized-amazon-nova-to-their-retail-moderation-policies-for-production-deployment/)

---
Canonical: https://techandbusiness.org/newswire/10uwuU6mIzOA6bptwWSTxP
Published: 2026-10-01T15:56:29.809Z
Story chronology: 2026-10-01T15:33:01.000Z
Retrieved: 2026-10-01T18:17:55.845Z
Publisher: Tech & Business (techandbusiness.org)