Merged BF16 model from a text-layer-only QLoRA ablation of nvidia/Nemotron-3.5-Content-Safety for contextual healthcare safety classification in Portuguese. Trained on 1,777 safe Revalida questions and a category-stratified sample of 3,554 unsafe Medical Malice questions.

The model emits User Safety: safe or User Safety: unsafe; category prediction and thinking were disabled. The upstream OpenMDW and Gemma terms continue to apply.