NVIDIANemotron 3.5
Nemotron 3.5 Content Safety
This 4B safety classifier is adapted from Google's Gemma 3 4B instruction model. It combines multimodal safety assessment with custom-policy reasoning.
What the release provides
NVIDIA trained a LoRA adapter and merged it into the released checkpoint. The card lists a 128K context and twelve supported languages. Its training combines general safety and domain-policy examples.Nemotron 3.5 Content Safety official model card
How to use it well
Practical guidance. Write the policy in terms that a reviewer can apply to the same input. Inspect false positives and missed cases separately for each language. Keep policy decisions auditable rather than treating the model's explanation as a substitute for the policy.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | nvidia/Nemotron-3.5-Content-Safety |
| Release license | OpenMDW model terms; review the linked Gemma base termsNemotron 3.5 Content Safety official model card |