nvidia/Nemotron-3.5-Content-Safety
The Nemotron 3.5 Content Safety model by NVIDIA is a 4 billion parameter small language model (SLM) based on Google's Gemma-3-4B-it, fine-tuned for multimodal, multilingual, and reasoning-oriented content safety. It functions as a content-safety moderator for LLM and VLM inputs and responses, supporting both standard taxonomy and custom safety policies with reasoning traces. This model unifies previous Nemotron 3 Content Safety Multimodal and Nemotron Content Safety Reasoning 4B capabilities, offering robust moderation across text and image inputs in multiple languages.
Loading preview...
Nemotron 3.5 Content Safety Model Overview
The Nemotron 3.5 Content Safety model, developed by NVIDIA, is a 4 billion parameter small language model (SLM) built upon Google's Gemma-3-4B-it. It is specifically fine-tuned for content safety moderation across multimodal (text and image) and multilingual inputs, incorporating advanced reasoning capabilities.
Key Capabilities
- Multimodal Moderation: Processes both text and images to assess safety, preserving the behavior of the Nemotron 3 Content Safety Multimodal model.
- Custom Policy Enforcement: Supports user-defined safety policies and domain-specific moderation criteria, a feature inspired by the Nemotron Content Safety Reasoning 4B model.
- Reasoning Traces: In custom policy mode, the model can produce a concise reasoning trace before delivering its final safety classification.
- Multilingual Support: Handles inputs in English, Arabic, German, Spanish, French, Hindi, Japanese, Thai, Dutch, Italian, Korean, and Chinese.
- Output: Provides safety labels ("safe" or "unsafe") for user inputs and responses, along with violated safety categories in standard taxonomy mode.
Good For
- LLM/VLM Content Moderation: Ideal for developers needing to moderate inputs and outputs of large language models and vision-language models.
- Custom Safety Requirements: Suitable for applications that require flexible, domain-specific safety definitions beyond standard taxonomies.
- Multilingual Applications: Effective for moderating content across a diverse range of languages.
- Commercial Use: The model is ready for commercial deployment, governed by the OpenMDW-1.1, Gemma Terms of Use, and Gemma Prohibited Use Policy.