nvidia/Nemotron-3.5-Content-Safety
The Nemotron 3.5 Content Safety model by NVIDIA is a 4.3 billion parameter small language model (SLM) built 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 safety taxonomy and custom policy enforcement with a context length of 32768 tokens. This model unifies multimodal moderation with custom policy adaptation, providing safety labels and optional reasoning traces for user inputs and model responses. It is designed for commercial use in applications requiring robust content moderation.
Loading preview...
Nemotron 3.5 Content Safety Model Overview
The Nemotron 3.5 Content Safety model, developed by NVIDIA, is a 4.3 billion parameter small language model (SLM) based on Google's Gemma-3-4B-it. It is specifically fine-tuned for content safety across multimodal, multilingual, and reasoning-oriented datasets. This model unifies the multimodal moderation capabilities of the Nemotron 3 Content Safety model with the custom-policy enforcement features of the Nemotron Content Safety Reasoning 4B model.
Key Capabilities
- Multimodal Content Moderation: Acts as a safety moderator for both text and image inputs, as well as LLM/VLM responses.
- Custom Policy Enforcement: Supports user-defined safety policies, allowing developers to implement domain-specific moderation criteria.
- Reasoning Trace Generation: In custom policy mode, it can produce a concise reasoning trace before classification, explaining its safety determination.
- Multilingual Support: Processes content in English, Arabic, German, Spanish, French, Hindi, Japanese, Thai, Dutch, Italian, Korean, and Chinese.
- Comprehensive Safety Taxonomy: Utilizes the Aegis Content Safety Dataset V2 taxonomy for standard safety classification.
- High Context Length: Supports a context length of up to 32K tokens.
Ideal Use Cases
- LLM/VLM Safety Moderation: Determining the safety of user prompts and model-generated responses in large language and vision-language models.
- Application-Specific Safety: Implementing tailored content safety rules for applications that require custom moderation criteria.
- Commercial Deployments: Ready for commercial use in various content moderation scenarios.