Nemotron 3 Content Safety 4B: Multimodal, Multilingual Content Moderation

Earlier safety models, which were text-only and mainly trained on English data, struggled with non-English and multilingual prompts, often missing cultural nuances. To address this, NVIDIA created the multimodal, multilingual Nemotron 3 Content Safety model. It was trained using the novel, culturally aligned multilingual safety data from the Nemotron Safety Guard Dataset v3. A multilingual safety model trained on this data has demonstrated superior performance on multilingual benchmarks.

The complexity of multimodal input—such as text paired with an image—presents significant challenges for safety models, as the meaning is often non-additive. For instance, an image of a benign household object (like a common kitchen knife) paired with the text "this is a great tool for cooking" is safe, but the same image paired with the text "I'm going to use this to harm someone" becomes a clear policy violation requiring immediate moderation.

Multimodal-multilingual content safety is challenging because it requires understanding cultural and linguistic context, especially in multimodal AI. A safety model must not only process multiple languages but also recognize how language and cultural context can alter the safety status of a prompt-image pair. For example, a prompt containing an image of a traditional religious symbol such as the Swastika paired with text describing a celebration might be perfectly acceptable in one language and culture (such as Indian), but when paired with an identical image and text in a different language (such as German) that carries a history of inter-group conflict, the combination could be interpreted as incitement to hate speech or discrimination, requiring immediate moderation. This sensitivity to cultural nuance is critical for accurate, globally deployed content safety models.

Nemotron 3 Content Safety is built on the Gemma‑3 4B‑IT vision‑language foundation model, which provides strong multimodal reasoning, instruction following, a 128K context window, and support for over 140 languages. NVIDIA fine‑tuned this base using a LoRA adapter, adding targeted safety classification behavior while keeping the model lightweight and efficient.

When a user provides text, an image, or both, the model encodes the visual and language features jointly and outputs a concise safety judgment. If an assistant response is included, the model evaluates the combined interaction to determine whether the response is safe in context, enabling it to catch violations that arise only from the interplay between request, image, and output.

It supports two inference modes: default low-latency safe/unsafe classification and safety category classification following the Aegis AI taxonomy. The training data blend consists of multilingual content safety data, human-annotated multimodal data, and synthetic data (SDG) which makes up about 10% of the total. Across benchmarks like Polyguard and VLGuard, the model achieved 84% accuracy on average, outperforming comparable models while maintaining roughly half the latency.