來源Hugging Face Blog•較早收集於 3m
Nemotron 3.5 Content Safety：可自定義的多模態企業級 AI 安全防護
💡NVIDIA 推出的全新企業級多模態安全防護欄，助您確保生產環境中的 AI 部署安全。
⚡ 30 秒速覽
有什麼變化
為多模態 AI 輸入提供可自定義的安全防護欄。
為什麼重要
此工具協助企業降低模型輸出有害或不當內容的風險，促進生成式 AI 在受監管產業中的安全採用。它為管理合規性與品牌安全提供了一種標準化方式。
下一步行動
評估您目前的安全性管線，並測試 Nemotron 3.5，看看它是否能取代或增強您現有的內容審核層。
誰應關注：Enterprise & Security Teams
關鍵要點
- •為多模態 AI 輸入提供可自定義的安全防護欄。
- •專為企業級部署需求而設計。
- •支援強大的內容過濾功能，以降低生產環境中的風險。
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 7 個來源。
🔑 增強重點摘要
- •Nemotron 3.5 Content Safety is a small language model (SLM) built upon Google's Gemma-3-4B-it, which NVIDIA fine-tuned using multimodal and multilingual datasets specifically for content safety.
- •The model is designed to act as a content-safety moderator for both the inputs (prompts and optional images) and the generated responses from Large Language Models (LLMs) and Vision Language Models (VLMs).
- •It supports 23 distinct safety categories and 12 languages, providing enterprises with customizable policy enforcement and 'reasoning trails' to facilitate auditing and adaptation of safety decisions to specific domain rules.
- •Beyond real-time inference-time guardrailing, Nemotron 3.5 Content Safety can also serve as a judge for evaluating and testing LLM safety, or its accompanying training dataset can be used to post-train other models for improved safety behaviors.
- •Nemotron 3.5 Content Safety is integrated into the broader NVIDIA Nemotron family of open models, which are optimized for agentic AI applications, and is deployable as an NVIDIA NIM microservice.
🛠️ 技術深入
- Base Model: Google Gemma-3-4B-it.
- Network Architecture: Transformer (Decoder-only).
- Vision Encoder: SigLIP, designed to process square images resized to 896 x 896 pixels.
- Total Parameters: 4 Billion (4B).
- Fine-tuning Method: LoRA (Low-Rank Adaptation), with weights subsequently merged back into the main Gemma-3-4b-it model.
- Training Data Modality: Multilingual Text and Images.
- Training Data Size: Less than a million images and less than a billion tokens.
- Data Sources: NVIDIA ThreatOps Team, Nemotron Safety Guard v3, Nemotron VLM Dataset V2, and synthetically generated data.
- Data Collection and Labeling Method: Hybrid approach combining automated, human, and synthetic methods.
- Context Window: Supports up to 128,000 tokens.
- Output: Provides a string containing safety labels for both the input (prompt and image) and the response (if present), with optional lists of violated safety categories and a reasoning trace.
- Deployment: Optimized for execution on NVIDIA GPU-accelerated systems, leveraging NVIDIA's hardware and software frameworks like CUDA libraries for enhanced training and inference performance. It is available as an NVIDIA NIM microservice.
🔮 前景展望基於引用來源的 AI 分析
NVIDIA will further integrate Nemotron 3.5 Content Safety with its broader NeMo framework and NIM microservices to offer more comprehensive, end-to-end AI agent safety solutions.
Nemotron 3.5 is already part of the Nemotron family for agentic AI and integrates with NeMo Guardrails and NIM, indicating a strategic direction towards a unified safety stack for complex AI systems.
The 'free' availability of Nemotron 3.5 Content Safety will drive its adoption as a foundational safety layer, increasing NVIDIA's influence in the enterprise AI safety market.
Offering the model for free lowers the barrier to entry for developers and enterprises, encouraging its use and potentially standardizing NVIDIA's safety taxonomy and integration patterns within AI applications.
⏳ 時間線
2023-11
NVIDIA introduced Nemotron-3 8B, the first public Nemotron-branded release, for enterprise chatbot and copilot development.
2024-06
NVIDIA released the Nemotron-4 340B family, intended for synthetic data generation and instruction tuning, with over 98% of alignment data synthetically generated.
2025-01
NVIDIA announced a broader Llama Nemotron family at CES, intended for enterprise reasoning and agentic AI tasks.
2025-12
NVIDIA announced the Nemotron 3 family (Nano, Super, Ultra), emphasizing efficiency and leading accuracy for agentic AI applications, with Nano released.
2026-03
NVIDIA formed the Nemotron Coalition, a group of AI labs collaborating on future open models.
2026-06-04
NVIDIA introduced Nemotron 3.5 Content Safety, a customizable multimodal enterprise AI safety tool.
📎 來源 (7)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： Hugging Face Blog ↗
每週電子報
每週一封，可隨時退訂。