SamsungSDS-Research/SGuard-ContentFilter-2B-v1
SGuard-ContentFilter-2B-v1 is a 2 billion parameter safety guardrail model developed by Samsung SDS, based on the IBM Granite 3.3 2B architecture, with a context length of 32768 tokens. It is specifically designed to detect harmful content and screen adversarial prompts in LLM interactions, providing multi-class safety predictions across five categories: Crime, Manipulation, Privacy, Sexual, and Violence. This model excels at identifying safety risks in both prompts and responses according to the MLCommons hazard taxonomy, offering adjustable thresholds for category-specific safety level control.
Loading preview...
Overview
SGuard-ContentFilter-2B-v1 is a lightweight, specialized safety guardrail model developed by the AI Research Team at Samsung SDS. Based on the IBM Granite 3.3 2B model, it is designed to detect harmful content and screen adversarial prompts in human-AI conversational settings. The model provides multi-class safety predictions and binary confidence scores, enhancing interpretability for downstream use.
Key Capabilities
- Harmful Content Detection: Identifies safety risks in LLM prompts and responses based on the MLCommons hazard taxonomy, covering categories like Crime, Manipulation, Privacy, Sexual, and Violence.
- Adjustable Sensitivity: Allows for category-specific safety level control through adjustable thresholds, enabling fine-tuning of detection sensitivity.
- Lightweight and Efficient: Maintains a small model size (2 billion parameters) while offering robust safety detection capabilities.
- Multilingual Support: Primarily fine-tuned on Korean and English data, while retaining some capability in other languages supported by the base Granite model.
- Integration: Can be used with any open-source or closed-source LLM.
Performance
SGuard-ContentFilter-2B-v1 demonstrates strong performance on various content safety benchmarks, achieving an average F1 score of 0.83 across Beavertails, HarmfulQA, OpenAI Moderation, ToxicChat, and XSTest. It outperforms Llama-Guard-4-12B and Kanana-Safeguard-8B on several of these benchmarks, and shows competitive results against Qwen3Guard-Gen-4B.
Limitations
- Does not guarantee 100% accuracy, especially with novel attack techniques or data near decision boundaries.
- Performance is not guaranteed for languages other than Korean and English.
- Not a general-purpose LLM; it does not continue conversations based on prior history.
- Recommended input length up to 8k tokens for reliable detection.