mncai/genguard-1.0-light

GenGuard 1.0 Light: Integrated AI Safety Guardrail

GenGuard 1.0 Light, developed by mncai, is a 7.9 billion parameter guardrail model built on google/gemma-4-E4B-it. It uniquely combines harmful content detection, custom policy enforcement, and attack detection capabilities within a single model, supporting a substantial maximum input length of 32,768 tokens.

Key Capabilities

• Comprehensive Harmful Content Detection: Identifies and categorizes harmful user inputs and model responses across 23 predefined safety categories (S1-S23).
• Custom Policy Support: Allows users to define and apply custom policies without retraining, enabling flexible content moderation based on specific service requirements.
• Robust Attack Detection: Effectively detects various attack types, including jailbreaks and prompt injections, enhancing the security of AI systems.
• Multilingual Support: Optimized for both Korean and English content, making it suitable for diverse applications.

Performance Highlights

GenGuard 1.0 Light demonstrates strong performance across multiple benchmarks, often outperforming comparable models. It achieves an average score of 84.3% across Korean Safety, English Safety, Attack Detection, Custom Policy, and Long Context evaluations. Notably, it leads in Korean Safety (87.3%), Attack Detection (93.7%), and Long Context (88.3%). The model also offers competitive inference speeds, with a median response time of 40ms and processing 59 requests per second under concurrent load.

Good for

• Developers needing a unified solution for AI safety and content moderation.
• Applications requiring dynamic custom policy enforcement.
• Systems vulnerable to prompt injection or jailbreak attacks.
• Services operating with both Korean and English content.