inclusionAI/Sing-Guard-4b

SingGuard: Policy-Adaptive Multimodal LLM Guardrail

SingGuard is a 4 billion parameter multimodal guardrail model developed by inclusionAI, designed for comprehensive safety assessment across various content types. Unlike traditional guardrails, SingGuard treats the active safety policy as a runtime input, enabling deployment teams to evaluate content against default categories or custom natural-language rules without requiring model retraining. This allows for highly flexible and adaptive moderation in practical settings.

Key Capabilities

• Unified Multimodal Moderation: Supports safety assessment for text, images, image-text combinations, and multilingual content, covering both query-side and response-side scenarios.
• Dynamic Policy Adaptation: Accepts custom safety rules via a policy argument, allowing the model to judge content exclusively against the provided rules.
• Strong Benchmark Performance: Achieves state-of-the-art average performance across six major benchmark categories, including multimodal, image-only, text query, text response, and multilingual safety.
• Dynamic Reasoning Flow: Features a fast first-token routing for immediate safety signals, with deeper reasoning for more precise judgments when needed.

Good For

• Content Moderation: Ideal for platforms requiring robust and adaptable moderation of user-generated content, model responses, and multimodal inputs.
• Custom Safety Policies: Developers who need to implement specific or evolving safety guidelines without constant model retraining.
• Multilingual Applications: Safeguarding content across diverse languages and modalities efficiently.