inclusionAI/SingGuard-2b

SingGuard-2b: Policy-Adaptive Multimodal Guardrail

SingGuard-2b is a 2 billion parameter multimodal guardrail model developed by inclusionAI, specializing in dynamic safety assessment. Unlike traditional guardrails with fixed taxonomies, SingGuard-2b allows deployment teams to supply active safety policies at runtime, enabling evaluation against custom natural-language rules without requiring model retraining. This flexibility is crucial for practical moderation settings where risks can arise from user queries, images, model responses, or cross-modal compositions.

Key Capabilities

• Unified Multimodal Moderation: Supports safety assessment across text, image, image-text, and multilingual inputs, covering both query-side and response-side scenarios.
• Dynamic Policy Adaptation: Accepts custom safety rules via a policy argument, judging content only against the specified rules and outputting both a safe/unsafe judgment and the matched risk category.
• Dynamic Reasoning Flow: Features a "fast-slow" mode for immediate safety signals followed by deeper reasoning for precise judgments, and a "fast" mode for compact binary judgments.
• 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.

Good For

• Developers needing a flexible and powerful guardrail for diverse content moderation tasks.
• Applications requiring real-time adaptation to evolving safety policies without model retraining.
• Ensuring safety across complex multimodal interactions in LLM-powered systems.