inclusionAI/Sing-Guard-8b
SingGuard-8b by inclusionAI is an 8 billion parameter policy-adaptive multimodal guardrail model designed for comprehensive safety assessment. It unifies moderation across text, image, image-text, and multilingual content, handling both query-side and response-side scenarios. A key differentiator is its runtime policy adaptation, allowing users to define custom natural-language safety rules without retraining, and it achieves state-of-the-art performance across six major safety benchmark categories.
Loading preview...
SingGuard-8b: Policy-Adaptive Multimodal Guardrail
SingGuard-8b is an 8 billion parameter multimodal guardrail model developed by inclusionAI, specializing in safety assessment across diverse content types. Unlike traditional guardrails with fixed taxonomies, SingGuard-8b features runtime policy adaptation, allowing deployment teams to define and apply custom natural-language safety rules dynamically without requiring model retraining. This flexibility enables evaluation against default categories or user-defined policies.
Key Capabilities & Features
- Unified Multimodal Moderation: Provides comprehensive safety assessment for text, images, combined image-text inputs, and multilingual content, covering both user queries and model responses.
- Strong Benchmark Performance: Achieves state-of-the-art average performance across six major benchmark categories, including multimodal safety, image-only safety, text query safety, text response safety, multilingual query safety, and multilingual response safety.
- Dynamic Reasoning Flow: Supports a "fast-slow" mode for safety judgments, offering an immediate signal followed by deeper reasoning for precise final categorization. A "fast" mode provides compact binary judgments and risk categories.
- Native Inference Compatibility: Integrates seamlessly with standard Transformers and vLLM chat-style message inputs, simplifying deployment.
When to Use This Model
SingGuard-8b is ideal for applications requiring robust and adaptable content moderation. It is particularly well-suited for scenarios where:
- Safety policies need to be updated or customized frequently without model retraining.
- Multimodal content (text, image, image-text) requires unified safety checks.
- Both user queries and model-generated responses need to be moderated for safety.
- High performance across a broad range of safety benchmarks is critical.