zrwang1211/SafeAtlas-Guard-2B

SafeAtlas Guard 2B: Multimodal Safety Model

SafeAtlas Guard 2B, developed by zrwang1211, is a 2 billion parameter multimodal safety model introduced in the paper "SafeAtlas-VL: Beyond Binary Multimodal Safety with Large-Scale Data and Guard Models." It is built upon the Qwen3-VL-2B-Instruct backbone and is specifically engineered to assess the safety of image content, image-grounded user requests, and assistant responses.

Key Capabilities

• Comprehensive Safety Evaluation: Provides a five-level ordered safety label (e.g., safe core, unsafe core) and a continuous risk score from 0 to 100.
• Multimodal Analysis: Evaluates safety across images, text requests, and text responses, offering auxiliary safety-judge predictions for request and response targets.
• Detailed Output: Returns various prediction fields including safety_label, risk_score, category, and teacher_predictions.
• Robust Training: Utilizes a two-stage training process, including full-parameter multimodal instruction tuning and specialized training for ordinal, harm-category, and teacher-simulation heads.
• Strong Benchmarks: Achieves an overall average F1 score of 80.43% across 11 benchmarks, including BeaverTails-V, SPA-VL, VLGuard, and HarmBench.

Good For

• Multimodal Safety Moderation: Ideal for automatically identifying and flagging unsafe content in applications involving images and text.
• Ordinal Risk Assessment: Provides granular risk scores and safety levels, useful for nuanced content moderation policies.
• Red-Teaming and Evaluation: Can be used to test the safety boundaries of other AI models and systems.
• Safety Alignment Research: A valuable tool for researchers working on improving AI safety and alignment.