DuoGuard

DuoGuard is a family of compact safety classifiers (0.5B, 1B, and 1.5B variants) built on Qwen 2.5 and Llama 3.2 backbones, introduced in February 2025 with an arXiv paper (2502.05163), open code, and public weights. It is trained with a two-player reinforcement learning framework in which a generator and the guardrail co-evolve, synthesizing progressively harder multilingual safety data to harden the classifier.

The model emits independent probabilities across 12 harm categories following an MLCommons-style hazard taxonomy, plus a jailbreak-prompt category, and is fine-tuned primarily for English, French, German, and Spanish with broader coverage inherited from its base models. Release date is era-level (February 2025) matching the code, paper, and weights release announcement.