HaloGuard 1.0: Open-weights AI safety classifier achieves SOTA performance

Researchers have introduced HaloGuard 1.0, an open-weights constitutional classifier designed for AI safety. This model demonstrates state-of-the-art performance on multilingual prompt-safety benchmarks, achieving high F1 scores with significantly smaller model sizes compared to existing leading open guard models. HaloGuard 1.0 utilizes a natural-language constitution of 46 policies to drive synthetic data generation and employs a two-tier harmless design to address false positives. The models are released as open weights, with continuous adversarial red-teaming to enhance their robustness against various attacks. AI

IMPACT Provides a more efficient and accessible tool for ensuring AI safety across multiple languages.

RANK_REASON The cluster describes the release of a new open-weights model for AI safety, detailed in a research paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• HaloGuard 1.0

• model release
• safety
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →