This research introduces RTP-LX, a dataset aimed at evaluating language models for toxicity in 28 languages. This corpus, enriched through participatory design, aims to tackle challenges in detecting culturally-specific toxic content, ensuring safer deployments of multilingual models globally.

• Designed for multilingual safety evaluation.
• Includes 28 languages.
• Utilizes participatory design for dataset creation.
• Focuses on detecting culturally-specific toxic language.
• Supports safer model deployments globally.

This study is crucial as it addresses the increasing necessity for models that can operate safely across different languages and cultures. It sets a foundation for future works to build upon, ensuring that as technology reaches a global audience, it remains inclusive and safe.