The research introduces RTP-LX, a unique dataset specifically targeting the detection of toxic content in 28 languages using Large Language Models (LLMs). The study emphasizes the importance of culturally sensitive and multilingual toxic content evaluation.### Key Findings from the Study
Participatory design for dataset creation ensures inclusivity and relevancy.
Majority of the evaluated S/LLMs show acceptable accuracy but struggle with cultural nuances like microagressions or bias, underscoring the challenge of scale in toxicity detection.
The release of this corpus aims to enhance the safe deployment of AI systems by mitigating harmful language use.
This work highlights the critical need for tools that can operate across cultural and linguistic boundaries to ensure ethical AI deployments.
The comprehensive nature of the dataset enables further research on improving LLMs’ performance in culturally complex scenarios, potentially directing new developments in AI safety and bias detection.