This project addresses the limitations and potential implications of Large Language Models (LLMs), both proprietary and open-source software (OSS), particularly their application in languages other than English. While LLMs have shown potential in comprehending natural languages and undertaking various intellectual tasks, these models are largely monolingual and display poor understanding of non-English languages. While LLMs have shown problem-solving skills and increased productivity, there are significant concerns regarding the generation of harmful and offensive content. LLMs have been observed to create harmful, toxic, biased, and nonfactual content intentionally or unintentionally. Additionally, proprietary models that have undergone safety training have also been jailbroken using different methods of prompt engineering. Models that are robust in English have been jailbroken using low-resource languages. Based on the aforementioned issues, this research proposes a study focusing on two key areas: (1) multilingualism in OSS LLMs, with an emphasis on exploring strategies to align LLMs for multilingual tasks, and (2) the development of safe and robust multilingual LLMs, addressing the potential for these models to generate harmful content in multilingual settings.
Current Team Members:
PI: Vahid Behzadan
Affiliate Organizations:
N/A
Tools and Datasets:
Nepali 33B Model: https://huggingface.co/saillab/Nepali_33B/
Persian 33B Model: https://huggingface.co/saillab/g33b_persian
Code: N/A
Publications:N/A