SeaLLMs - Large Language Models for Southeast Asia

Explore the development of SeaLLMs, a groundbreaking series of language models designed specifically for Southeast Asian languages, in this 40-minute seminar presented by Phi, a senior research engineer at DAMO Academy, Alibaba Group. Learn how SeaLLMs address the linguistic bias in large language models by focusing on low-resource and regional languages. Discover the innovative approach of building upon Llama-2 and enhancing it through continued pre-training, specialized instruction, and alignment tuning. Gain insights into the comprehensive evaluation demonstrating SeaLLM-13b models' superior performance across various linguistic tasks and assistant-style instruction-following capabilities compared to similar open-source models. Understand how SeaLLMs outperform ChatGPT-3.5 in non-Latin languages like Thai, Khmer, Lao, and Burmese while remaining lightweight and cost-effective. Delve into the speaker's extensive background in multilinguality in large language models and translation technologies, as well as his goal to democratize AI for under-represented communities.