
SeaLLMs introduces a family of language models tailored to Southeast Asian languages, including Burmese, addressing gaps in support from English-focused models. Through continual pre-training on regional language data and vocabulary expansion, the models are adapted to better understand local languages and process their scripts more efficiently. Instruction tuning strengthens their ability to follow user requests, while alignment training incorporates local cultural norms, preferences, and legal considerations.
The paper evaluates both language capabilities and assistant-style responses, reporting improvements over comparable open models on the tasks tested. The authors also report stronger performance than ChatGPT-3.5 in selected non-Latin language evaluations. These findings offer useful approaches for developing Burmese and other regional language models that better serve local communities.