World's first geographic multimodal premiered in China

The research team established a comprehensive geographic corpus covering four major categories and sixteen subcategories, providing 32.3 billion geographic tokens for the large model's self-supervised learning. They also crafted over 40,000 high-quality geographic instructions for model fine-tuning.

"Compared to general-purpose large language models, Sigma Geography is more familiar with the language patterns, professional terminology, and domain knowledge of geography, leading to an accuracy improvement of around 31 percent on the geographic benchmark test set," said Su Fenzhen, deputy director of the Institute of Geographic Sciences and Natural Resources Research.

Su compared Sigma Geography with Chat-GPT in answering geographic professional queries and found that Sigma Geography provided more accurate and comprehensive responses, tailoring answers to the roles of different questioners.

The research team explained that this was achieved through their innovative user profile, precise discrimination and response technology. Consequently, Sigma Geography can fully consider the cognitive and expressive differences in geographic scientific knowledge systems among geography enthusiasts, geography students, and researchers, providing tailored solutions to geographic professional questions that align with users' knowledge structures through a combination of text and images.