The Design and Development of Multimodal Ancient Chinese Large Language Model AI Jiusi 2.0

Abstract:  [Objective/Significance] With the rapid development of generative artificial intelligence (AIGC), various large models have gradually evolved from initially being capable of processing only single text modality in large language models to being able to handle multi-modal data such as text, images, voice, and video. However, domestic large language models targeting the ancient Chinese language field still mainly focus on improving the performance of ancient Chinese language processing tasks, and are mainly centered on single text modality information processing. There is still considerable room for development in terms of the knowledge understanding and question-answering interaction capabilities of large language models, as well as in multi-modal information processing. Based on this, Huazhong University of Science and Technology has newly launched the ancient Chinese multi-modal large language model “AI Jiusi 2.0”, which not only masters ancient Chinese professional knowledge but also possesses ancient Chinese application capabilities and supports multi-modal data processing, aiming to set an example for the development of multi-modal ancient Chinese large language models. [Method/Process] This paper details the dataset construction, computing power upgrade, model training, and interface optimization of “AI Jiusi 2.0”, and showcases the performance of the new version “AI Jiusi” in ancient Chinese language knowledge and language ability. [Result/ Conclusion] The newly upgraded “AI Jiusi 2.0” demonstrates significant advantages in the understanding of ancient Chinese texts and ancient Chinese knowledge question-answering, and has already acquired a certain ability to understand ancient Chinese characters (oracle bone