The open-source Traditional Chinese expert model project, "TAiwan Mixture of Experts" (Project TAME), was jointly initiated by Chang Chun Group, Pegatron Corporation, Chang Gung Memorial Hospital, Unimicron, and Tech News, in collaboration with the Department of Computer Science and Engineering, the Department of Information Management, and Luguo Technology at National Taiwan University. The project, trained with the assistance of NVIDIA, utilizes local culture and terminology, combined with industry expertise, enabling Taiwanese industries to rapidly introduce generative artificial intelligence technology applications.
For example, if a typical enterprise wants to train a 10 billion parameter model from scratch, it usually costs NT$3.8 million and takes up to 576 hours. However, with Project TAME's 700 billion parameter model, the cost of implementing it is only NT$1600 million. At the same time, training of internal corporate data can be completed in just 3.5 hours, significantly reducing implementation costs and time, thereby enabling rapid optimization of various AI technology applications such as corporate operations management, personnel training, product services, and customer service.
This Project TAME Traditional Chinese Expert Model open source project is led by Associate Professor Chen Minnong of the Department of Computer Science at National Taiwan University, and his laboratory colleagues and corporate partner development teams. Through technical assistance from the NVIDIA Developer Program, experts from multiple vertical industries contributed professional field data to pre-train nearly 5 billion characters (tokens) to develop a large-scale Traditional Chinese language expert model.
With no contractual constraints, industry and academia have spontaneously collaborated, and in just a few months, the Project TAME Traditional Chinese model has achieved initial, concrete, and significant results. It leads the world in various Traditional Chinese-related indicators, and has even achieved excellent scores in Taiwan's "University College Entrance Examination, Bar/Traditional Chinese Medicine Examination, Tour Guide License, Driver's License, and Taiwan Localization Test." At the same time, in Taiwan's 39 comprehensive evaluations and nearly 3,000 questions, Project TAME's scores exceeded all other models, with an accuracy rate 6.8% higher than the second-place Claude-Opus model and 9.3% higher than OpenAI's GPT-4o.
After the Project TAME Traditional Chinese Expert Model is officially released, it will be available as open source.Provide externallyAssociate Professor Chen Minnong of the Department of Computer Science and Engineering at National Taiwan University pointed out that there may be common intersections between different data that can be used. This is better for the development of language models and can also allow experts from different industries to contribute data from their own fields.
