離開
進入《TAIPEI TIMES》 New traditional Chinese LLM beats GPT-4o
The words artificial intelligence are pictured in an illustration taken on May 4. Photo: Reuters
By Kayleigh Madjar / Staff writer, with CNA
An open-source traditional Mandarin large language model （LLM） localized for Taiwan was released yesterday, achieving higher scores in the Taiwanese bar exam than GPT-4o.
The project, called Taiwan Mixture of Experts （TAME）, was initiated by Chang Chun Group, Pegatron, Unimicron, Chang Gung Memorial Hospital, TechOrange and National Taiwan University.
Using Nvidia’s Taipei-1 supercomputer in Kaohsiung, it was trained on nearly 500 billion tokens of data provided by the partners in their specialized fields.
On its first attempt taking last year’s bar exam, Project TAME did better than 89 percent of test takers, LegalSign.ai chief executive officer Steven Chen （陳啟桐） said.
After only a few months, it has already received outstanding marks when taking exams for university students, legal professionals, traditional Chinese medicine practitioners, tour guides and drivers, Project TAME said.
Out of 39 evaluations and nearly 3,000 questions, Project TAME outperformed all other available models, with an accuracy rate 6.8 percent higher than the second-place Claude-Opus and 9.3 percent higher than GPT-4o.
The model is available to use for free online at https://twllm.com and would be open source, organizers said.
An interactive chatbot version for the public was created by AMPIC.
It is also available for download by enterprises and developers on Github.
The new LLM is the result of 350,000 GPU hours and 1,285 working hours by 31 engineers, Project TAME leader and Pegatron associate vice president of AI development Andrew Hsiao （蕭安助） said.
Its two specialties are localization for Taiwan and application for industry, he said.
It was only released today, but enterprises would be able to verify and fine-tune the model as time goes on to verify whether it is suitable for use in different fields, he added.
新聞來源：TAIPEI TIMES
