A state publishes weights in traditional Chinese
On 15 April 2024 Taiwan's TAIDE programme gave the public its LX-7B model built on Llama 2, and in under half a month it was downloaded more than six thousand times. When Llama 3 appeared on 19 April, the team assembled a version on it in four days and released Llama3-TAIDE-LX-8B-Chat-Alpha1 on 29 April.
Why it matters
Here the state acts neither as regulator nor as buyer of compute but as publisher of weights: the model is made in traditional Chinese on Taiwanese archives so that its text carries neither simplified script nor mainland vocabulary. The four days between somebody else's release and its own show that the cost of this kind of sovereignty is the cost of further pretraining, not the cost of a model.
The National Science and Technology Council pulled together industry, universities and research institutions in early 2023. A year produced two models on Llama 2: LX-7B as the commercial version and LX-13B for academic and research use. The Council names five tasks where they excel, writing articles, writing letters, writing summaries, English to Chinese translation and Chinese to English translation, and adds the ability to hold multi-turn dialogue and avoid inappropriate responses. Data was collected by the Policy Research and Information Center and the computing environment prepared by the National Center for High-performance Computing, both of NAR Labs. The year was showcased on 3 May 2024. What the record does not claim. The page names no training corpus size, so the hundreds of billions of characters the report gave are not here. Nor is there any measured comparison against base Llama: the five tasks are named as strengths, without numbers, and that is the programme's own assessment. Confidence is therefore medium. The report named LX-7B and 8B and four tasks; in fact the year produced 7B and 13B on Llama 2, the 8B came separately on Llama 3, and there are five tasks.