Taiwan-LLM

Why we included this project

Chinese LLMs are usually built around Simplified Chinese and mainland conventions, which leaves Taiwan's Traditional Mandarin speakers underserved. Taiwan-LLM is a set of models fine-tuned on Traditional Mandarin and English data with Taiwan's usage patterns and cultural context in mind, including the 70B-parameter Llama-3-Taiwan-70B, which delivers state-of-the-art results on Traditional Mandarin NLP benchmarks. The repo doubles as a working toolkit: it ships Axolotl fine-tuning configs, links to hosted demos and model weights, and documents partnerships with organizations in medical, manufacturing, and legal fields, so teams can adapt the models to their own domain instead of starting from a generic base. It is research-grade software, so expect to benchmark it against your own data before relying on it in production.