Hugging Face yentinglin/Llama-3-Taiwan-70B-Instruct is a 70 billion parameter instruction-tuned language model developed by yentinglin, based on the Llama-3 architecture. It is fine-tuned on a large corpus of Traditional Mandarin and English data, excelling in language understanding, generation, reasoning, and multi-turn dialogue for these languages. The model features an 8K context length and demonstrates strong performance on various Traditional Mandarin NLP benchmarks, including legal, manufacturing, medical, and electronics domains.
Overview
yentinglin/Llama-3-Taiwan-70B-Instruct is a 70 billion parameter language model built upon the Llama-3 architecture, specifically fine-tuned for Traditional Mandarin and English. Developed by yentinglin, this model leverages a substantial corpus of both general and industrial-specific data, including legal, manufacturing, medical, and electronics domains.
Key Capabilities
- Bilingual Proficiency: Strong capabilities in Traditional Mandarin (zh-tw) and English (en).
- Comprehensive NLP: Excels in language understanding, generation, reasoning, and multi-turn dialogue.
- Domain-Specific Knowledge: Enhanced performance in specialized fields due to targeted training data.
- Function Calling: Supports function calling, with recommendations for constrained decoding for JSON mode.
- Long Context: Features an 8K context length, with a 128K version also available.
Performance Highlights
The model demonstrates competitive performance on various Traditional Mandarin NLP benchmarks. For instance, it achieves 74.76% on TMLU and 80.95% on Taiwan Truthful QA, often outperforming or closely matching other large models like Claude-3-Opus and GPT-4o on specific Traditional Mandarin tasks. It also shows strong results on Legal Eval and TMMLU+ benchmarks. The model was trained using the NVIDIA NeMo Framework on NVIDIA DGX H100 systems.
Good for
- Multiturn Dialogue: Engaging in natural and extended conversations.
- RAG (Retrieval Augmented Generation): Enhancing responses with retrieved information, as demonstrated by its web search integration.
- Structured Output & Entity Recognition: Generating formatted outputs, understanding language nuances, and identifying entities, particularly useful with constrained decoding for JSON output.
- Taiwanese Contexts: Ideal for applications requiring deep understanding and generation in Traditional Mandarin, especially within legal, medical, and industrial sectors relevant to Taiwan.