Overview
yentinglin/Llama-3-Taiwan-8B-Instruct is an 8 billion parameter language model built on the Llama-3 architecture, specifically fine-tuned for Traditional Mandarin and English. Developed by yentinglin, this model leverages a substantial corpus of high-quality Traditional Mandarin and English data, encompassing general knowledge and industrial-specific information from legal, manufacturing, medical, and electronics domains. It features an 8K context length and is released under the Llama-3 license.
Key Capabilities
- Bilingual Proficiency: Strong performance in both Traditional Mandarin (zh-tw) and English (en).
- Domain-Specific Knowledge: Enhanced understanding and generation in specialized fields like legal, manufacturing, medical, and electronics due to its training data.
- Core LLM Functions: Demonstrates robust capabilities in language understanding, generation, reasoning, and multi-turn dialogue.
- Function Calling: Supports function calling, making it suitable for structured output generation when combined with constrained decoding (e.g., JSON mode).
Good For
- Multiturn Dialogue: Engaging in extended, coherent conversations in Traditional Mandarin and English.
- Retrieval-Augmented Generation (RAG): Effective for tasks requiring information retrieval and synthesis, as demonstrated by its web search integration example.
- Structured Output & Entity Recognition: Generating formatted outputs, performing language understanding, and identifying entities, particularly useful with constrained decoding for JSON output.
- Taiwanese Contexts: Excels in benchmarks relevant to Taiwan, such as TMLU, Taiwan Truthful QA, and Legal Eval, outperforming several larger models in specific Traditional Mandarin tasks.