Scholarlink

<p>This paper introduces TCNNet-9B, a specialized Traditional Chinese language model developed to address the specific requirements of the Taiwanese networking industry. Built upon the open-source Yi-1.5-9B architecture, TCNNet-9B underwent extensive pretraining and instruction finetuning utilizing a meticulously curated dataset derived from multi-source web crawling.&#xa0;The training data encompasses comprehensive networking knowledge, DIY assembly guides, equipment recommendations, and localized cybersecurity regulations. Our rigorous evaluation through custom-designed benchmarks assessed the model’s performance across English, Traditional Chinese, and Simplified Chinese contexts.&#xa0;The comparative analysis demonstrated TCNNet-9B’s superior performance over the baseline model, achieving a 2.35-fold improvement in Q&amp;A task accuracy, a 37.6% increase in domain expertise comprehension, and a 29.5% enhancement in product recommendation relevance. The practical efficacy of TCNNet-9B was further validated through its successful integration into Hi5’s intelligent sales advisor system. This research highlights the significance of domain-specific adaptation and localization in enhancing large language models, providing a valuable practical reference for future developments in non-English contexts and vertical specialized fields.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Localized large language model TCNNet 9B for Taiwanese networking and cybersecurity

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月20日

查看全文
  • Jiun-Yi Yang,
  • Chia-Chun Wu

摘要

This paper introduces TCNNet-9B, a specialized Traditional Chinese language model developed to address the specific requirements of the Taiwanese networking industry. Built upon the open-source Yi-1.5-9B architecture, TCNNet-9B underwent extensive pretraining and instruction finetuning utilizing a meticulously curated dataset derived from multi-source web crawling. The training data encompasses comprehensive networking knowledge, DIY assembly guides, equipment recommendations, and localized cybersecurity regulations. Our rigorous evaluation through custom-designed benchmarks assessed the model’s performance across English, Traditional Chinese, and Simplified Chinese contexts. The comparative analysis demonstrated TCNNet-9B’s superior performance over the baseline model, achieving a 2.35-fold improvement in Q&A task accuracy, a 37.6% increase in domain expertise comprehension, and a 29.5% enhancement in product recommendation relevance. The practical efficacy of TCNNet-9B was further validated through its successful integration into Hi5’s intelligent sales advisor system. This research highlights the significance of domain-specific adaptation and localization in enhancing large language models, providing a valuable practical reference for future developments in non-English contexts and vertical specialized fields.