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

## TL;DR

This paper introduces TCNNet-9B, a large language model tailored for the Taiwanese networking and cybersecurity industry, showing improved performance in specialized tasks.

## Contribution

The novel contribution is the development of TCNNet-9B, a localized Traditional Chinese language model optimized for networking and cybersecurity domains in Taiwan.

## Key findings

- TCNNet-9B achieved a 2.35-fold improvement in Q&A task accuracy compared to the baseline.
- The model showed a 37.6% increase in domain expertise comprehension and 29.5% improvement in product recommendation relevance.
- TCNNet-9B was successfully integrated into Hi5’s intelligent sales advisor system.

## Abstract

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.