Enhancing Network Engineering Capabilities through LLM Fine-Tuning with Automatically Generated Datasets

Claudiu Trăistaru1, Florin Pop2, Costin Bădică3, Cătălina Mancaş3 and and Ionuţ Murareţu3

  1. University of Science and Technology POLITEHNICA Bucharest, University of Craiova
    Craiova, Romania
    claudiu.traistaru@edu.ucv.ro
  2. University of Science and Technology POLITEHNICA Bucharest, National Institute of Research and Development in Informatics (ICI), and Academy of Romanian Scientists
    Bucharest, Romania
    florin.pop@upb.ro
  3. Computer and Information Technology Department, University of Craiova
    Craiova, Romania
    {costin.badica,catalina.mancas, ionut.muraretu}@edu.ucv.ro

Abstract

The paper presents a method for automatically generating domain-specific datasets to fine-tune open-source LLMs in network engineering. Our objective is to address the increasingly complex nature of network configuration and management jobs by supplying LLMs with high-quality training data. We evaluated datasets generated using open-source LLMs, including DeepSeek-R1 671B, LLaMA 3.1 70B, Qwen 2.5 72B, and Mixtral 8x7B, analyzing the quality of unprocessed knowledge data and the efficacy of cleaning and deduplication methods. The resulting dataset addresses various subjects related to routing, security, and network services. After-ward, we fine-tuned smaller LLaMA 3.2 1B, LLaMA 3.2 3B and Qwen 2.5 1.5B models using Low-Rank Adaptation, thereby minimizing computational demands while maintaining the quality of domain knowledge.

Digital Object Identifier (DOI)

https://doi.org/10.2298/CSIS250416082T

Publication information

Volume 23, Issue 1 (January 2026)
Year of Publication: 2026
ISSN: 2406-1018 (Online)
Publisher: ComSIS Consortium

Full text

DownloadAvailable in PDF
Portable Document Format

How to cite

Trăistaru, C., Pop, F., Bădică, C., Mancaş, C., Murareţu, a. I.: Enhancing Network Engineering Capabilities through LLM Fine-Tuning with Automatically Generated Datasets. Computer Science and Information Systems, Vol. 23, No. 1, 535-560. (2026), https://doi.org/10.2298/CSIS250416082T