Network Self-Configuration based on Fine-Tuned Small Language Models

TL;DR

This work presents SLM_netconfig, a hybrid on-premise framework that uses fine-tuned small language models and an agent-based architecture to translate natural-language network intents into valid configurations. By pairing dataset-engineered, domain-specific fine-tuning with a built-in verifier and structured prompting, it achieves higher syntax and goal accuracy with reduced translation time compared to prior LLM-based systems. The approach emphasizes privacy, efficiency, and scalability, demonstrated through a two-model setup (requirement-configuration and question-configuration) and a rigorously constructed vendor-document dataset pipeline. The results indicate that small, domain-tuned models can rival larger LLMs for autonomous network configuration, with practical benefits for on-site deployment and multi-vendor adaptability.

Abstract

As modern networks grow in scale and complexity, manual configuration becomes increasingly inefficient and prone to human error. While intent-driven self-configuration using large language models has shown significant promise, such models remain computationally expensive, resource-intensive, and often raise privacy concerns because they typically rely on external cloud infrastructure. This work introduces SLM_netconfig, a fine-tuned small language model framework that uses an agent-based architecture and parameter-efficient adaptation techniques to translate configuration intents expressed as natural language requirements or questions into syntactically and semantically valid network configurations. The system is trained on a domain-specific dataset generated through a pipeline derived from vendor documentation, ensuring strong alignment with real-world configuration practices. Extensive evaluation shows that SLM_netconfig, when using its question-to-configuration model, achieves higher syntactic accuracy and goal accuracy than LLM-NetCFG while substantially reducing translation latency and producing concise, interpretable configurations. These results demonstrate that fine-tuned small language models, as implemented in SLM_netconfig, can deliver efficient, accurate, and privacy-preserving automated configuration generation entirely on-premise, making them a practical and scalable solution for modern autonomous network configuration.

Figures (11)

• Figure 1: SLM_netconfig Operation.
• Figure 2: Fine-tuning network configuration models
• Figure 3: Syntax Accuracy of SLM_netconfig models - Question vs Requirement
• Figure 4: Format Accuracy of SLM_netconfig - Question vs Requirement
• Figure 5: Goal Accuracy of SLM_netconfig - Question vs Requirement