Network Intent Translation Using LoRA Fine-Tuned LLMs

Intent translation, a key enabler of Intent-Based Networking (IBN), converts high-level requirements into low-level deployable configurations. While traditional translation schemes targeted specific implementations and were difficult to generalize, Large Language Model (LLM) based translations rely on prompt engineering using proprietary models, such as ChatGPT. This entails significant financial costs and raises privacy concerns, as sensitive infrastructure data may be exposed in the prompts. In this work, we propose a cost-efficient intent translation framework using open-source LLMs and Parameter-Efficient Fine-Tuning (PEFT). Fine-tuning open-source models preserves privacy as they can be deployed locally without sharing sensitive information. We investigate PEFT-tuning general-purpose LLMs (Llama 2 and Mistral) against a code-specialized LLM (Code Llama 2) for intent-to-vendor translation using a domain-specific dataset. We propose a comprehensive evaluation procedure that combines both numerical score-based and human-centered approaches. We compare the performance of the fine-tuned models to prompt engineering with ChatGPT-4o mini. Results show that the PEFT-tuned Code Llama 2 outperformed ChatGPT and other models in terms of understanding the user's intent and generating deployable configurations. These findings demonstrate the potential of open-source LLMs as a scalable, cost-efficient, and confidentiality-preserving solution for IBN deployments.