Fine-tuning large language models for automating NEF API calls

Thumbnail Image

Files

master_Khan_Zainab_2025.pdf (2.35 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-03-23

Department

Major/Subject

Security and Cloud Computing

Mcode

Degree programme

Master's Programme in Security and Cloud Computing

Language

en

Pages

56

Series

Abstract

To support easier integration and better adaptability, Representational State Transfer (REST) Application Programming Interfaces (APIs) are used in many domains. They help system developers keep each functionality and feature isolated and wellmaintained at the same time. Within the telecommunication domain, Service-Based Architecture (SBA) exposed the control plane functionality and data repositories of the 5G network with the help of Network Functions (NFs). These NFs used REST APIs to further expose their functionalities; one such example is Network Exposure Function (NEF) API. These APIs are growing exponentially with the addition of new NFs and underlying functionalities. Hence, it started to become cumbersome for system administrators to grasp the technical feasibility of each one of them and invoke the correct API to address user inquiries. To solve this issue, this thesis demonstrates a fine-tuning experiment on open-source Large Lan uage Models (LLMs) to observe the performance of fine-tuned models in comparison to their non-fine-tuned counterparts as well as one closed-source model, i.e., GPT-4. Although there have been multiple research experiments around the usage of Generative Artificial Intelligence (GenAI) and LLMs to understand and invoke REST APIs, but none of them were focused on specifically telecommunication-related APIs and also not all of them worked around open-source LLMs. Hence, this thesis fine-tuned two open-source models, namely, Phi-2 and Mixtral, on a dataset consisting of JavaScript Object Notation (JSON) objects of API requests and answers, and evaluated the performance of Phi-2 via two metrics, i.e., GPT4Ref and BertScore, to assess accuracy and similarity of answers with ground truths. The results appeared to be significantly improved in terms of accuracy and similarity with the realization of how effective it can be to use LLMs on domain-specific problems, given the right form of data and sufficient computational resources.

Description

Supervisor

Hellas, Arto

Thesis advisor

Papadimitratos, Panagiotis
Thakur, Mukesh

Keywords

generative AI, large language models, fine-tuning, REST APIs, parameter-efficient fine-tuning, automation

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202505193782

Collections

[dipl] Perustieteiden korkeakoulu / SCI