Implementing zero-shot learning in Code-LLM for effective code explanation generation

Abstract

As software systems, particularly expansive codebases such as Systems on Chip (SoCs), grow increasingly complex, there is a heightened demand for improved code interpretability tools. This thesis explores the development of a Code Explanation Generation Tool leveraging the CodeLlama 13B Instruct model in a Zero-Shot learning approach. Without additional fine-tuning, this endeavor utilizes efficient prompt engineering and internal model parameter manipulation. We assess the semantics and readability of the generated explanations through a human-annotated dataset crafted by SoC engineers within the Nokia group, specifically for this thesis, alongside the publicly available Code-NL pair dataset from the CoNala Corpus.

The generated code explanations are designed to enhance problem localization and comprehension within Nokia Mobile Networks Solutions’ SoC codebases. The study further investigates how the use of connectives affects the coherence and clarity of explanations and their impact on the trustworthiness of these explanations. Additionally, we evaluate the model’s capacity to manage System Verilog and VHDL code without specific tuning.

This research advances the application of Large Language Models (LLMs) in software development and lays the foundation for a searchable vector database for SoC specifications, aiming to boost productivity and accuracy in Nokia’s operations.