A Model For Automated Code Debugging Using Small Language Models

Abstract

The purpose of this paper is to assess the performance of small language models in fixing golang concurrency bugs with a quantitative and qualitative analysis[1]. Comparing the results of the bugs fixes against the results of Gfix and Copilot using the gfix dataset. There have been a lot of models and frameworks that have been made to try and automate the repair of bugs. However very few attempts have been made to fix concurrency bugs and none was specifically made using a language model. As such this paper attempted to find if it is possible to solve concurrency bugs using small language models. Making use of the Gcatch/Gfix dataset to produce fixes for bugs and comparing them against copilot and GFix. We find that small language models are better than GFix in the quantity and variety of bugs they can fix but perform worse as compared to copilot.