Institutional Repository of Vilnius University
Didžiųjų kalbos modelių panaudojimas programinio kodo peržiūrų automatizavimui /
Abstract
dc:descriptionThis paper explores the ability of large language models (LLM), in this case “GPT‐4 Turbo” model, to detect software code defects by comparing its effectiveness with traditional static code analysis tools “Snyk Code” and “SonarQube”. The ability to detect code defects was tested using 109 publicly available code samples from different programming languages and it was found GPT‐ 4 Turbo model detected the main code vulnerability defect in 104 out of 109 files (96%), while “Snyk Code” found the main vulnerability in just 62 out of 109 files (57%), and “SonarQube” found them in just 7 files of 74 files tested (7%). For additional testing, a new set of Python code sample dataset with various defects was created. In this dataset, the “GPT‐4 Turbo” model detected main security weakness in all 10 files, “Snyx Code” detected 3 out of 4 main weaknesses, and “SonarQube” detected only 1 out of 10 weaknesses. These results demonstrate that large language models are capable to detect defects in small code bases, and are able perform even better than some static code analysis tools.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Petrikas, Mantas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD210648339&prefLang=en_US