@phdthesis{digilib72924,
           month = {August},
           title = {RANCANG BANGUN VISUAL STUDIO CODE EXTENSION CODEVA UNTUK CODE GENERATOR JAVA SECARA OTOMATIS MENGGUNAKAN LARGE LANGUAGE MODELS (LLMs)},
          school = {UIN SUNAN KALIJAGA YOGYAKARTA},
          author = {NIM.: 21106050042 Muhammad Arif Rakhman Azizi},
            year = {2025},
            note = {Dr. Agung Fatwanto, S.Si., M.Kom.},
        keywords = {Large Language Models, Fine Tuning, Ekstensi, Unsloth},
             url = {https://digilib.uin-suka.ac.id/id/eprint/72924/},
        abstract = {The rapid advancement of technology has spurred a significant demand for
software, yet the Coding phase, particularly in a complex language like Java,
remains a time-consuming and error-prone process. To address this, a Visual
Studio Code Extension named CodeVa was developed to automatically generate
Java code using Large Language Models (LLMs).
Developed using the Agile Extreme Programming method ology, CodeVa
integrates a fine-tuned Open-source model, Qwen2.5 Instruct 7B. This model was
specifically trained for Java code generation using the Parameter-Efficient Finetuning
(PEFT) technique, Low-Rank Adaptation (LoRA), and the Unsloth
framework.
The fine-tuning process resulted in substantial improvements across various
performance metrics, including an increase in the CodeBLEU score from 0.3204 to
0.3521 and significant gains in ROUGE scores ROUGE-1 from 0.1736 to 0.3705,
ROUGE-2 from 0.0251 to 0.2278, and ROUGE-L from 0.0991 to 0.3252 and
Pass@10 using kulal version and chen version from 73.17\% to 73.78\%. Usability
testing of CodeVa yielded a System Usability Scale (SUS) score of 81.0,
categorizing it as "excellent" and "acceptable." This indicates that CodeVa is a
highly effective tool that can significantly enhance developer productivity by
facilitating faster and more efficient Java Coding .}
}