This study aims to improve code generation performance by applying parameter-efficient fine-tuning using Quantized Low-Rank Adaptation (QLoRA). Currently, large language models (LLMs) in code generation continue to face deployment challenges in low-resource environments, particularly due to high computational demands. The core problem addressed in this study is the inefficiency and limited adaptability of pre-trained models in producing correct code under constrained resource conditions, which results in decreased output quality and restricts accessibility for low-resource users. While previous approaches have employed fine-tuning on large-scale datasets to mitigate these issues—yielding improvements in generalization—they remain hindered by substantial memory usage and computational cost. This study analyzes a compact fine-tuning pipeline utilizing QLoRA, applied to the Qwen2.5-Coder-0.5B-Instruct model, to address these constraints and improve generation accuracy with minimal resource consumption. The proposed system was fine-tuned using two benchmark datasets—CodeExercise-Python-27k and Tested-22k-Python-Alpaca—and demonstrated performance improvements of up to 7.3% on HumanEval and 4.3% on HumanEval in pass@1 metrics, compared to the base model. These findings confirm that fine-tuning with specific datasets, with lightweight methods like QLoRA, significantly enhances the effectiveness of compact LLMs in code generation, contributing to advancements in software engineering, AI-assisted learning, and low-resource-constrained development platforms.