E-commerce Sentiment Analysis Using Fine-tuned LLaMA3 Models: A QLoRA - based Approach

With the rapid expansion of e-commerce platforms and the surge of user-generated content, accurate sentiment analysis of consumer reviews has become essential for business intelligence and customer relationship management. Traditional methods struggle with the linguistic complexity and diversity of online reviews. To address this challenge, this study proposes a fine-tuned LLaMA3 model using the QLoRA (Quantized Low-Rank Adaptation) method for e-commerce sentiment analysis. Experiments were conducted on a dataset of 4,846 Amazon reviews annotated with three sentiment categories: positive, negative, and neutral. Results reveal substantial improvements through domain-specific fine-tuning. While the pre-trained LLaMA3 baseline achieved only 0.37 overall accuracy with a strong bias toward neutral classification, the fine-tuned model reached 0.86 accuracy with balanced performance across all classes. Notably, the F1-score for the “negative” class increased from 0.01 to 0.87, while “positive” and “neutral” classes achieved 0.90 and 0.81, respectively. The QLoRA approach ensures computational efficiency, enabling fine-tuning on consumer-grade hardware. This research highlights the importance of domain-specific adaptation for large language models and provides a practical framework for deploying scalable sentiment analysis in e-commerce.