Analisis Efektivitas Fine-Tuning dan Prompt Engineering Berbasis Llama 3.1 pada Deteksi Depresi di Media Sosial

Abstrak

Detecting depression has become an important concern in addressing mental health issues. According to WHO, more than 300 million people suffer from depression. Large Language Models offer great potential to address this issue, however the full fine-tuning process is often hampered by heavy computational requirements, and LLMs that are not specifically configured for a particular context can result in biased and inaccurate outcomes. This study aims to analyze the effectiveness of Prompt Engineering and Fine-Tuning using QLoRA in improving the accuracy of depression detection. Utilizing the Llama-3.1-8B-Instruct model on social media datasets, this research compares model performance in two scenarios consist of the application of direct prompting strategies on the base model and the application of QLoRA fine-tuning. Evaluation results demonstrate that the Chain-of-Thought strategy improved baseline accuracy from 81.4% to 84.4%, but still exhibited significant bias towards the 'Severe' class. In contrast, the QLoRA Fine-Tuning approach proved superior, achieving 92.4% accuracy with balanced F1-Scores across classes, effectively eliminating detection bias in the 'Minimum' class. These findings confirm that while prompting techniques can enhance baseline performance, QLoRA provides a more accurate, stable, and objective solution for depression detection tasks.