Finetuning Large Language Models for Suicide Risk Level Assessment on Social Media

Suicide remains a pressing global issue, and early detection of individuals at risk is vital for prevention. Traditional methods of suicide risk assessment rely heavily on questionnaires and interviews. However, social media offers a rich source of realtime data, providing an opportunity for automated suicide risk detection. In this study, we propose a novel method leveraging large language models (LLMs) to assess suicide risk based on social media activity. Our approach integrates RoBERTa, Llama 3.1, and Gemma 2 through a voting mechanism to enhance predictive performance. Results indicate that our model achieves an F1 score of 0.7006, outperforming traditional machine learning and deep learning models. We also observe that while fine-tuning large models on limited data yields better results than pre-trained models, increasing model size does not always improve performance. The use of a voting mechanism significantly improves the accuracy of the predictions, demonstrating the utility of multi-model approaches in suicide risk detection.