Developing an Efficient Suicide-Detection System: How Small Models Compare to LLMs

Abstract: As language evolves across generations and suicide rates continue to rise, automated detection of suicidal ideation in social media text has become increasingly important for early intervention. LLMs demonstrate strong accuracy, but their computational cost and large size limit real-world uses and implementation. This study evaluates whether smaller models can achieve comparable performance for suicide detection. A diverse set of models, ranging from thousands to trillions of parameters, was accessed across four datasets from Reddit and Twitter, two widely used social media platforms. All traditionally trained models were smaller than the LLMs. Results show that smaller models can achieve performance comparable to or exceeding LLMs while requiring substantially fewer computational resources. Model performance does not scale directly with parameter count, although large models tend to be more consistent across domains. These findings suggest that lightweight models can support deployable, real-time suicide detection systems, which are capable of flagging at-risk individuals and facilitating mental health intervention.