Large Language Models for Mental Health Counselling

Mental health challenges are becoming increasingly prevalent worldwide, there is a growing demand for scalable and accessible support systems. This paper explores the application of Large Language Models (LLMs) in providing personalized counselling for mental health. This study evaluates the performance of three pre-trained models—T5-Small, T5-Base and BART-large on a dataset of mental health counselling queries. Metrics such as BLEU, ROUGE, and perplexity were used to assess the quality of model-generated responses before and after fine-tuning. Further, we evaluated the finetuned models using BERTScore. Results reveal that fine-tuning significantly improves model coherence and relevance, with noticeable reductions in perplexity. The findings highlight the potential of LLMs in assisting mental health professionals by providing empathetic and contextually relevant responses.