Advancements in Large Language Models (LLMs) have led to significant progress in text generation and understanding of human cognition. Despite these advancements, their application in mental health counselling remains under-explored. This paper focuses on framing a proper dataset for fine-tuning lightweight, small models for mental health counselling. In this study, BART has been fine-tuned, and the ability of the lightweight model to deliver empathetic, relevant, and contextually aware responses has been analysed, highlighting the potential that such models hold for real-time, scalable mental health support. The results indicate that BART can generate coherent and emotionally supportive responses, with efficient response time and perplexity performance. The paper also explores the efficiency of lightweight model deployment in real-world mental health applications. This work demonstrates the promise of using small-scale models (LLMs) to address mental health concerns and lays the groundwork for further improvements in fine-tuning and model evaluation.