Recent advances in artificial intelligence have seen natural language processing (NLP) applications used increasingly for mental health support, such as question-answering systems, conversational agents, and chatbots. However, these applications often require high-quality dialogue datasets, which are scarce for certain under-resourced mental health conditions like gambling disorders. Privacy concerns and the sensitive nature of mental health issues make real-world data collection difficult and time-consuming. We propose a framework that uses publicly available social media posts as seeds to generate empathy-driven multi-turn dialogues between help-seekers and supporters in mental health conditions that have been under-researched in NLP. Through a case study on gambling disorder, we demonstrate the effectiveness of our approach in producing higher-quality dialogues than standard role-playing prompts. Our work showcases an effective dataset generation method for fine-tuning mental health language models alongside downstream applications like dialogue system development, particularly for low-resourced mental health disorders like gambling disorders.