GPT-empowered question-answer dataset for informative and empathetic support for Korean childhood cancer survivors

• Developed the first QA dataset for Korean childhood cancer survivors (3,876 pairs). • Used GPT-4 Turbo and expert feedback to ensure informative and empathetic answers. • Built a knowledge graph to detect and reduce AI hallucinations (accuracy: 97.9%). • Introduced pseudo-scoring to evaluate and regenerate low-quality AI responses. • Pilot-tested QA chatbot achieved high survivor satisfaction and reusability intent. Despite improvements in survival rates, childhood cancer survivors in South Korea still face significant challenges in accessing the psychological and informational support they need. To address these challenges, we developed the Korean Childhood Cancer Survivor Question-Answer (KCCSQA) dataset in which contains 3,876 question–answer pairs. The questions were sourced from websites, academic articles, and an online survey, where 119 childhood cancer survivors contributed 1,283 questions. We used GPT-4 Turbo to generate the responses, followed by an expert evaluation by 11 specialists to ensure factual accuracy, complementarity, comprehensibility, and empathy. The overall quality of the GPT-generated responses was rated 4.98 out of 6, indicating a high level of quality. To enhance the dataset, we integrated a relational knowledge graph to mitigate hallucinations in the AI-generated answers, achieving a performance of 0.979 in hallucination detection. Additionally, a pseudo-scoring system was implemented for continuous quality assessment. The dataset’s effectiveness was evaluated through a pilot study involving 14 childhood cancer survivors, who interacted with a retrieval-based QA system using a single-turn chatbot format. The mean satisfaction rating was 4.36 on a 6-point Likert scale, and all participants expressed a willingness to use the system again.