EduEVAL-DB: A role-based dataset for pedagogical risk assessment of educational explanations
EduEVAL-DB: A Role-Based Dataset for Pedagogical Risk Evaluation in Educational Explanations
AI summary
EduEVAL-DB is a role-based dataset for pedagogical risk assessment of educational explanations, containing 854 explanations corresponding to 139 questions from ScienceQA. Each question has 1 human teacher explanation and 6 generated by LLM-simulated teacher roles, covering K-12 science, language, and social science, and is annotated with binary labels via a five-dimensional pedagogical risk rubric with expert review. Preliminary validation compares Gemini 2.5 Pro with a local Llama 3.1 8B and tests risk detection after supervised fine-tuning.