AIriskEval-edu: A New Dataset for Evaluating the Risks of AI-Mediated K-12 Instructional Explanations

The AIriskEval-edu-db2 dataset is used to train and evaluate LLM audit models to perform interpretable risk assessment of K-12 instructional explanations. It contains 1639 explanations for 170 ScienceQA questions, from real human teachers and LLM-simulated teachers with different instructional risks, and 785 entries are annotated with risk localization and descriptions. The experiments compare a local Llama 3.1 8B with proprietary models to test whether, after fine-tuning, it can approach or surpass frontier models while protecting privacy.