L-RISK: A reproducible annotation and evaluation method for learning-risk in generative AI for intelligent tutoring systems

Abstract

This dataset supports research on evaluating generative AI (GenAI) hallucinations in educational contexts. It accompanies the L-RISK framework, a standardized methodology for annotating and assessing hallucinations based on their impact on learners’ mental models rather than solely on factual correctness. Grounded in Mental Model Theory, Conceptual Change Theory, and Cognitive Load Theory, the dataset operationalizes a pedagogically informed error taxonomy, an ordinal learning-risk severity scale, and task- and prompt-aware evaluation procedures. The materials include annotated examples from a pilot application in supply chain education and are intended to support reproducible research on learning-oriented hallucination evaluation and responsible GenAI deployment in education.