A dataset of human and AI-generated rubric evaluations for formative programming assessment.

Formative assessment in education is often constrained by the time required for rubric design and manual evaluation, limiting the frequency of feedback provided to students. The HARMOGEN-R dataset was created to support research on automated rubric generation and evaluation using artificial intelligence under teacher supervision. It contains evaluations from a Data Structures and Algorithms programming course in which 770 student responses were assessed with five rubrics: one human-created and four generated by large language models (LLMs) using reasoning-enhanced structured and free-form approaches. Each response was evaluated by both human instructors and LLMs, with complete reasoning traces and generated feedback reports recorded to document assessment logic and communication with learners. The dataset also incorporates a variance-based quality control system that flags evaluations exceeding predefined consistency thresholds. This parallel evaluation design enables comparisons of rubric performance, evaluator agreement, and assessment reliability. The dataset is intended for reuse in studies of rubric generation, feedback quality, and automated formative evaluation.