Raw student programming-exam submissions and instructor grades (Kotlin and Dart) — restricted-access dataset

This dataset contains 961 real student programming-exam submissions in Kotlin and Dart, written in French, together with the grades assigned by a human instructor. Each record includes the examination question, the student's submitted code, the question's maximum score, and the instructor's awarded and normalised scores. The data supports the study "Open-weight large language models match proprietary systems for out-of-the-box grading of programming exams, but neither approaches human inter-rater reliability," which benchmarks 24 large language models as automated graders against the human instructor and a second human rater. Access is restricted because the records contain student-authored code collected under institutional consent that does not permit public redistribution. A scores-only version — sufficient to reproduce all published statistical results — is openly available in the companion record. Access to this restricted record is granted to editors, reviewers, and researchers on reasonable request, subject to institutional approval.