eKlausur Foundation Models Benchmark 2026

This dataset accompanies a study on automatic recognition of handwritten letter answers in printed exams. It contains anonymised answer-table crops from 61 exams, ground-truth answer information, model output files, and aggregate evaluation tables comparing YOLO-based detectors with vision-language foundation models. The benchmark contains 742 scanned answer-table pages and 3141 evaluated answer positions. It supports the analysis of overall recognition accuracy as well as fairness-relevant false-negative and false-positive behaviour in automated exam grading. Names and matriculation numbers are not included. The released images are cropped answer-table regions rather than full exam scans.