A Correction-Aware Synthetic Dataset for Training Language Models in Academic-Regulation Question Answering

We construct 9,958 examples for academic regulation question answering, pairing draft answers with review feedback and final answers under PASS, FAIL, and OOS labels. Preliminary experiments show that correction performance depends on the quality of the review signal: supplying ground-truth feedback enables effective correction, while errors in learned feedback limit the combined system.

Paper Encouragement Award · 장려논문상 — Korean Association of Data Science, June 19, 2026.