2026-10-14T06:43:35Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33940903 2026-09-18T22:44:03Z category_28849 item_type_3 month_year_09_2026

Data and code for 'When Retrieval Can Hurt' (Arabic MCQ generation, 2×2 human evaluation) abderrahman azi (25074097) Natural language processing Self-Refinement Modern Standard Arabic (MSA) MCQ exams <p dir="ltr">This deposit holds the research data and analysis code behind a study on automatically writing Arabic multiple-choice questions with large language models. The work tests whether adding retrieved reference passages and an automatic self-revision step improves question quality, using a four-condition experiment across 31 driving-education topics judged by a panel of raters in three stakeholder roles.</p><p><br></p><p dir="ltr">Included here are the model-produced questions for every condition, the panel's anonymised scores (raters labelled by role only), the computed text-similarity and audit measures, and the scripts that run the statistics and build the figures. Retrieved passages appear as file-and-page references rather than as text, and the copyrighted source lessons are not shared; only generated items and derived results are provided.</p> 2026-09-18T22:44:03Z Dataset Dataset 10.6084/m9.figshare.33940903.v1 https://figshare.com/articles/dataset/Data_and_code_for_When_Retrieval_Can_Hurt_Arabic_MCQ_generation_2_2_human_evaluation_/33940903 CC BY 4.0