POSCOMP Dataset

This is a curated dataset of past POSCOMP exam multiple-choice questions, up to 2024. Description This dataset contains 1,340 multiple-choice questions (MCQs) sourced from historical editions of the Brazilian National Exam for Admission to Graduate Studies in Computing (POSCOMP). It was originally created to support research in automated question generation and educational assessment, and its process of creation was documented in a paper (titled Smarter Questions, Smaller Models: RAG-Enhanced Multiple-Choice Question Generation for POSCOMP) to the published in the 2025 Brazilian Symposium on Computers in Education. The questions are drawn from 21 past POSCOMP exams, covering the period from 2002 to 2024. The dataset is comprehensive, omitting only 130 questions from the original exams that were dependent on images. The 1,340 questions are categorized into POSCOMP's three main areas of knowledge: Mathematics: 391 questions Computational Foundations: 507 questions Computational Technology: 442 questions As no publicly available, structured dataset of POSCOMP exams existed, this one was built from scratch. The process involved extraction from text content from historical exam PDF files, as well as automatic and extensive manual structuring. Each question in the dataset is accompanied by a rich set of metadata11: year: The year the exam was administered. number: The original question number in the exam. stem: The text of the question. options: A list containing all answer choices (distractors and the correct answer). key: The official correct answer. knowledge_area: The broad area (MT, CF, or CT). area: A more specific thematic division within the knowledge area. subarea: A granular topic classification within the area. The area and subarea fields were generated using an LLM-based approach (Llama 3.3-70B) with Self-Consistency and Self-Refine techniques, and were then validated by researchers, reacihng strong agreements.