The Stack Overflow Recommendations Dataset (SORD) – A Large-Scale Curated Dataset of Recommendations Related Stack Overflow Questions, Answers and Comments

Developer discussions particularly on programming related questions answering (Q&A) sites, contain useful information, which, if mined and analysed carefully can be transformed into insightful recommendations for developers about which software to use or prefer over others, matching with one’s requirements for different software development activities. However, there is a long way to go before such a system can be realized. As a first step in this direction, we empirically explored the developers’ discussions on Stack Overflow, one of the most popular Q&A sites among developers, with a particular emphasis on mining software recommendations related insights. We considered the Stack Overflow data dump published in October 2025 containing complete data of the site since 2008. The data extraction process started with the conversion of the gigantic XML files to SQL Server database table records. We then applied a keywords-based filtering approach to identify potential recommendation related queries in questions, and recommendations in answers and comments related to any aspect of software development. The set of keywords comprised of 19 keywords including the term ‘recommend’ along with its 14 synonyms and 4 antonyms. The extracted dataset contains 73.9k (0.31%) questions containing such terms in the question title, 1.1M (4.69%) questions containing them in the question body, 2.2M (6.18%) answers and 1.9M (2.08%) comments because of exact keyword matching. The results further increase in case of substring matching. When enriched with additional metadata e.g. Users, Badges, Votes and Tags (which are available in the Stack Overflow data dump), the raw dataset presented in this paper can become highly useful for the empirical software engineering and machine learning research community for training models and developing recommendation systems for software engineering. The extracted answers and comments can be mined to extract implicit developer preferences from which ratings can be inferred whereas the extracted recommendation related questions with accepted answers can be transformed into a benchmark for retrieval evaluation of software recommendation systems for developers. The dataset is hosted on Figshare using DOI 10.6084/m9.figshare.30948506 and can be accessed via https://doi.org/10.6084/m9.figshare.30948506 or the GitHub repository.