One of the difficulties when testing algorithms for information retrieval, is having benchmark datasets to determine their effectiveness, since for several indicators used in this discipline, it is necessary to know in advance results for a query given and that the same results are evaluated by experts on the subject under investigation. This paper describes the collaborative implementation of a dataset created from the metadata and full-text of articles from nine areas of knowledge, with the participation of fourteen researchers at master or doctoral level, supported with the bibliographical reference manager Mendeley. Final dataset with metadata and full text, consists of 1678 papers and the result of 38 queries of two, three and four words.