Recod.ai LLM authorship dataset

The dataset developed as part of this Scientific Initiation project is integrated into the Horus Project, conducted by the Pattern Recognition and Image Processing Laboratory (Recod.ai) at the University of Campinas (UNICAMP). The main objective of this dataset is to support research on text authorship attribution, focusing on the analysis and differentiation between human-written texts and those generated by large language models (LLMs). The dataset consists of texts produced by multiple human authors and two language models — ChatGPT and Mistral — covering a wide range of topics, styles, and text lengths, with a particular focus on Twitter posts. Each sample includes the text itself, along with the identifier of the author or generating model. A two-stage feature extraction process was performed to facilitate subsequent analysis, and the extracted features are also included as part of the dataset. In Phase 1, the extracted features include semantic, structural, syntactic, and stylistic characteristics. In Phase 2, these same categories are preserved, but LLM-specific indicators are additionally incorporated, enabling a deeper analysis of the writing patterns of generative models. Furthermore, the dataset also includes a subset highlighting the most influential features identified in each phase, both per model and in a general overview, supporting a more detailed interpretation of the results and the development of robust authorship attribution systems.