APPRISE: A Persona-Conditioned Dataset of Synthetic App Reviews Paired with GitHub Issues

Mobile app reviews and software issue trackers describe overlapping maintenance concerns from different perspectives, but large-scale labeled datasets linking the two are scarce. Existing resources either provide review classifications without paired issues, smaller-scale review–bug matches without persona diversity, or app-store data without the contrastive structure required to train modern retrievers. We present APPRISE, a publicly released dataset of 13,579 persona-conditioned synthetic user reviews paired with 9,435 real GitHub issues across four open-source Android applications: Brave Browser, Signal Android, AnkiDroid, and K-9 Mail. The synthetic reviews are generated by a teacher large language model conditioned on a 10-persona taxonomy that captures the stylistic diversity of app-store feedback. APPRISE further provides 73,984 hard-negative contrastive triplets mined with a dual within-app and cross-app strategy, complete issue metadata supporting status- and closure-reason analyses, and a stratified 400-sample LLM-based source-alignment quality audit. The dataset is released under the MIT License with reproduction scripts, prompt templates, persona definitions, validation scripts, mining scripts, and a datasheet. Full dataset DOI:https://doi.org/10.5281/zenodo.20091031 Companion GitHub repository:https://github.com/SoftALL/APPRISE Hugging Face triplets dataset:https://huggingface.co/datasets/SoftALL/APPRISE-triplets Contact:Ogtay Hasanov — [e-mail] Ezzini — [e-mail]