Indonesian Sentiment Dataset with Rating-Derived Labels, Preprocessing Variants and a Human-Annotated Subset for M-Pajak Google Play Reviews

Indonesian-language user reviews of M-Pajak, the mobile tax-service app operated by Indonesia's tax authority (Direktorat Jenderal Pajak), collected from Google Play in April-May 2026 (6,986 reviews). Contents:mpajak_reviews.csv: each review in its original form and in four cumulatively preprocessed forms (case folding with symbol removal; slang normalisation; NLTK stopword removal; Sastrawi stemming), its star rating, a sentiment label derived from the rating (1-2 stars negative, 3 neutral, 4-5 positive), a stratified 80/10/10 split, test-fold numbers for two repetitions of stratified five-fold cross-validation, and flags for repeated texts and encoding artefacts. mpajak_cv_splits.csv: the train, validation or test role of every review in each of the ten cross-validation folds. mpajak_annotated_subset.csv: 400 reviews, sampled by rating label with the neutral class oversampled, that two annotators labelled independently from the text alone without seeing the rating. It holds both annotators' labels, the adjudicated final label with its rationale, and sampling weights for estimates about the full dataset. data_dictionary.csv, the Indonesian annotation guideline, README.md (full documentation), LICENSE.txt and SHA256SUMS.txt. No reviewer names, review identifiers or timestamps are included. The dataset is shared for non-commercial research and teaching under CC BY-NC 4.0; the review texts remain the work of their authors.