Why was the dataset created? To enable cross-jurisdiction evaluation of free-text 311 service-request classification. No public benchmark combined citizen-authored 311 free text from more than two US cities with a shared label space and fixed splits; MC311 fills that gap and supports the study of domain shift, taxonomy non-comparability, and ambiguity-aware evaluation.
Who created it and funded it? The paper's authors, using public municipal open data; no external funding is associated with the dataset assembly.
What do the instances represent? Each instance is one citizen 311 service request: a free-text description, the city's native service category, and the harmonized 14-class label. About 156,000 content instances span seven cities (Baton Rouge, Bloomington IN, Richmond VA, Auburn WA, Gainesville FL, Honolulu, San Francisco).
How many instances and what is the split? ~156k content rows after harmonization, administrative-category removal, and an informativeness filter. A frozen, seeded, stratified test set holds up to 500 requests per city (3,502 total); the remainder is training. A deduplicated split and, for the four cities with timestamps, a temporal split are also released.
Is any information missing? Yes. Three cities lack usable request timestamps in the collected fields; San Francisco free text is available only via its Open311 API (its bulk export is placeholder-filled). These limitations are documented in the paper.
Are there labels? The label is the harmonized civic class derived from the city's native category by an auditable rule mapping (545 native categories to 14 classes). Labels are the cities' own routing categories and carry measured noise (about 14% not text-entailed per an LLM judge).
Does it contain sensitive or confidential data? The text is public record but contains personal data: street addresses (retained as intrinsic to a request), and, before scrubbing, phone numbers and emails (removed). The Homelessness class concerns a vulnerable population.
How was the data acquired? Programmatically, from each city's Socrata open-data portal or
Open311 API, after a verification pass identifying which portals expose citizen free text. Retrieval scripts
(collect_311.py) and the exact dataset identifiers are released.
Over what timeframe? The requests span each city's available open-data history; collection was performed in August 2026. Exact per-city date ranges are recoverable from the released rows where timestamps exist.
Were people involved? No human annotation was used to build the labels; they derive from the cities' own categories via the released rule mapping. Human validation of the mapping and of label ambiguity is identified as future work.
What preprocessing was done? Native categories were mapped to 14 classes; administrative and routing-only categories were excluded; an informativeness filter removed call-center shorthand and stubs (~8% of rows); phone numbers and email addresses were scrubbed. The raw per-city dumps are regenerable from the released collection scripts.
Is the raw data available? The harmonized content data, the mapping file, the frozen and robustness splits, and the collection scripts are released; the very large raw per-city dumps are regenerable rather than shipped.
What is the dataset for? Cross-jurisdiction and in-city text classification, ambiguity-aware evaluation, taxonomy-harmonization studies, and domain-generalization research on administrative text.
What should it not be used for? Individual-level surveillance, enforcement targeting, or any re-identification of reporters. Because labels reflect each city's operational conventions and reporting is a biased signal of need, the dataset should not be read as a neutral measure of citizen demand.
How is it distributed and licensed? Via the public repository github.com/ApartsinProjects/311. Code is MIT; the data derives from municipal open-data portals under their respective public/open licenses (for example San Francisco's Open311 under PDDL), listed per source.
Who maintains it and how are updates handled? The authors. Corrections and versioned updates are tracked in the repository; every reported number regenerates from the released predictions and scripts.