A quality-controlled PM2.5 dataset with frozen cross-city evaluation splits for six Central Asian cities

Daily PM2.5 observations from 7 instruments across 6 Central Asian cities (Almaty, Ashgabat, Bishkek, Dushanbe, Khujand, Tashkent), screened by seven pre-registered quality-control rules, deduplicated and timezone-verified. Five instruments are US diplomatic-post BAM/FEM-class reference monitors; two are Clarity Node-S low-cost sensors, labelled as such. Each station-day is paired with satellite retrievals, chemistry-transport forecasts, reanalysis meteorology and static geography, every predictor annotated with a measured acquisition latency. The release contains frozen, checksummed evaluation splits — blocked-temporal with a 240-hour purge gap, and leave-city-out over 6 folds — with a baseline ladder, reference model outputs and row-level predictions, plus the complete pipeline that regenerates every number. The underlying PM2.5 observations are not redistributed: per-provider OpenAQ licence terms differ and have not been verified for redistribution. Split definitions and reference results are self-contained and verifiable offline. Benchmark version 1.1.0. Accompanies a Scientific Data Data Descriptor.