Edge-CCIIL benchmark: per-run forecasts, results tables, figures and CAMS/ERA5 inputs for multi-site air-quality forecasting (v1.0.0)

Data and results accompanying the article Compressed serverless cyclic learning for multi-site air-quality forecasting across heterogeneous monitoring networks: a reproducible benchmark (submitted to Environmental Modelling & Software). Code: Zenodo software record DOI 10.5281/zenodo.23806727. Contents: 729 per-run JSON files (504 final runs and supplementary runs; metrics, per-site errors, learning curves, transmitted bytes and seed-wise test forecasts), validation-only tuning runs, all result tables (CSV), statistical tests (Wilcoxon/Holm, Diebold-Mariano/HLN, Friedman), edge emulation outputs, preprocessing audit outputs, all figures (PDF and 600 dpi PNG), and the JTB grid-cell inputs (CAMS global composition forecasts and ERA5 reanalysis retrieved through Open-Meteo, CC BY 4.0; NASA POWER/MERRA-2). Not included because of third-party terms: the ISPU DKI Jakarta daily index file and NOAA ISD observations; download instructions and checksums are given in the code repository. Contains modified Copernicus Atmosphere Monitoring Service and Copernicus Climate Change Service information (2022-2026).