The reference database is composed by 277 hours of Global Horizontal Irradiance (GHI) measurements at minute resolution from the Carpentras Baseline Surface Radiation Network (BSRN) station during the years 2012 and 2013. Each hourly interval was manually classified into one of eight variability classes, ranging from clear-sky to overcast conditions. The dataset was developed to support the reproducibility of the variability classification method presented in the accompanying publication (https://doi.org/10.1016/j.renene.2026.125387). The Carpentras reference database is made available as a NetCDF file with the following file structure groups: Minute_data: contains 1-minute irradiance measurements (GHI, Direct Normal Irradiance - DNI, and Diffuse Horizontal Irradiance - DHI) and corresponding clear-sky estimates from the McClear model, together with minute timestamps and minute class labels. Hourly_variability_indices: contains all variability indices computed for each hourly window, hourly classes labels, and timestamps. The labeling of the timestamps is at the end of the time interval (e.g. the timestamp at 10h00 considers data between 09:00 and 09:59 for the variability indices calculations). Variability_indices_probability: provides the variability indices and its KDE-based probability density estimates (1024 grid points) in each class, enabling reproduction of the probabilistic classification step. Satellite imagery from MSG/APOLLO_NG (Schroedter-Homscheidt et al., 2022) and MODIS (Platnick et al., 2014) was used to visually associate variability classes with typical cloud conditions. These images are provided as PDF files for each variability class (from Class 1 to Class 8). Further details about the interpretation of these images are found in the Supplementary Material of the accompanying paper. The code repository with the classification algorithm will be linked here once publicly released.