Machine-learning approaches for soil moisture modelling are expanding rapidly, but their comparison remains difficult because studies often use different target data, predictors, preprocessing methods, and evaluation strategies. Here we present SoMoBench, a quasi-global benchmark dataset that provides standardised predictor inputs and soil moisture targets for data-driven modelling. The targets are derived from in situ observations provided by the International Soil Moisture Network and Brazil’s CEMADEN network. Sub-daily measurements are aggregated to daily values and assigned to a common 0.25° grid for three nominal soil layers (0–10, 10–30, and 30–50 cm). To reduce site-dependent offsets while retaining observation-derived temporal variability, each station series is adjusted to the mean and standard deviation of collocated ERA5 soil moisture. Target series are constructed only for grid cells containing at least one eligible station, and no spatial interpolation is applied. For each grid cell and layer, the longest station record is used as the primary series, and temporal gaps are filled with validated secondary-station record from the same grid cell when available. The current release spans 2001–2024 and includes 1,445 grid cells with at least one target layer, along with daily ERA5 meteorological predictors for 2000–2024. We characterise the spatial and temporal properties of the targets and illustrate their use with climatology-based and Ridge regression models. The release also includes a Python example notebook and baseline performance results to support reproducibility. SoMoBench is designed as a reusable community resource for observation-informed benchmarking and model intercomparison in soil moisture modelling. The dataset is available for peer review through Zenodo; access information is provided in the Data availability section.