A Community Baseline and Benchmark for Machine-Learning Satellite Precipitation Retrievals Authors: Simon Pfreundschuh (Colorado State University); Malarvizhi Arulraj (University of Maryland); Yashraj Upase (Indian Institute of Technology Hyderabad); Veljko Petković (University of Maryland); Jackson Tan (University of Maryland); Lisa Milani (University of Maryland) Corresponding author email: simon.pfreundschuh@colostate.edu Abstract: Recent machine-learning (ML) approaches for satellite precipitation retrieval show promising performance and are increasingly considered for operational use. However, differences in training data, validation strategies, and sensor characteristics make it difficult to compare published methods or to distinguish genuine methodological advances from dataset-dependent accuracy variations. To address this limitation, the International Precipitation Working Group (IPWG) has developed a standardized benchmark for precipitation detection and estimation from satellite observations. The dataset combines multi-sensor satellite measurements, ancillary information, and gauge-corrected ground-radar reference data, and provides a consistent evaluation protocol across diverse climate regimes. Building on this framework, we introduce a publicly available baseline retrieval and systematically assess key design choices in neural-network-based precipitation retrievals. Across architectures, training strategies, and model complexities, we find that incorporating spatial context provides the largest performance gains, whereas specific architectural refinements yield comparatively small improvements. The resulting baseline achieves robust performance across sensor types for both precipitation detection and quantitative estimation. By providing both a uniform benchmark and an open baseline model, this work enables reproducible comparison of ML retrievals and clarifies which methodological developments lead to meaningful improvements. The framework establishes a reference point for future algorithm development and supports more transparent evaluation of machine-learning precipitation retrievals. Within this benchmark, performance gains are dominated by the use of additional observational context rather than architectural complexity, suggesting that future improvements will primarily arise from better exploitation of available information rather than further model refinement.