Accurate urban flash flood forecasting relies on well-spatialized rainfall data distribution. This study introduces and utilizes the TTI-HydroMet dataset, a publicly available and unique collection for the Tamanduateí River Watershed, in Sao Paulo (Brazil). The dataset includes rainfall measurements from 23 rain gauge stations, stage observations from a hydrological gauge near the outlet, and quantitative precipitation estimates at 1-km radar resolution, accumulated in 10-minute precipitation fields over 10 years. The weather radar data presents missing values for only 0.3 % of timestamps during rainfall events observed by rain gauges. The Spearman correlation coefficient between weather radar and rain gauges varies from 0.675 (full period) to 0.949 (a specific event). It was used to assess the predictive capacity of Machine Learning (ML) hydrological models trained on accumulated rainfall data from rain gauges and estimated by a weather radar. Using an advanced cross-validation framework, two representative algorithms (LinearSVR and XGBRegressor) were tested across different rainfall source configurations and showed strong performance at lead times up to 120 minutes. The Nash–Sutcliffe Efficiency index ranges from 0.781 to 0.996. The statistically comparable performance of ML models driven by radar and rain gauge rainfall indicates that radar-based ML approaches can represent a viable alternative for short-term stage forecasting in regions lacking rain gauge networks.