HFlow320 is a compact synthetic event-based optical-flow dataset focused on human motion at 320x320 resolution. It was generated with a modified BlinkSim/Blender pipeline using MakeHuman characters, motion clips from the Carnegie Mellon University Graphics Lab Motion Capture Database, and varied HDRI backgrounds. High-frame-rate renders were converted into event streams with DVS-Voltmeter, while low-frame-rate renders provide optical-flow labels and validity masks. The dataset is designed as a software-focused benchmark for rapid event-flow model development, especially for compact and resource-conscious baselines. The dataset contains 144 animated clips split into 100 training clips, 20 validation clips, and 24 test clips. These correspond to 22,160 training, 4,466 validation, and 5,563 test optical-flow frames, or approximately 742 s, 150 s, and 186 s of motion respectively. The motion set includes actions such as walking, running, jumping, dancing, gestures, and daily activities. Camera viewpoints and backgrounds are randomized while keeping the scene centered on a single human actor. The test split uses a held-out human model and disjoint animations to reduce appearance and motion leakage. Each sample includes event data, optical-flow labels, validity masks, and optional RGB frames for visualization or multimodal use. Event streams and flow labels are stored in HDF5 format, while RGB frames are provided separately as PNG images. Evaluation should report the event window used for prediction and apply the provided validity masks consistently. The primary reported metrics are masked endpoint error and masked average angular error. HFlow320 is intended to complement larger driving- and general-scene event-flow datasets by providing a smaller, controlled benchmark centered on localized, non-rigid human motion. Its main purpose is to support quick training, ablation studies, and reproducible comparison of event-based optical-flow methods, including lightweight software baselines such as EmFlow. It is not meant to replace broad real-world benchmarks such as MVSEC or DSEC, but to provide a tractable setting for studying compact model behavior under articulated human motion.