Wireless World Model - Beam Prediction Dataset

Dataset Overview This dataset provides beam predictor training data for the study “Wireless World Models for Future AI-Native Networks”. It contains simulated channel state information (CSI) matrices sampled across multiple time steps, enabling research on learning-based wireless channel modeling and prediction. The dataset corresponds to the experiments reported in the associated paper. The data are generated in 4 realistic urban environments reconstructed from geographic information: the Forbidden City of Beijing, the Central Business District of Beijing, Place de l’Étoile (Paris), and an urban area in Munich. For each environment, multiple datasets are collected using base stations deployed at different locations. There are 2 central frequency settings, 2.6GHz and 6.6GHz, with samples matching the same UE position. Within each scenario, channel realizations are densely sampled over time and synchronized with the spatial coordinates of user equipment (UE), which are provided in the accompanying position (pos) files. This design yields time-continuous CSI–position pairs rather than independent channel snapshots. Typical UE velocities considered in the simulations include 5 km/h, 30 km/h, and 60 km/h. Data Generation Method The CSI data were generated using the Sionna ray-tracing framework, which provides physically consistent wireless channel simulation. The ray-tracing environment models realistic urban propagation conditions including reflections, diffractions, and multi-path propagation. File Organization dataset/ ├── etoile_26.zip │ ├── H/ │ │ ├── samples_H_etoile_bs0_5kmh_2048_seed12445.npy │ │ ├── ... │ │ └── samples_H_etoile_bs1_60kmh_2048_seed68002.npy │ └── pos/ │ ├── samples_pos_etoile_bs0_5kmh_2048_seed12445.npy │ ├── ... │ └── samples_pos_etoile_bs1_60kmh_2048_seed68002.npy ├── cbd_26.zip ├── ... └── etoile_66.zip Data Format Each dataset file is stored in NumPy (.npy) format. The CSI data file samples_H_xxx_bsx_xxkmh_2048_seedxxxxxxxxxx.npy has the following shape: (2048, 2, 16, 32, 32) where 2048 : number of channel samples2 : real and imaginary components16 : temporal snapshots32 : base station antennas features32 : user equipment antennas and frequency subbands features The real and imaginary parts are stored separately along the second dimension. The position data file samples_pos_xxx_bsx_xxkmh_2048_seedxxxxxxxxxx.npy has the shape: (2048,) The file contains 2048 dictionary entries, where each dictionary corresponds to one CSI sample. Each dictionary includes the following keys: ue_traj: the trajectory of the user equipment (UE), storing the UE positions at 16 time steps. cityname: the name of the urban scenario in which the data are generated. bs_pos: the spatial coordinates of the base station used in the simulation. Usage Download and unzip all data files within etoile_26.zip, cbd_26.zip, munich_26.zip and forbidden_city_26.zip into src/datasets/processed_data/output1, unzip all data files within etoile_66.zip, cbd_66.zip, munich_66.zip and forbidden_city_66.zip into src/datasets/processed_data/output2. Then run the script python -m app.beam_pred.prepreprocess, python -m app.beam_pred.nomalize_dataset and python -m app.beam_pred.split_dataset to perform dataset preprocessing.