The machine learning dataset used in the work of "On the detection of stellar wakes in the Milky Way: a deep learning approach". The .tar contains .h5 files for four different subhalo target masses such that each file is generated from a unique simulation seed. Each file contains data from a single simulation run and is comprised of 100 samples which are generated by sampling 1% of all star particles from a particular simulation snapshot. Prior to sampling particles, we split the simulation box into three equal slices in the Z-coordinate and bin the data onto the X-Y plane with 32 bins on each axis. Each sample is then characterised by a 2D histogram of shape (32, 32, 12) with the last dimension containing the training features which are the overdensity, mean X-Y velocity, X-Y velocity dispersion and divergence in X-Y for the three Z-slices.