Robust robotic grasping of objects has broad industrial applications. The reliability of data-driven grasping methods is influenced by the variability of object shapes encountered during training. Most existing objects datasets suffer from human selection bias, lack variability, or are non-reproducible. This paper presents a physically reproducible 3D-printable object dataset for training and evaluating grasping algorithms. It contains exact 3D meshes of 50 objects for simulation and printing purposes. The various objects in the dataset were found using the MAP-Elites algorithm, optimising the variability of objects according to two grasping metrics. We used a Variational AutoEncoder (VAE) as a generative model for voxelgrid object models, which were then converted to meshes and simplified using Volumetric Hierarchical Approximate Convex Decomposition (V-HACD). The dataset is publicly available online, and can be ordered from any 3d-printing service according to given specifications. We hope it will become a standard benchmarking dataset for the robotic grasping community.