This dataset was generated using the following active learning scheme: 1) candidate structures were relaxed by a partially-trained MTP model, 2) structures for which the MTP had to perform extrapolation were passed to DFT to be re-computed, 3) the MTP was retrained, including the structures that were re-computed with DFT, 4) steps 1-3 were repeated until the MTP no longer extrapolated on any of the original candidate structures. The original candidate structures for this dataset included 40,000 unrelaxed configurations with BCC, FCC, and HCP lattices.