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 about 375,000 binary and ternary structures, enumerating all possible unit cells with different symmetries (BCC, FCC, and HCP) and different number of atoms.