This dataset comprises predicted protein structures from the paper "Ultrafast end-to-end protein structure prediction enables high-throughput exploration of uncharacterized proteins". Structures were predicted using DMPfold2.BFD_1.3M.hdf5 contains all the models from the set of 1.3M that were generated. The models can be retrieved from this file using the provided hdf5_extract.py script and the list of IDs in bfdfold_1.3M_target_ids.csv.Also provided are tarballs of the models and sequence alignments for the 5193 Pfam families modelled in the paper, as well as for the set of 255 Pfams with released structures used for comparisons against DMPfold1 and C-I-TASSER.train_data.tar.bz2 contains the data used to train the DMPfold2 neural network. Further scripts and instructions are available on the associated GitHub page: https://github.com/psipred/DMPfold2