I am trying to reproduce a machine learning model that has been developed here. As one of the datasets, they use single-domain proteins with full length 3D-structures solved. Since I don't have a background in biology, I am not sure how to filter the PDB for these kind of proteins. Any help would be appreciated.
If you want to reproduce the model in that paper, you should email the corresponding author asking for the exact set of PDB IDs/chain IDs/residue ranges they used. There is too much flexibility in the criteria as given to get the exact set any other way.
If you want to assemble a similar set for your own work, the RCSB PDB advanced search is very powerful and would let you filter e.g. by max sequence length or for only monomers.
However the PDB doesn't deal with domains (regions that evolve and fold independently) directly. A PDB chain may contain part of a domain, one domain, or multiple domains.
As a side issue, please remember to remove overlap of homologous proteins between the training and test sets. People entering the field often neglect to do that. See for example this comment piece.