To train DiffPhore, we developed two 3D ligand-pharmacophore datasets, LigPhoreSet and CpxPhoreSet, which incorporate pharmacophore models derived from ligand conformations and protein-ligand complex structures, respectively. These datasets encompass a broad range of chemical and pharmacophore diversity. LigPhoreSet and CpxPhoreSet include both perfectly and imperfectly matched samples, providing a robust training framework for DiffPhore to learn diverse binding scenarios.