This dataset contains graph representations of 73,821 unique transition metal complexes from the tmQMg dataset ready for use in Δ-machine learning frameworks. The graph representations were generated with the HyDGL Python package according to the u-NatQG specification and are based on electronic structure data at two different levels of theory: Geometry optimization: GFN2-xTB // Single-point refinement: LSDA/LANL2DZ Geometry optimization: GFN2-xTB // Single-point refinement: PBE0-D3BJ/def2-TZVP The corresponding benchmark graphs obtained in previous work are also supplied.