ChiralSPE: a symmetry-audit benchmark for molecular foundation potentials (DFT energies + forces for enantiomer pairs)

ChiralSPE is a benchmark dataset for trustworthiness under physical symmetry constraints in modern molecular foundation models / interatomic potentials. It targets a simple but strict requirement: for an isolated molecule under parity-symmetric electronic structure theory (standard DFT labels), a structure X and its mirror image R·X must be isoenergetic, and their forces must transform covariantly under reflection. This makes chirality an unusually sharp “unit test” for whether a model’s learned potential energy surface respects fundamental invariances—beyond average MAE on random test sets. This release provides paired enantiomers with OMol25-aligned DFT labels (total energy + per-atom forces). Dataset size and layout The archive contains 46,865 files totaling 371.30 MiB, organized as one geometry file and one force file per structure, plus three canonical index/manifest artifacts at the root. 23,430 structures are present (i.e., 11,715 enantiomer pairs). File formats and schemas 1) energy_forces_chiral.jsonl (recommended “single-file” loader) A line-delimited JSON file with one record per structure (23,430 lines). It contains both energies and forces plus pair metadata, and pointers to per-structure files. Observed keys include (not exhaustive):index, pair_index, enantiomer_rank, enantiomer_of, is_mirrored, mirror_axis, chirality, chiral_centers, e_tot_Ha, e_tot_eV, forces, n_atoms, charge, multiplicity, plus CMRT metadata (Literature, RT, Column, Speed, i-PrOH_proportion) and file pointers (forces_file, etc.). 2) energy_index_chiral.csv (tabular index for analysis / joins) A 23,430-row CSV providing the core numeric labels and pointers. Columns include e_tot_Ha, e_tot_eV, forces_file, record_file, pair_index, enantiomer_rank, SMILES, and other run metadata such as timestamp, status, t_grad_s, q_mulliken_sum, etc. 3) manifest_energy_force_chiral.csv (canonical mapping from indices → files + pair metadata) A 23,430-row CSV used as a ground-truth manifest for reconstructing the dataset from per-structure files. Columns include:index, pair_index, enantiomer_rank, enantiomer_of, SMILES, chirality, chiral_centers, is_mirrored, mirror_axis, xyz_path, plus provenance fields (energy/force source folders, variants). geometries/<index>.xyz (one file per structure) Standard XYZ with a metadata-rich comment line. Sample comment fields include: index=<int> (dataset index) smiles=<SMILES> pair_index=<int> and rank=<0|1> chir=<R|S> optional mirror markers such as mirrored axis=0 ff=MMFF94s and a logged energy=<float> (pipeline trace) forces/<index>.json (one file per structure) JSON dict storing the per-atom force array (Nx3) and metadata. It contains the same pair/chirality/CMRT fields used for joins (e.g., pair_index, enantiomer_rank, chirality, chiral_centers, Literature, RT, etc.).