UltraMS SpectraVerse dataset: molecule-level 8:1:1 split of the SpectraVerse tandem mass spectral library

A frozen, molecule-disjoint conversion of the SpectraVerse tandem mass spectral library, released by UltraMS. This is a data resource, not a standalone benchmark. Table. Spectraverse.csv contains 464,346 spectra with columns mzs, intensities, smiles, precursor_mz, adduct, collision_energy, fold (707,447,509 bytes; SHA-256 06e084501e91f92cc642c3d60de2863266ff3d0db958354a011e32851bef3976). Split. Train / validation / test contain 370,362 / 47,825 / 46,159 spectra and 33,846 / 4,231 / 4,231 unique exact SMILES (8:1:1 at the molecule level). The split unit is the exact SMILES string: every spectrum of a SMILES string lies in exactly one fold and no exact SMILES crosses folds (0 of 42,308). The partition is not formula-disjoint: structural isomers share a molecular formula and formulas recur across folds. Use. The manuscript uses the table as the spectrum source of the negative-ion structure-family evaluation (Figure 2c; 661 molecule-spectrum pairs after negative-ion selection and MoNA exclusion) and as one of the three structure-annotated sources (with MassSpecGym and MSnLib) for the contrastively enhanced UltraMS training, chemical-class evaluation and UltraAtlas reference spectra. The frozen assignment is distributed as splits/spectraverse_compound_split.csv (42,308 rows). The spectra are the public SpectraVerse library (v1.0.1, Zenodo 10.5281/zenodo.17870921, CC BY 4.0; Gupta, Qiang, Chung, Herbst and Skinnider). UltraMS contributed the CSV representation, the frozen molecule-level split and the release. Released under CC BY 4.0 with attribution to the SpectraVerse authors.