# UltraMS SpectraVerse dataset

Version **1.0.0** · 2026-11-04 · license **CC BY 4.0** · Zenodo DOI **10.5281/zenodo.23964105**
(see [NOTICE.md](NOTICE.md))

A frozen, molecule-disjoint conversion of the
[SpectraVerse](https://doi.org/10.5281/zenodo.17870921) tandem mass spectral library,
released by UltraMS. **This is a data resource, not a standalone benchmark.** The
manuscript uses it in two roles:

1. as the spectrum source of the negative-ion structure-family evaluation (Figure 2c:
   661 molecule-spectrum pairs after negative-ion selection and exclusion of molecules
   present in the MoNA adaptation data set), and
2. as one of the three structure-annotated sources (together with MassSpecGym and MSnLib)
   for the contrastively enhanced UltraMS training, the chemical-class evaluation and the
   UltraAtlas reference spectra.

## 1. The table

| item | value |
| --- | --- |
| file | `Spectraverse.csv` |
| rows | 464,346 spectra |
| bytes | 707,447,509 |
| SHA-256 | `06e084501e91f92cc642c3d60de2863266ff3d0db958354a011e32851bef3976` |
| columns | `mzs, intensities, smiles, precursor_mz, adduct, collision_energy, fold` |
| peak format | comma-separated strings, one spectrum per row |

The `fold` column is the frozen train / validation / test assignment:

| fold | spectra | unique exact SMILES |
| --- | ---: | ---: |
| train | 370,362 | 33,846 |
| validation | 47,825 | 4,231 |
| test | 46,159 | 4,231 |
| total | 464,346 | 42,308 |

## 2. Split definition

- The split unit is the **exact SMILES string** (molecule level): every spectrum of a given
  SMILES string falls in exactly one fold, and no exact SMILES occurs in more than one fold
  (0 crossing strings out of 42,308).
- The partition is 8:1:1 at the molecule (exact-SMILES) level: 33,846 / 4,231 / 4,231
  unique SMILES.
- **This split is not formula-disjoint.** Structural isomers share a molecular formula, and
  molecular formulas do recur across folds in this library. Downstream work must describe
  the partition as molecule/SMILES-disjoint, not as a formula-disjoint partition.

The frozen assignment is distributed as
[`splits/spectraverse_compound_split.csv`](splits/spectraverse_compound_split.csv)
(42,308 rows: `smiles, fold, n_spectra`), so the data can be used without reconstructing
the assignment.

## 3. Usage

```python
import pandas as pd

base = "https://huggingface.co/datasets/dsadd4/UltraMS-Spectraverse-Dataset/resolve/v1.0.0/"
spectra = pd.read_csv(base + "Spectraverse.csv")
split = pd.read_csv(base + "splits/spectraverse_compound_split.csv")

test = spectra[spectra.fold == "test"]
assert split.groupby("smiles").fold.nunique().max() == 1
```

The UltraMS repository consumes this table in
`benchmarks/spectral_properties/prepare_library_spectra.py --dataset spectraverse`
(published size/SHA-256 and fold counts are verified before use).

## 4. Provenance and licence

The spectra are the public SpectraVerse library (SpectraVerse v1.0.1, Zenodo
[10.5281/zenodo.17870921](https://doi.org/10.5281/zenodo.17870921), CC BY 4.0;
Gupta, Qiang, Chung, Herbst and Skinnider). UltraMS contributed the CSV
representation, the frozen molecule-level train / validation / test split and its
distribution. See [NOTICE.md](NOTICE.md) for the full attribution.

## 5. Files

```text
README.md, NOTICE.md, SHA256SUMS, VERIFY.md, stats.json
Spectraverse.csv                          (hosted with the release; not duplicated in the archive)
splits/spectraverse_compound_split.csv(.gz)
hf_dataset_card.md, hf_upload_commands.sh, zenodo_metadata.json
```