Hugging Face's logo

Datasets:
dsadd4
/
UltraMS-MSnLib-Benchmark

Tasks:
Other
Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
1M - 10M
Tags:
chemistry
mass-spectrometry
metabolomics
small-molecules
tandem-ms
spectrum-library-search
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
UltraMS-MSnLib-Benchmark / MANIFEST.md
dsadd4's picture
dsadd4
Add the full manifest README alongside the dataset card
df67944 verified
preview code
|
Raw History Blame Contribute Delete
25.1 kB
# UltraMS MSnLib benchmark
Version **1.0.0** · 2026-11-04 · license **CC BY 4.0** (see [NOTICE.md](NOTICE.md))
Two frozen, independent benchmark splits of the [MSnLib](https://doi.org/10.1038/s41592-025-02813-0)
tandem mass spectral library, built and released by UltraMS:
* **Split A — UltraMS MSnLib formula-level split**: the complete library (560,084 spectra,
all adducts) split by molecular formula into train / validation / test
(16,885 / 2,110 / 2,112 formulas, 446,667 / 55,980 / 57,437 spectra). Formula-disjoint.
This is the population for single-spectrum candidate ranking, multiple-spectrum voting,
label-blind grouping and embedding analyses.
* **Split B — strict-adduct molecule-level split**: the strict `[M+H]+` and `[M-H]-`
subsets (447,956 source rows) re-split 70 / 15 / 15 at the level of RDKit canonical
isomeric SMILES, separately per polarity, with frozen queries and libraries for
spectrum-to-spectrum library search. Molecule-disjoint within each polarity.
Both splits are frozen benchmark definitions with published checksums; neither is a
redistribution of the MSnLib spectra.
---
## 1. At a glance
| | Split A (formula-level, all adducts) | Split B `[M+H]+` | Split B `[M-H]-` |
| --- | ---: | ---: | ---: |
| compounds / molecules | — (21,107 formulas) | 41,348 | 21,293 |
| spectra | 560,084 | 301,636 | 140,193 |
| train | 16,885 formulas · 446,667 spectra | 28,944 compounds · 211,745 spectra | 14,907 · 98,401 |
| validation | 2,110 formulas · 55,980 spectra | 6,202 compounds · 44,881 spectra | 3,193 · 20,927 |
| test | 2,112 formulas · 57,437 spectra | 6,202 compounds · 45,010 spectra | 3,193 · 20,865 |
| test task | candidate ranking / voting (57,437 queries) | 6,202 queries vs 295,434 library spectra | 3,193 vs 137,000 |
| queries with an exact-adduct library positive | — | 6,201 | 3,160 |
| validation task | — | 6,202 vs 250,424 | 3,193 vs 116,135 |
Split A test fold: 57,437 spectra, 2,112 formulas, 5,003 distinct MSnLib source SMILES,
4,889 distinct RDKit canonical isomeric SMILES. The **official** multiple-spectrum voting
unit is the MSnLib source SMILES string: 5,002 groups covering 57,436 spectra (canonical
view: 4,888 groups).
Split B split seeds: 42 (`[M+H]+`) / 43 (`[M-H]-`); query seed 119. Split A has no seed:
its fold assignment is frozen in the `fold` column of `MSnLib.csv` and replayed exactly by
`split_A_full_msnlib/build_split_a.py`.
## 2. Disjointness: what is claimed — and what is not
| split | disjointness unit | claimed |
| --- | --- | --- |
| A | molecular formula | every spectrum of a formula lies in exactly one fold; no formula occurs in two folds (0 conflicts out of 21,107) |
| B | RDKit canonical isomeric SMILES, per polarity | the train / validation / test compound sets of each polarity are pairwise disjoint |
**A and B are different objects and must not be combined into a global-disjointness claim.**
About 41 % of split-B compounds are in a *different* split-A fold, and a spectrum can be a
split-A test spectrum while its compound is a split-B train compound. Concrete numbers:
* split B `[M+H]+` test compounds: 654 are split-A test, 4,922 are split-A train, 626 are
split-A validation;
* split B `[M-H]-` test compounds: 303 are split-A test, 2,541 are split-A train, 349 are
split-A validation;
* 31,570 split-A test spectra are `[M+H]+` rows of split B and 13,413 are `[M-H]-` rows.
The full cross-tabulation (compound level and spectra level, per polarity, plus the
relation of the split-A test fold to split B) is in [figures_mapping.md](figures_mapping.md) §2.
Never report a cross-split score as a held-out result: train and evaluate inside one split.
## 3. Package layout
```text
ultrams_msnlib_benchmark_v1/
├── README.md this file
├── figures_mapping.md which manuscript panel uses which split (+ cross-split statistics)
├── VERIFY.md full verification record and reproducible commands
├── NOTICE.md MSnLib attribution, provenance and licensing
├── SHA256SUMS checksums of every file except itself
├── stats.json machine-readable release summary (all counts + file hashes)
├── split_A_full_msnlib/ split A (formula-level, full library)
│ ├── README.md
│ ├── VERIFY.md split-A verification record
│ ├── REPORT.json split-A machine-readable report
│ ├── build_split_a.py replay of split A from MSnLib.csv + hash check
│ ├── formula_split.csv(.gz) 21,107 formulas → fold, n_spectra
│ ├── row_formula_fold.csv(.gz) 560,084 rows → formula, fold
│ ├── test_spectra.csv(.gz) 57,437 test spectra
│ ├── val_spectra.csv(.gz) 55,980 validation spectra
│ ├── test_molecules_voting.csv(.gz) 5,002 official source-SMILES voting units
│ ├── test_molecules_voting_canonical_view.csv(.gz) 4,888 canonical-SMILES alternative view
│ └── msnlib_rt.csv.gz 560,084 row_index → retention time (seconds)
├── splits/ split B compound tables (70/15/15, per polarity)
│ ├── pos_compound_split.csv(.gz) 41,348 compounds
│ └── neg_compound_split.csv(.gz) 21,293 compounds
├── queries/ split B query spectra
│ ├── pos_val_queries.csv(.gz) 6,202
│ ├── pos_test_queries.csv(.gz) 6,202
│ ├── neg_val_queries.csv(.gz) 3,193
│ └── neg_test_queries.csv(.gz) 3,193
├── libraries/ split B library spectra
│ ├── pos_val_library.csv(.gz) 250,424
│ ├── pos_test_library.csv(.gz) 295,434
│ ├── neg_val_library.csv(.gz) 116,135
│ └── neg_test_library.csv(.gz) 137,000
├── protocol/ frozen split-B protocol arrays and manifest
│ ├── manifest_summary.json
│ ├── protocol_{pos,neg}.npz
│ ├── smiles_{pos,neg}.json
│ ├── adducts_{pos,neg}.json
│ ├── derived/row_metadata_{pos,neg}.npz
│ └── README.md
├── rebuild/
│ ├── build_library_manifest.py the split-B producer (replays MSnLib.csv)
│ ├── baseline_common.py its helper module
│ ├── verify_release.py self-contained verifier for the whole package
│ ├── LICENSE-APACHE-2.0.txt
│ └── README.md
├── hf_dataset_card.md Hugging Face dataset card draft
├── hf_upload_commands.sh upload commands, dry-run by default, not executed
└── zenodo_metadata.json Zenodo deposition metadata draft
```
Every plain `.csv` has a byte-identical `.csv.gz` twin; `rebuild/verify_release.py` checks
this, all counts, all orderings, the cross-split statistics and the checksums.
## 4. Split A — UltraMS MSnLib formula-level split
### 4.1 Definition
For every MSnLib spectrum the neutral molecular formula is computed from the source SMILES
with RDKit (`rdMolDescriptors.CalcMolFormula(Chem.MolFromSmiles(smiles))`). The frozen
`fold` column of `MSnLib.csv` assigns each formula to exactly one fold; the split is the
formula-grouped 80/10/10 partition that results. The rule was verified for all 560,084
spectra: 21,107 formulas, none in more than one fold. This is the frozen benchmark
definition; `build_split_a.py` replays it from `MSnLib.csv` and reproduces every table
byte-for-byte.
### 4.2 Tables
| file | rows | columns |
| --- | ---: | --- |
| `formula_split.csv` | 21,107 | `formula, fold, n_spectra` — ordered by fold (train, valid, test) then formula |
| `row_formula_fold.csv` | 560,084 | `row_index, formula, fold` — 0-based `MSnLib.csv` data row → formula → fold |
| `test_spectra.csv` | 57,437 | `row_index, smiles, precursor_mz, adduct, collision_energy, fold, canonical_isomeric_smiles` |
| `val_spectra.csv` | 55,980 | same schema for the validation fold |
| `test_molecules_voting.csv` | 5,002 | `group_id, smiles, n_spectra` — **official** voting units (MSnLib source SMILES) |
| `test_molecules_voting_canonical_view.csv` | 4,888 | `group_id, canonical_isomeric_smiles, n_spectra` — alternative canonical view |
| `msnlib_rt.csv.gz` | 560,084 | `row_index, rt` — per-spectrum retention time in seconds (gzipped only, see §4.5) |
`row_index` is the 0-based data row of `MSnLib.csv` (physical line = data row + 2). The
tables do not carry peak data; join on `row_index` against `MSnLib.csv` (see §8).
### 4.3 Official voting units — read this before using the voting tables
The **official multiple-spectrum voting units are the 5,002 MSnLib source-SMILES groups
covering 57,436 spectra**, exactly the grouping used by the published multiple-spectrum
voting task (Figure 3d and the Figure 3a voting readouts, `n_molecules = 5002`). The test
fold contains 5,003 distinct source SMILES strings; one occurs in only one spectrum, hence
5,002 groups. This table is `test_molecules_voting.csv` (columns `group_id, smiles,
n_spectra`) and it is verified member-for-member against the repository voting table
(`benchmarks/molecular_identification/data/multiple_spectrum_voting_per_group.csv.gz`).
An **alternative canonical view** is provided as `test_molecules_voting_canonical_view.csv`:
the same test spectra grouped by RDKit canonical isomeric SMILES give 4,888 groups covering
the same 57,436 spectra (114 source-string groups merge into chemically identical
molecules; the single excluded molecule is `CC(C)(C)OC(=O)NCC1(CC(=O)O)CCCCC1`). This view
is not the official voting unit.
No data was altered to force either count. See `split_A_full_msnlib/VERIFY.md` §7 and
`REPORT.json` → `anomalies`.
### 4.5 Retention times (RT)
`split_A_full_msnlib/msnlib_rt.csv.gz` is a compact companion table needed by the
label-blind grouping and multiple-spectrum voting analyses (Figure 3d): `row_index` is the
0-based `MSnLib.csv` data row (0 … 560,083) and `rt` is the retention time in seconds, kept
in the source text representation.
RT is **source metadata from the MSnLib records** (CC BY 4.0, Zenodo
10.5281/zenodo.11163380; see `NOTICE.md`). The table was exported from
`MSnLib_with_rt.csv` (755,984,910 bytes, SHA-256
`a5b7c38228287c5c1de401050445319b331ee9a5f17993d6e104e9a734fe2964`) after verifying
row-by-row that its first seven columns are identical to `MSnLib.csv` in the same order
(560,084/560,084 rows, 0 mismatches; see `split_A_full_msnlib/VERIFY.md` §5.4). The full
756 MB CSV is not duplicated; only this 1.4 MB gzipped table is shipped.
```python
import pandas as pd
test = pd.read_csv("split_A_full_msnlib/test_spectra.csv")
rt = pd.read_csv("split_A_full_msnlib/msnlib_rt.csv.gz") # row_index, rt
test_rt = test.merge(rt, on="row_index", how="left")
assert test_rt["rt"].notna().all()
```
The same `row_index` join works for every row of `row_formula_fold.csv`.
#### Provenance of split A
Source rows, formulas and folds were computed read-only in an isolated data-preparation
environment with RDKit 2026.03.5. The split-A tables were frozen from that run; cross-checks against the
historical per-fold CSV files, against the MSnLib original JSON `formula` fields, and
against the repository voting table are recorded in `split_A_full_msnlib/VERIFY.md`.
## 5. Split B — strict-adduct molecule-level split
### 5.1 Definition
Split B keeps only the strict `[M+H]+` (303,709 source rows) and `[M-H]-` (144,247 source
rows) spectra. After the frozen spectrum preprocessing (drop `m/z ≤ 0` peaks, require at
least 3 peaks, normalise to the base peak, keep the 150 most intense peaks and sort by m/z)
2,073 positive and 4,054 negative rows are dropped, leaving **301,636** and **140,193**
preprocessed spectra. Compound identity is RDKit canonical isomeric SMILES (frozen with
RDKit 2026.03.5). Per polarity the sorted compound list is shuffled with
`random.Random(42)` (`[M+H]+`) / `random.Random(43)` (`[M-H]-`) and split 70 / 15 / 15;
one query spectrum per validation/test compound is drawn with a fresh `random.Random(119)`.
### 5.2 Tables
| table | schema |
| --- | --- |
| `splits/{pos,neg}_compound_split.csv` | `polarity, compound_id, canonical_isomeric_smiles, split, n_spectra, n_frozen_queries, n_raw_smiles_variants, first_msnlib_csv_data_row` |
| `queries/{pos,neg}_{val,test}_queries.csv` | `polarity, split, query_index, compound_id, canonical_isomeric_smiles, adduct, msnlib_csv_data_row, msnlib_csv_line, precursor_mz, n_peaks, collision_energy, source_smiles_raw, n_library_positives` |
| `libraries/{pos,neg}_{val,test}_library.csv` | same as queries with `library_index` and `compound_split` instead of `query_index`; row order is the frozen library order |
### 5.3 Protocol
1. Score every query against the **complete** library of its task (no candidate short list).
2. A retrieved library spectrum is correct iff its canonical isomeric SMILES equals the
query's (exact same-adduct spectrum of the same molecule).
3. Empirical FDR = `(accepted − correct) / accepted`; identification recall =
`correct / n_queries_with_library_positive` (6,201 / 3,160 for the two test tasks).
Queries without a library positive enter the FDR denominator but not the recall
denominator; `n_library_positives` is published per query.
4. The **test library** is every preprocessed spectrum of the polarity except the selected
test queries (it therefore still contains further spectra of the test molecules; a query
never occurs in its own library). The **validation library** contains train and
validation compounds only, so test compounds never influence validation scores.
5. Validation is for model selection only; the test fold is evaluated once.
The split-B tables deliberately do **not** carry the formula-level fold of the spectra:
that label belongs to split A. To obtain each split-B row's split-A fold, join
`msnlib_csv_data_row` (queries/libraries) or `first_msnlib_csv_data_row` (compound split)
against `split_A_full_msnlib/row_formula_fold.csv`; the packaged verifier does exactly this
for the cross-split statistics.
The frozen row-index arrays behind these tables are in `protocol/`; the CSV tables are a
human-readable view. `rebuild/build_library_manifest.py` replays them bit-for-bit from
`MSnLib.csv`, and `rebuild/verify_release.py` checks the tables against the arrays and the
frozen manifest.
### 5.4 Cross-collision-energy subtask
The cross-collision-energy library-search task uses split B test queries that were measured
at two or more collision energies: **6,161 positive** and **2,999 negative** queries (out of
6,202 / 3,193 split-B test queries). The task pairs a low-energy query spectrum with a
high-energy reference library (or vice versa) and asks for the same-molecule hit; the
query identities are a subset of the split-B test queries published here. The CE values in
MSnLib are 15, 20, 30, 45, 60 and 75 eV. The cross-CE pairing manifest and readouts are
part of the UltraMS repository, not of this split package.
## 6. Relation to split A
Split A and split B are independent; see §2 for the disjointness statement and
[figures_mapping.md](figures_mapping.md) §2 for the full cross-tabulation. Summary: 58.9 %
positive and 59.4 % negative split-B compounds keep the same fold in split A; 654 positive
and 303 negative split-B test compounds are also split-A test compounds; 31,570 / 13,413
split-A test spectra are `[M+H]+` / `[M-H]-` rows of split B.
## 7. Manuscript figures
[figures_mapping.md](figures_mapping.md) maps every figure panel to the split it uses:
* **A**: Figure 2b, Figure 2j, Figure 3a, 3b, 3d, 3g, 3a_sp_4, 3a_sp_5 and the
single-spectrum baselines Top-1 table;
* **B**: Figure 3e, 3f, 3h, 3h_sp_1 and the cross-collision-energy task (6,161 / 2,999);
* **not this dataset**: NPLIB1 and MassSpecGym comparison panels (independent external
benchmarks); Figure 3c / 3i case studies are not split-defined.
Candidate-ranking tables (the per-formula candidate lists behind Figure 3a/3b/3d/3g) are
part of the UltraMS repository, not of this split package; this package provides the frozen
queries, identities and folds they are evaluated on.
## 8. Data acquisition and rebuilding
### 8.1 The spectra are not part of this manifest
This package publishes splits and identifiers, not peak arrays. Every table is joined back
to the spectra through `row_index` / `msnlib_csv_data_row`, which point into `MSnLib.csv`;
both large files are hosted in the same Hugging Face dataset (see the managed-file table
below) rather than duplicated in this manifest.
| item | value |
| --- | --- |
| file | `MSnLib.csv` — one row per MSnLib spectrum |
| rows | 560,084 data rows (560,085 physical lines, no embedded newlines) |
| bytes | 752,707,651 |
| SHA-256 | `be54b33039cf18675751ea3c6354b592177c899a0cd5df303ae616330133ec07` |
| columns | `mzs, intensities, smiles, precursor_mz, adduct, collision_energy, fold` |
| CSV format | peaks stored as comma-separated strings (`mzs`, `intensities`), one spectrum per row |
**Managed files on Hugging Face.** The dataset is planned for Hugging Face under the
placeholder id **`dsadd4/UltraMS-MSnLib-Benchmark`** (dataset card and upload commands are
drafted here as `hf_dataset_card.md` and `hf_upload_commands.sh`; the upload has **not**
been executed). The repository will host:
| hosted file | bytes | SHA-256 | purpose |
| --- | ---: | --- | --- |
| every file of this manifest (`split_A_full_msnlib/`, `splits/`, `queries/`, `libraries/`, `protocol/`, `rebuild/`, docs; 59 files) | ≈ 185.2 MiB total | see `SHA256SUMS` | the two benchmark splits, verifier, docs |
| `split_A_full_msnlib/msnlib_rt.csv.gz` | 1,483,069 | `994053612db874f7ac7985f278e04cc3463f3b93cd649a195bd167c4c6c7d3d8` | compact per-spectrum retention time (`row_index, rt` in seconds); joins `MSnLib.csv` rows 1:1 and supplies the RT used by the label-blind grouping analysis (Figure 3d) |
| `MSnLib.csv` | 752,707,651 | `be54b33039cf18675751ea3c6354b592177c899a0cd5df303ae616330133ec07` | the full source CSV: one row per MSnLib spectrum; every `row_index` / `msnlib_csv_data_row` / `first_msnlib_csv_data_row` pointer resolves here |
| `MSnLib_candidates.json` | 1,368,264,026 | `125a1cdcc57a3fcef8cc76d5322ac0c434cd09b0bfb9e0c9d28bddf052f45fcc` | candidate SMILES lists behind the single-spectrum candidate-ranking task: a JSON object mapping a query SMILES string to its candidate SMILES list (9,886 query entries) |
The large files are hosted because every table points into them: spectra (`MSnLib.csv`) and
candidate rankings (`MSnLib_candidates.json`). The upload script keeps them in a separate
step and remains dry-run by default.
It can also be rebuilt from the **MSnLib original release** (Brungs et al., Zenodo
[10.5281/zenodo.11163380](https://doi.org/10.5281/zenodo.11163380), CC BY 4.0): the release
provides 18 per-file MS2 JSONL record lists with exactly 560,084 records in total (equal to
the CSV data-row count), each record carrying `peaks`, `smiles`, `precursor_mz`, `adduct`
and `collision_energy`. The UltraMS CSV conversion writes one row per record in the frozen
conversion order with the columns above and carries the frozen fold labels. The
authoritative row → formula → fold mapping is shipped independently as
`split_A_full_msnlib/row_formula_fold.csv`, so downstream work never depends on
reconstructing the conversion order.
### 8.2 Rebuilding split A
```bash
python3 split_A_full_msnlib/build_split_a.py \
--csv /path/to/MSnLib.csv \
--output-dir replay_split_a
```
Needs Python ≥ 3.10 and RDKit 2026.03.5 (the frozen canonicalisation runtime). The script
verifies all counts and all six table SHA-256 values; with a different RDKit patch release
the canonical-string tables can differ textually while every count still matches (the
script reports that, or downgrades it to a warning with `--lenient-rdkit`).
### 8.3 Rebuilding split B
See [rebuild/README.md](rebuild/README.md). In short:
```bash
python3 rebuild/build_library_manifest.py \
--csv /path/to/MSnLib.csv --skip-reference-regression \
--output-dir replay_manifest
```
This regenerates `protocol/*.npz`, `smiles_*.json`, `adducts_*.json` and
`manifest_summary.json` (needs NumPy, h5py, RDKit). The CSV tables are then rebuilt from the
protocol arrays through `protocol/derived/row_metadata_*.npz` (see `protocol/README.md`).
## 9. Verification
```bash
python3 rebuild/verify_release.py # self-contained; no network, no MSnLib download
sha256sum -c SHA256SUMS # or: shasum -a 256 -c SHA256SUMS (macOS)
```
`rebuild/verify_release.py` re-derives every published count, ordering hash, disjointness
statement, query/library partition, per-query library positive, cross-split statistic and
file checksum from the shipped files; the complete transcript is in [VERIFY.md](VERIFY.md).
Split-A-specific evidence (including the read-only replay and the voting-unit check) is in
`split_A_full_msnlib/VERIFY.md`.
## 10. Known properties and caveats
1. **Two splits, two disjointness units.** Split A is formula-disjoint; split B is
compound-disjoint within each polarity. Do not merge them into one global-disjointness
claim and do not reuse a fold label across splits (§2).
2. **Voting units: official 5,002 source-SMILES groups / 57,436 spectra, alternative
canonical view 4,888 groups / 57,436 spectra.** The official unit is the one used by the
published `n_molecules = 5002` grouping; both tables are shipped (§4.3).
3. **Retention times are split-A row metadata.** `split_A_full_msnlib/msnlib_rt.csv.gz`
(`row_index, rt` in seconds) is source metadata from the MSnLib release, verified
row-aligned to `MSnLib.csv`; it is not part of split B and carries no fold label of its
own.
4. **Queries without a library positive are expected** in split B: 1 positive and 33
negative *test* queries (and 8 / 30 *validation* queries) are singleton compounds whose
only spectrum is the query. They enter the FDR denominator but not the recall
denominator; `n_library_positives` is published per query.
5. **Singleton compounds are retained** in split B (39 positive and 202 negative compounds
have exactly one spectrum). Filtering them changes the benchmark.
6. **Compound identity is frozen text** (RDKit 2026.03.5 canonical isomeric SMILES). Other
RDKit patch releases can serialise the same molecule differently; compare the published
strings, not freshly canonicalised ones. Molecular formulas are stable across RDKit
patch releases.
7. **Precursor m/z** is the source CSV text value; the split-B protocol stores it as
float32 and all 441,829 preprocessed rows agree exactly after the cast.
8. **Collision energy is source metadata**, not a selection criterion; all six values
(15, 20, 30, 45, 60, 75 eV) occur in both polarities.
9. **Split B libraries are large by design** (≈47.6 positive / ≈42.9 negative spectra per
query on average); ranking must cover the full library, there is no candidate short list.
10. **Split A includes all adducts** (11 distinct adducts in the test fold), not only the
strict `[M+H]+` / `[M-H]-` subset used by split B.
## 11. Provenance and licensing
The upstream MSnLib release (spectra and molecule metadata) is CC BY 4.0 (Brungs et al.,
Zenodo 10.5281/zenodo.11163380). The split definitions, tables and scripts added by UltraMS
are released under CC BY 4.0 as well. `rebuild/build_library_manifest.py` and
`rebuild/baseline_common.py` are Apache-2.0 files copied from the UltraMS public release.
See [NOTICE.md](NOTICE.md) for the full attribution and licensing statement, and
`protocol/manifest_summary.json` for the frozen split-B manifest (one path field is
redacted to a placeholder; every count, seed and hash is byte-identical).
## 12. Citation
If you use this benchmark, please cite MSnLib and UltraMS:
* MSnLib: Brungs C., Schmid R., Pluskal T. *MSnLib: efficient generation of open
multi-stage fragmentation mass spectral libraries.* Nature Methods (2025),
doi:10.1038/s41592-025-02813-0; data: Zenodo 10.5281/zenodo.11163380 (CC BY 4.0).
* UltraMS: Li Y., Zhong Y., Liu P., Xia J. *UltraMS* (manuscript in preparation;
code: <https://github.com/Dsadd4/UltraMS>).
Machine-readable citation metadata is in [`zenodo_metadata.json`](zenodo_metadata.json) and
[`hf_dataset_card.md`](hf_dataset_card.md).
## 13. Versioning
`1.0.0` is the first joint release of splits A and B. Its content is defined by
`stats.json`, `SHA256SUMS` and the frozen `protocol/manifest_summary.json`. Any change to
the splits, the compound/formula identity or the library scope requires a new version
directory and new manifests — the frozen tables must stay frozen so that published
FDR/recall/ranking numbers remain comparable.