README.md

---
license: cc-by-4.0
pretty_name: msnlib
configs:
- config_name: full
  default: true
  data_files:
  - split: train
    path: data/msnlib/*.parquet
- config_name: compounds
  data_files:
  - split: train
    path: data/compounds/*.parquet
---

# msnlib

Pinned source: https://zenodo.org/records/21105617 . This mirror contains 4,532,018 source MGF records with original peak order and intensities, plus original source files. No molecule standardization, chemistry filter, peak normalization, or new benchmark split is applied. `train` is a storage convention.

Metadata values and duplicate metadata keys remain in `source_metadata_json`; source identifiers and record indices are preserved. Raw charge, precursor, adduct, and collision-energy text are retained rather than imputed. Extra peak fields are preserved. Empty/negative-intensity source spectra are flagged and retained.

MS2 and MSn exports overlap by design: source MS2 files include pseudo-MS2 spectra, and MSn files also contain individual higher fragmentation stages. Do not count merged spectra or duplicate export views as independent measurements. `SPECTYPE` is retained and pseudo-MS2 flagged. The `source_file` suffix distinguishes MS2-only/MSn views and polarity. Source compound metadata and both JSON/MGF representations remain under `source/`. The `compounds` configuration exposes the source detected-compound Parquet unchanged.

Published file MD5 checksums, row counts, peak-vector shapes, and finite values were checked. Full chemical annotation validity and overlap with other datasets remain unaudited. Cite the original source authors and publication described in `source/source_metadata.json`.