harrylau/ms-gpt-18

Decoder pretraining corpora: 18 controlled variants

This repository releases the 18 molecular decoder pretraining files used in the controlled pretraining study in MassSpecGym in the Wild: three nominal scales (2.5M, 10M, and 50M molecules) and six selection policies per scale.

The files support fingerprint-to-SMILES decoder training and experiments on the effect of exposure to MassSpecGym test structures. The include_half and include_all variants intentionally include test structures for these controlled experiments; their results should be interpreted under that exposure condition.

Files and scales

The original filenames and file contents are preserved. The historical 3M_ filename prefix refers to the approximately 2.5M-record group in the manuscript.

Filename prefix	Nominal scale	Records in each released file
3M_	2.5M	2,499,999; exclude_tani060 has 2,499,998
ds_10M_	10M	10,000,000
ds_50M_	50M	50,000,000

Each prefix has these six suffixes: exclude_exact, exclude_tani060, exclude_tani070, exclude_tani080, include_half, and include_all, followed by .morgan4096.smiles.train.

Across all variants, the files contain 374,999,993 records and occupy 123,961,153,007 bytes (approximately 124 GB). Variants overlap by construction; the aggregate record count is not a count of distinct molecules. Exact per-file row counts, byte sizes, and SHA-256 checksums are provided in manifest.json.

Selection policies

The following definitions describe the construction conditions in the associated study. The reference is the MassSpecGym test set.

Variant	Construction condition
exclude_exact	Excludes molecular connectivity identities matching any test molecule, using the first 14 characters of the InChIKey.
exclude_tani080	Retains molecules with maximum fingerprint Tanimoto similarity to test molecules strictly below 0.80.
exclude_tani070	Retains molecules with maximum fingerprint Tanimoto similarity to test molecules strictly below 0.70.
exclude_tani060	Retains molecules with maximum fingerprint Tanimoto similarity to test molecules strictly below 0.60.
include_half	Replaces non-test molecules in the size-matched exact-exclusion construction with 50% of the unique test structures (1,585 structures).
include_all	Replaces non-test molecules with all unique test structures (3,170 structures).

The manuscript defines the similarity comparisons using ECFP4 Morgan fingerprints (radius 2, 2048 bits). The stored decoder-conditioning fingerprints are 4096 bits, as described below. The release preserves the existing study files; no additional filtering, deduplication, or regeneration was performed for this upload.

Record format

Each file is UTF-8 text, with no header, one molecule per line, and two tab-separated columns:

<space-tokenized canonical SMILES>\t<space-separated active fingerprint indices>

• Column 1: tokenized SMILES. Concatenating the whitespace-separated tokens reconstructs the SMILES string.
• Column 2: active-bit indices of a Morgan radius-2, 4096-bit fingerprint, encoded as integers from 0 to 4095. This is a sparse list of indices, not a 4096-character binary string.
• The .train extension denotes a decoder training corpus; the files are not model weights or raw mass spectra.

For example, read records without loading a whole file into memory:

filename = "3M_exclude_exact.morgan4096.smiles.train"
with open(filename, encoding="utf-8") as source:
    for line in source:
        tokenized_smiles, fingerprint = line.rstrip("\n").split("\t")
        smiles = "".join(tokenized_smiles.split())
        active_bits = [int(bit) for bit in fingerprint.split()]
        # Consume one molecule at a time.

The representation follows the MolForge fingerprint-to-molecule training format.

Download

Download a selected variant with the Hugging Face CLI:

hf download harrylau/ms-gpt-18 \
  3M_exclude_exact.morgan4096.smiles.train \
  --repo-type dataset --local-dir ./ms-gpt-18

Download the complete collection, approximately 124 GB:

hf download harrylau/ms-gpt-18 --repo-type dataset --local-dir ./ms-gpt-18

Data provenance and licensing

The nominal 2.5M group is derived from the composite molecular collection containing DSSTox, HMDB, COCONUT, and MOSES, as described in the associated manuscript. The 10M and 50M groups are sampled from the approximately billion-molecule ZINC20/UniChem source pool associated with SAFE-GPT.

Please retain attribution to the original databases and comply with their applicable source terms. This repository does not assert a new blanket license over those databases. The other license metadata reflects these source-specific terms; the software license of a training implementation should not be interpreted as relicensing its molecular data. SAFE-GPT's dataset card declares Apache-2.0, while the SAFE project's data-license notice states CC BY 4.0.

Relevant source publications include Grulke et al. (2019, DSSTox), Wishart et al. (2022, HMDB), Sorokina et al. (2021, COCONUT), Polykovskiy et al. (2020, MOSES), Irwin et al. (2020, ZINC20), Chambers et al. (2013, UniChem), and Noutahi et al. (Gotta be SAFE). Please also cite the associated MassSpecGym in the Wild study when using these controlled variants.

For reproducible use, record this repository's commit revision and the file checksum in manifest.json.