msms-models: weights and frozen inputs for MassSpecGym in the Wild

Model weights and frozen inputs for the implemented entries of MassSpecGym in the Wild: Uncovering and Correcting Evaluation Pitfalls in AI-Driven Molecule Discovery. The code, configurations and per-entry instructions are in github.com/harrylaucngd/msms-models.

Download into the root of a checkout; the paper configurations expect this layout:

bash environments/install.sh && source .venv/bin/activate
bash paper/prepare_data.sh
hf download harrylau/msms-models --local-dir . --include 'checkpoints/*' 'data/*'
bash paper/table3/mist_4m_aug/run.sh

Path	Contents
data/paper/posteriors/	MIST fingerprint posteriors: the Table 1 encoders, MIST 4M-aug with MIST-CF or ground-truth formulas (Table 2), and MIST 4M-aug for Tables 3–5 and Figure 2
data/paper/table2/	MIST-CF top-1 formulas behind the Table 2 posteriors
data/paper/table3/	Table 3 test queries with imputed collision energies, the forward-model candidate pool, and the MIST vMassSpecGym subformula trees
data/paper/table5/mist_vdiffms/	DiffMS subformula trees
checkpoints/mist_msg_aug.ckpt	MIST 4M-aug (MassSpecGym training set + GLACIER synthetic spectra)
checkpoints/mist_vmassspecgym/	MIST vMassSpecGym (Zenodo 11580401)
checkpoints/mist_vdiffms/	MIST released with DiffMS (Zenodo 15122968)
checkpoints/molforge/	MolForge decoder and SentencePiece models (CC BY-NC-SA 4.0)
checkpoints/frigid/	FRIGID DLM (Zenodo 19685145), SAFE tokenizer and NGBoost length predictor
checkpoints/iceberg/, checkpoints/glacier/	ICEBERG pair for FRIGID refinement and FOAM; ICEBERG pair and GLACIER trained on the energy-completed split for Table 3
checkpoints/chemformer/, checkpoints/chemformer_two_stage/	Chemformer two-stage retrieval (inflated and corrected)
checkpoints/dreams_chemberta/	DreaMS + ChemBERTa alignment (inflated and corrected)

The decoder pretraining sets of the paper's Figure 2 are in harrylau/ms-gpt-18.

Each component keeps the license of its original method; see the code repository's third-party notices. The inflated checkpoints are the paper's shortcut-learning examples, not valid benchmark models.