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

Model weights and frozen inputs needed to reproduce the retrieval (Table 3) and de novo (Table 5) rows of MassSpecGym in the Wild: Uncovering and Correcting Evaluation Pitfalls in AI-Driven Molecule Discovery. The code, configurations and per-row instructions are in github.com/harrylaucngd/msms-models.

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

hf download harrylau/msms-models --local-dir .
bash paper/prepare_data.sh
bash paper/table3/mist_4m_aug/run.sh

Path	Contents
data/paper/table3/	Table 3 test queries with imputed collision energies, and the forward-model candidate pool
data/paper/posteriors/	MIST 4M-aug fingerprint posteriors used by the Table 3 and Table 5 rows
data/paper/table5/mist_vdiffms/	DiffMS subformula trees for the vDiffMS diagnostic rows
checkpoints/mist_msg_aug.ckpt	MIST 4M-aug (MassSpecGym training set + GLACIER synthetic spectra)
checkpoints/molforge/	MolForge decoder and SentencePiece models (CC BY-NC-SA 4.0)
checkpoints/frigid/	SAFE tokenizer and NGBoost length predictor paired with the FRIGID DLM
checkpoints/iceberg/, checkpoints/glacier/	ICEBERG pair used for FRIGID refinement; ICEBERG pair and GLACIER retrained 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)

Downloaded separately from their original sources:

• FRIGID DLM.ckpt: Zenodo 19685145.
• The DreaMS SSL checkpoint: Zenodo 10997887.
• The DiffMS MIST: Zenodo 15122968.

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. Inflated checkpoints reproduce the paper's shortcut-learning diagnostics and are not valid benchmark models.