Skip to content

Latest commit

 

History

History
21 lines (11 loc) · 1.5 KB

MODEL_CARD.md

File metadata and controls

21 lines (11 loc) · 1.5 KB

GLIDER model card

Intended use

Prediction of counterpoise-consistent mutual response ESP and induced dipole for complexes containing a neutral organic solute and structured multi-water clusters.

Inputs and outputs

Input is atomic identity and geometry. The frozen broad MACE-POLAR features and unfitted M/L response prior feed a site-response head. Outputs are net-neutral induced site charges and equivariant site dipoles, from which response ESP and molecular dipole are reconstructed.

Training

Fourteen development chemistries, 48 environments, five seeds. Direct response ESP and dipole supervision; no energy, force, hydration or prospective labels. Chemistry-blocked model selection. Exact settings are in the checkpoint manifest and frozen source.

Validation

Prediction-before-label panels of 12, 24 and 20 chemistries absent from GLIDER response supervision. Panel III used a deterministic label-blind selection from pinned FreeSolv v0.52 after excluding historical exact matches and Morgan similarity at least 0.70. FreeSolv hydration values were not accessed.

Limitations

Neutral organics, three-water environments and four designed regimes only. Not a conservative potential, not a force field, not validated for MD or free energy. MACE-POLAR checkpoint licenses apply to the prerequisite encoders. The broad encoder may have seen exact identities or analogues during foundation pretraining; the transfer claim concerns absence from GLIDER response supervision, not pretraining-naive chemistry.