Molecular LPGM

Reference implementation and documentation

Final Fragment Noise and Brownian Atoms models for GEOM-Drugs and QM9. Both methods condition on atom count and the BRICS fragment partition, with learned same/different-fragment attention biases in every transformer layer. Neither receives clean coordinates, atom labels, bonds or geometric templates at generation. Fragment Noise uses a single fixed global angular variance of 6 in both datasets.

Dataset	Width	Layers	Fragment Noise	Brownian Atoms
GEOM-Drugs	768	16	fragment.pt	bm.pt
QM9	512	12	fragment.pt	bm.pt

All four checkpoints contain final 800,000-update raw and EMA weights (decay 0.9999), architecture, process parameters, and final loss weights. The final 100,000 updates are learning-rate warmdown. Optimizer state is omitted. Use --ema in the reference sampling commands, or load_checkpoint(..., use_ema=True) in Python. SHA-256 hashes and parameter counts are in checkpoints.json.

The source is a centered Gaussian with standard deviation 3 Å. Brownian Atoms uses the scalar clock u(t) = 1 - (1 - t**2.3)**3.1 and variance coefficient beta = 14.8 Ų; these parameters approximately match the conditional endpoint spread. It has one Cartesian coordinate head and the same atom-label process as Fragment Noise.

10K-sample results and archives use final EMA weights and 2,000 square-root-spaced steps. The same ADiT validity, valid-and-connected, and uniqueness protocol is used for both datasets. GEOM evaluation uses test cases; QM9 pools validation and test cases. See the comparison protocol for conditioning, objective and training details.