# GEOM-Drugs with explicit hydrogens

Published completed epochs: 50, 100.

Continued from the best checkpoint after 48 completed epochs. The training subset contains at most 72 atoms including hydrogen.

The original Lightning checkpoints include current weights, EMA weights and optimizer state. The inference loader constructs only the selected network (EMA by default).

```python
from morph.inference import load_pretrained

model = load_pretrained("geom-explicit-h", epoch=100, device="cuda")
trajectories, validity, vocab, coordinate_std = model.generate(n_mols=16)
```

Coordinates returned by generation are in Angstrom. Config, vocabularies and training distributions are included; no dataset is needed for inference. Training settings and source are in `training_bundle.tar.gz`.