## MolWeaver pretraining datasets

This directory contains the finalized LMDB inputs used for MolWeaver
pretraining.

### Free ligands

`ligands/` contains five heavy-atom molecular training shards. Shards 1–4
contain 20,000,000 records each. The last 3,000 records were removed from
shard 5 and placed in `valid.lmdb`, leaving 19,997,000 training records in
shard 5 and 99,997,000 training records overall.

Each record contains a SELFIES representation, heavy-atom identities, ten
aligned Cartesian conformers, and five RDKit-derived properties: QED,
synthetic accessibility score, molecular weight, MolLogP, and TPSA.

### Protein pockets

`pockets_center_frame_cartesian/` contains 3,123,776 protein-pocket training
records in a deterministic heavy-atom Cartesian frame. There is no pocket
validation LMDB in this release.

### CrossDocked complexes

`crossdock/` contains paired heavy-atom pocket-ligand records with ligand
SELFIES, pocket and ligand coordinates, and Vina docking scores. The training
split contains 326,896 records after excluding seven examples whose
conservative sequence length exceeded 4,356. The validation split contains
3,000 records.

### Loading

```python
import lmdb
import pickle

env = lmdb.open(
    "ligands/shard_1.lmdb",
    readonly=True,
    subdir=False,
    lock=False,
    readahead=False,
)
with env.begin() as txn:
    length = pickle.loads(txn.get(b"length"))
    record = pickle.loads(txn.get(b"0"))
```

Pickled records should only be loaded from a trusted dataset source.