README.md

metadata

pretty_name: MolWeaver Ligands 100M
task_categories:
  - text-generation
  - feature-extraction
tags:
  - chemistry
  - molecules
  - selfies
  - conformers
  - rdkit

MolWeaver Ligands 100M

This dataset contains 100,000,000 globally unique heavy-atom molecular records split across five LMDB shards. Each shard contains 20,000,000 records and all records are assigned to the training split.

Record schema

Each numeric LMDB key contains a pickled Python dictionary with:

• smi: canonical heavy-atom molecule encoded as SELFIES.
• atoms: heavy-atom symbols in decoded SELFIES atom order.
• coordinates: float32 NumPy array shaped (10, n_heavy, 3) containing ten original Cartesian conformers. Coordinate atom i matches atoms[i].
• qed: RDKit QED.
• sa_score: RDKit Contrib synthetic accessibility score.
• molecular_weight: RDKit average molecular weight.
• mol_log_p: RDKit MolLogP.
• tpsa: RDKit topological polar surface area.

Explicit hydrogen atoms and hydrogen coordinates are not included. Standard implicit hydrogens are used by RDKit when calculating molecular properties.

Files

ligands/shard_1.lmdb
ligands/shard_1_metadata.json
...
ligands/shard_5.lmdb
ligands/shard_5_metadata.json
ligands/valid.lmdb
ligands/valid_metadata.json

Every shard_*.lmdb stores numeric keys b"0" through b"19999999" and a pickled b"length" value equal to 20_000_000.

The local deduplication registries used during generation are not uploaded; they are not needed to train from the finalized records.

valid.lmdb uses the same record schema and contains 10,000 unique SELFIES generated from unused source slice 6. All 10,000 entries were checked against the finalized training registries and have no overlap with the 100M training records. Its numeric keys are b"0" through b"9999", and its pickled b"length" value is 10_000.

Loading

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"))

Pickle should only be loaded from a trusted dataset source.

Uniqueness

SELFIES were deduplicated exactly within each shard and separated across shards by SHA-256 hash ownership. All five finalized registries contained exactly 20,000,000 entries with no registry-only extras, establishing 100,000,000 unique SELFIES records in total.