# AutoencoderDataset

Data for training **molvae** — a SELFIES molecular VAE (matryoshka latent) for Bayesian
optimization of mechanophores. Two layers:

- **`raw/`** — the source chemical databases (SMILES/SELFIES, pre-tokenized), as collected.
- **`mixes/`** — the *derived* training datasets: weighted, shuffled, materialized token-shard
  "mixes" built from the raw sources. Each mix is documented by its **config + README + build
  code**, so it is fully reproducible. The multi-GB token shards themselves live on **S3**
  (`s3://mechanophore/molvae/tokens/<name>/`) — the training home — and are mirrored here.

Everything shares one tokenizer (`raw/tokenizer.json`, vocab 1458; cls=1, sep=2, pad=3).

## Raw sources (`raw/`)
Uniform schema per parquet: `sequence` (SELFIES string) + `tokenized_sequence` (int16,
`[cls, …ids, sep, pad…]`). Highlights:

| source | rows | notes |
|--------|-----:|-------|
| `gdb/gdb13` | 977M | exhaustive ≤13-heavy-atom enumeration — dense strained/spiro/bridged rings |
| `gdb/GDB-13s`, `gdb/gdb11`, `gdb/GDB4c3D`, `gdb/GDBMedChem`, `gdb/GDB17…` | 6.6M–99M | GDB variants (strained/ring-rich, small) |
| `mcule_full/…` | 42M | purchasable; **long molecules** (median 46, to 154 tok) |
| `emolecules/ordersc/…` | 36M | purchasable; **long + diverse** (to 168 tok) |
| `Kaggle/chembl_22_clean…` | 1.6M | drug-like (short) |
| `zinc22-2d-all/…` | ~1.9B | ZINC22 SELFIES by heavy-atom band H04–H49 (source of the `zinc-tw2b-v2` corpus) |

## Mixes (`mixes/`)
See [`mixes/README.md`](mixes/README.md) for the catalog, composition tables, and the exact
rebuild command. Builder: [`mixes/mix_dataset.py`](mixes/mix_dataset.py) (from the molvae repo).

| mix | mols | on S3 | what |
|-----|-----:|-------|------|
| `zinc-tw2b-v2` | 1.855B | `…/tokens/zinc-tw2b-v2` | the ZINC-only base corpus (H04–H49) |
| `mix-phase1-40M` | 40M | `…/tokens/mix-phase1-40M` | v1 pilot: ZINC-heavy (40%) + GDB4c3D/GDBMedChem/ChEMBL |
| `mix-v2-40M` | 40M | `…/tokens/mix-v2-40M` | v2 pilot: strained-heavy (GDB) + **long** (mcule/eMol) |
| `mix-v2-1B` | 1B | `…/tokens/mix-v2-1B` | v2 full: the strained-rich pretrain mix |

## Held-out
The ~132 real mechanophores are a **held-out reconstruction test set** (not in any training mix;
every build dedups against them). They live on S3 (`…/eval/mechano_test.jsonl`), not here.