QM9/METADATA.md

# QM9 dataset and TD-jumps split

## Summary

This directory contains the with-hydrogen QM9 molecular geometries and the
exact train/validation/test split used by the QM9 experiments in the TD-jumps
paper, *Trans-Dimensional Generative Modeling via Jump Diffusion Models*.

The split was reconstructed from the released jump-diffusion code at commit
`b2e31007d2c5792dddc1bf03f2e92bf889f42fd8`. The stored split arrays were
verified byte-for-byte against an independent reconstruction of the TD-jumps
row arrays.

## Origin and filtering

The molecular data come from the official GDB9/QM9 XYZ archive:

- Archive: <https://ndownloader.figshare.com/files/3195389>
- Archive SHA-256:
  `3a63848ac80691bdb8d41834b575afad345b9300d7a2db0c38adb7f6eaa8360c`
- Official uncharacterized-molecule list:
  <https://ndownloader.figshare.com/files/3195404>
- Exclusion-list SHA-256:
  `3aa5115d540b356de94791d4a74c3bf1ed91c469ecf52a4f5d7cc0506fe02e24`

The original archive contains 133,885 molecules. The same 3,054
uncharacterized molecules excluded by the released TD-jumps loader were
removed, leaving 130,831 molecules. `qm9.npz` stores those retained molecules
in their original archive order after filtering; the split files contain the
TD-jumps permutation as row indices into that packed file.

## Exact TD-jumps split

The released TD-jumps code applies NumPy's legacy `RandomState`/MT19937
permutation with seed 0 to the 130,831 retained molecules, then divides that
permutation as follows:

| File | Molecules | Permutation positions |
|---|---:|---:|
| `train_rows.npy` | 100,000 | 0--99,999 |
| `valid_rows.npy` | 17,748 | 100,000--117,747 |
| `test_rows.npy` | 13,083 | 117,748--130,830 |

Each split file is a one-dimensional `int64` NumPy array of zero-based rows in
`qm9.npz`. The three arrays are disjoint, contain every row exactly once, and
retain the original seed-0 permutation order used by TD-jumps. They are not
GDB9 molecule IDs; the corresponding one-based GDB9 ID for a packed row is
available in `qm9.npz` as `source_ids[row]`.

For example:

```python
import numpy as np

with np.load("qm9.npz", allow_pickle=False) as qm9:
    offsets = qm9["molecule_offsets"]
    coordinates = qm9["coords"]
    atomic_numbers = qm9["atomic_numbers"]
    train_rows = np.load("train_rows.npy", allow_pickle=False)

    row = int(train_rows[0])
    left, right = offsets[row : row + 2]
    first_training_molecule = {
        "coords": coordinates[left:right],
        "atomic_numbers": atomic_numbers[left:right],
    }
```

## Packed data schema

`qm9.npz` is a compressed NumPy archive containing:

| Array | Shape | Dtype | Meaning |
|---|---:|---|---|
| `coords` | `(2_359_210, 3)` | `float32` | Concatenated XYZ coordinates in angstroms |
| `atomic_numbers` | `(2_359_210,)` | `int16` | Concatenated atomic numbers |
| `molecule_offsets` | `(130_832,)` | `int64` | Half-open offsets into the atom arrays |
| `source_ids` | `(130_831,)` | `int64` | One-based IDs in the official GDB9 archive |
| `vocabulary_atomic_numbers` | `(5,)` | `int16` | Model vocabulary order `[6, 1, 7, 8, 9]` |

The atom order inside each molecule is the order in the official XYZ file; the
packed dataset does not reorder atoms. In 130,372 molecules all hydrogens follow
all heavy atoms, 417 molecules have hydrogens intermingled with heavy atoms,
and 42 molecules contain no hydrogen. The released TD-jumps training
configuration subsequently randomizes the real-atom ordering on each dataset
fetch (`shuffle_node_ordering=true`), but that augmentation is not baked into
these files.

## File checksums

SHA-256 checksums of the files in this directory:

| File | SHA-256 |
|---|---|
| `qm9.npz` | `ddba2129feb7f1676ae0137820d08e489555b8d928c2b943f80be6673dc2ef26` |
| `train_rows.npy` | `aae8ded4c106664ca06029f697bd6f4b9d2ebd12f1ab8b0506950aa79d3613f1` |
| `valid_rows.npy` | `ac3a8de760f21a278a64375c2d1aa01468ba73a13cf42613a6511882a41871e7` |
| `test_rows.npy` | `94bd03e0e51002a954030efe7610e8b81a44dbd51f57d1a9a96d50d0f21c4990` |

The reconstruction verified all 130,831 retained molecules against the
MurrellLab QM9 serialization: element arrays matched exactly and coordinates
were bitwise identical after conversion to `float32`.