QHFlow — original checkpoints
This repository mirrors all 10 original QHFlow checkpoints from the
_qhflow1 Google Drive collection.
The original file names, folder structure, and checkpoint bytes are preserved.
The collection contains 4 QH9, 3 MD17, and 3 rMD17 checkpoints
(5,351,745,377 bytes in total).
These are the original QHFlow models. QHFlow2 checkpoints are available separately: QH9, MD17, and rMD17.
Files
The source uses the historical folder name qm9/ for its QH9 checkpoints;
this name is retained for traceability. In that folder, geo denotes the
QH9Dynamic geometry split. No model sizes or training configurations are
inferred from checkpoint filenames.
| Original directory | Checkpoint | Size (MB, decimal) |
|---|---|---|
md17/ethanol |
weights-epoch=39-val_loss=0.00.ckpt | 580.84 |
md17/malondialdehyde |
weights-epoch=39-val_loss=0.00.ckpt | 580.84 |
md17/uracil |
weights-epoch=39-val_loss=0.00.ckpt | 580.84 |
qm9/QH9Dynamic-300k-geo |
weights-epoch=33-val_loss=0.0000000.ckpt | 331.35 |
qm9/QH9Dynamic-300k-mol |
weights-epoch=32-val_loss=0.0000000.ckpt | 331.35 |
qm9/QH9Stable-random |
weights-epoch=78-val_loss=0.0000000.ckpt | 331.35 |
qm9/QH9Stable-size_ood |
weights-epoch=78-val_loss=0.0000000.ckpt | 331.35 |
rmd17/rmd-aspirin |
weights-epoch=39-val_loss=0.0000000.ckpt | 761.28 |
rmd17/rmd-naphthalene |
weights-epoch=39-val_loss=0.0000000.ckpt | 761.28 |
rmd17/rmd-salicylic_acid |
weights-epoch=39-val_loss=0.0000000.ckpt | 761.28 |
Download
from huggingface_hub import hf_hub_download
checkpoint = hf_hub_download(
repo_id="ksusu/QHFlow1",
filename="qm9/QH9Stable-random/weights-epoch=78-val_loss=0.0000000.ckpt",
)
Use the original QHFlow architecture and dataset-specific configuration to load these checkpoints. This mirror does not convert them to QHFlow2 or provide new evaluation results.
Provenance and integrity
checkpoint-manifest.json records every source
Drive file ID, URL, byte count, and SHA-256 digest.
SHA256SUMS provides the checkpoint checksums.
Downloaded file sizes were checked against Drive metadata and all PyTorch ZIP
archive CRCs were verified without deserializing the checkpoints. Uploaded
weights are verified against the Hugging Face LFS SHA-256 digests.
The source folder contains checkpoint files only; it supplies no separate license document or configuration files. This mirror adds no new license terms.