README.md

metadata

license: mit
library_name: pytorch
tags:
  - diffusion
  - flow-matching
  - molecular-generation
  - structure-based-drug-design
  - reinforcement-learning
  - dpo

MolFORM (checkpoints + sampling outputs)

This Hugging Face repository hosts released PyTorch .pt checkpoints and sampling outputs for MolFORM (Multi-modal Flow Matching for Structure-Based Drug Design, arXiv:2507.05503).

• Hub repo: daiheng/MolFORM
• Paper: https://arxiv.org/abs/2507.05503
• Codebase: https://github.com/daiheng-zhang/SBDD-MolFORM

What’s included

Checkpoints

• MolFORM-RL (NFT-Vina-SA): checkpoints/molform-rl/final.pt (~32MB) \
  • sha256: 1ccada153b1ba3d92f8d574e5e82c43b692b04bf2d6a817818cd473ab4381fab
• MolFORM-DPO: checkpoints/molform-dpo/best_vina_21500.pt (~33MB) \
  • sha256: 61031cf18b3faf20dcd9490d40813a4687071468fc6db09b42ff4c8692762900

Sampling outputs

• MolFORM-RL sampling (500 steps, 100 samples):
  sampling/sampling_nft_500steps_vinasa_final_job1879195/ (~4.4GB, 206 files)
  Includes:
  • result_*.pt: per-shard sampling results (100 files)
  • sample.yml: sampling config (seed=2021, num_steps=500)
  • eval_results/: evaluation artifacts (metrics, plots, example pickles)

Note: sample.yml was generated on our cluster and contains an absolute checkpoint path; update model.checkpoint to your local path after downloading.

Download

CLI:

huggingface-cli download daiheng/MolFORM checkpoints/molform-rl/final.pt --local-dir .
huggingface-cli download daiheng/MolFORM checkpoints/molform-dpo/best_vina_21500.pt --local-dir .
huggingface-cli download daiheng/MolFORM sampling/sampling_nft_500steps_vinasa_final_job1879195/sample.yml --local-dir .

Python:

from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download("daiheng/MolFORM", "checkpoints/molform-rl/final.pt")

PyTorch >= 2.6 note (important)

These .pt files were saved with torch.save and include Python objects (e.g., EasyDict configs).
In PyTorch 2.6+, torch.load defaults to weights_only=True, which can fail on these files.

If you trust the checkpoint source, load with:

import torch
ckpt = torch.load("final.pt", map_location="cpu", weights_only=False)

For the recommended environment used by the codebase, see the conda env in the code repository (PyTorch 1.10.x).

Citation

If you use MolFORM, please cite the paper (see arXiv page above).