MultiMFM

Checkpoints and data for Multimodal Meta Flow Maps: inference-time steering of pretrained generators, for molecules and DNA.

Code · Project page

Download

The files are laid out like the code repository, so fetch them from inside a checkout:

git clone https://github.com/shiyiwang958/MultiMFM && cd MultiMFM
python scripts/download_artifacts.py            # everything, ~1.5 GB, checksum-verified
python scripts/download_artifacts.py --only qm9 # or just one of qm9, geom, dna

Contents

checkpoints/
  qm9/   base_model, mfm_student_ce_esd, property_regressors          47 MB
  geom/  base_model, mfm_student, fingerprint_head                   422 MB
  dna/   base_model, dmfm_student*, c0_regressors                    499 MB
data/
  dna/        loop-seq cyclizability windows                         385 MB
  geom/       GEOM-Drugs fingerprints (steering targets)             173 MB

Each folder also holds the arguments its model was trained with. What each model does is in the code repository's per-experiment guides (docs/).

License

MIT. Uses TFG-Flow (Lin et al., ICLR 2025), GEOM-Drugs (Axelrod & Gómez-Bombarelli, 2022; CC0) and loop-seq data from Basu et al. (Nature, 2021); see the code repository's THIRD_PARTY.md.