Scaffold-First Diffusion — checkpoint registry

Final checkpoints for Scaffold-First Diffusion (SFD). Each checkpoint is self-contained (its Lightning hyper_parameters.config holds the full architecture); a per-dataset stats/<family>.stats.pt blob (~1 KB) supplies the tokenizer scalars. Dataset families present: moses, planar, qm9, sbm.

Load and sample

from sfd.hub import from_pretrained
model = from_pretrained("qm9_riskopt_motif", repo_id="EliasHossain/scaffold-first-diffusion")   # or set SFD_HF_REPO
from sfd.sampling import ScheduleAwareConfidenceSampler
rows = ScheduleAwareConfidenceSampler(model, num_steps=128, temperature=0.9).sample(100, 32)

Or from the CLI:

export SFD_HF_REPO=EliasHossain/scaffold-first-diffusion
python -m sfd.cli.sample --model qm9_riskopt_motif --n 100 --out smiles.txt

Models

model	dataset	size
moses_full_motif_riskopt	moses	37 MB
moses_full_uniform	moses	37 MB
moses_off	moses	37 MB
moses_sfd_motif	moses	37 MB
off_full_s42	qm9	37 MB
planar_uniform	planar	38 MB
qm9_invfreq	qm9	37 MB
qm9_random_role	qm9	37 MB
qm9_riskopt_motif	qm9	37 MB
qm9_riskopt_noexp	qm9	37 MB
sbm_uniform	sbm	41 MB
sfd_motif_full_s42	qm9	37 MB

Uploaded by scripts/upload_to_hf.py.