nanofold-smiles

Character-level transformer models trained on ZINC-250K for drug-like molecule generation.

Best model: L6-E80-modern — 468K params, 97.4% valid SMILES.

See the experiment report and source code.

Usage

from nanofold.smiles.model import load_checkpoint
from nanofold.smiles.sample import generate

model, tokenizer = load_checkpoint("p2-l6-e80-modern.pt")
molecules = generate(model, tokenizer, n=100, temperature=0.8)

Checkpoints

File	Config	Params	Validity
p2-l6-e80-modern.pt	6L, 80d, SwiGLU+RoPE	468K	97.4%
p2-l8-e80-modern.pt	8L, 80d, SwiGLU+RoPE	622K	98.0%
p1-l4h4e256.pt	4L, 256d, baseline	3.2M	98.8%

All 24 checkpoints from the scaling experiment are included.