YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

foundation β€” model bundle

Raw small-molecule foundation model (causal LM over SMILES). Pretrained on general molecules only β€” NO DPP-4 data, no property conditioning. Decontaminated against the DPP-4 fine-tune set and the 12 FDA gliptins.

Use it for: Fine-tune this on a new target/disease, or sample general drug-like SMILES from it.

Files

File What it is
model.keras the weights (whole Keras 3 model), best by val loss
vocab.txt 621 tokens in id order
tokenizer_config.json ids, property tokens, bins, max lengths
model_config.json architecture, provenance, dependencies
history.csv per-epoch loss / accuracy / perplexity
metrics.json stage metrics (if the stage produced them)
infer_from_bundle.py generate SMILES from this bundle
finetune_from_foundation.py fine-tune this bundle on a new target
train_llm_foundation.py module the two helpers import
MANIFEST.json sha256 + size of every file here

Inference

pip install "tensorflow==2.17.0" "keras==3.5.0" "keras-nlp==0.15.1" rdkit pandas
python infer_from_bundle.py --bundle . --n 1000 --out generated.csv

Add --condition POTENT,MW_350_400,LOGP_1_2,HBD_1,HBA_5_6,TPSA_90_120,ROT_4_6 for a property-conditioned run (fine-tuned bundle only; the foundation model ignores conditioning because it never saw those tokens).

Fine-tuning this on another disease

Only the foundation bundle should be the starting point for a new target β€” starting from finetuned_dpp4 would carry DPP-4 bias into the new target.

python finetune_from_foundation.py \
    --bundle /path/to/foundation \
    --csv my_target_actives.csv \      # columns: smiles[, pIC50][, source]
    --out_dir ./runs_mytarget \
    --epochs 30 --patience 4

It reuses this bundle's tokenizer unchanged (same 621-token vocab, same property tokens), so no vocabulary surgery is needed, and it writes a new bundle at ./runs_mytarget/model_weights/finetuned_<name>/ in this same layout.

Hard rule carried over from the DPP-4 run: keep the new target's actives out of any pretraining data, and keep reference/approved drugs in the train split so the model actually learns them.

Downloads last month
17
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support