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

Check out the documentation for more information.

MolMistral β€” DPP-4 Fine-Tuned

DPP-4-specific fine-tune of the MolMistral SMILES foundation model (12 layers, hidden 512, 21.8M parameters, vocab 621, max_len 128). Causal language model over SMILES with conditional generation via 7 property control tokens.

Provenance

  • Base model: Delower/molmistral-foundation-v61 β€” the verified foundation trained by SPARCq job 822 (Cheaha, A100 80GB) on 5,441,246 decontaminated SMILES (ChEMBL / MOSES / BindingDB / PubChem / ZINC). Base weights sha256 2076717dee20df2bf9e04985a0fbaef9aa93801b242ba6e6d1ffdb85b938e802, verified in-notebook before training.
  • Fine-tune run: Kaggle notebook delower2021/molmistral-finetune-potent, Version 1, GPU T4 x2, 2026-11-17. Training code: train_llm_foundation.py (shipped in this bundle; the fine-tune stage only β€” the foundation was trained unconditionally).

Fine-tuning data

  • DPP-4 POTENT ONLY β€” 5,747 compounds (pIC50 β‰₯ 7.0), split 5,173 train / 574 validation.
  • No moderate compounds. No FDA/regulatory gliptins (the 12 are excluded from this fine-tune by design).
  • Conditional training: each sequence is [BOS] [7 property tokens] SMILES [EOS] β€” [POTENT] plus MW / LOGP / HBD / HBA / TPSA / ROT bins computed per molecule β€” with condition dropout 0.15, so the model also works unconditionally.
  • Batch 128, max_len 128, LR 5e-5, epoch ceiling 120, patience 25, target val_loss 0.20 (early stop on target).

Training result

Stopped at epoch 3 when the validation loss reached the 0.20 target:

epoch train loss val loss
1 0.3428 0.2412
2 0.2090 0.2016
3 0.1841 0.1973 (best)

Best val_loss 0.1973167 (perplexity 1.218) at epoch 3. best_finetune_dpp4.keras is the best-val checkpoint. Full history: history_finetune.csv; machine-readable cards: model_card_finetune.json, finetune_metrics_summary.json.

⚠️ Two-weights rule β€” read before using

These weights are DPP-4-specific. Never use them as the base for fine-tuning on another disease or target. For any other disease, start from the disease-free foundation: Delower/molmistral-foundation-v61. The companion scripts enforce this β€” infer_finetuned.py and generate_finetuned.py refuse any bundle that does not verify as the fine-tuned one.

Usage

Inference / scoring check (verifies the fine-tuned role, tokenizer round-trip, and teacher-forced loss/perplexity on built-in drug SMILES):

python infer_finetuned.py --bundle .
python infer_finetuned.py --bundle . --csv my_smiles.csv --out scores.csv

Conditional generation (default: 500 molecules at each of T = 0.8, 1.0, 1.1; raw, unfiltered SMILES β€” RDKit validation is a separate downstream stage):

python generate_finetuned.py --bundle . --out dpp4_generated_raw.csv
python generate_finetuned.py --bundle . --no_condition --n_per_arm 100

Default condition (the fine-tune set's modal property profile):

POTENT,MW_400_450,LOGP_2_3,HBD_1,HBA_5_6,TPSA_60_90,ROT_4_6

Files

file contents
best_finetune_dpp4.keras fine-tuned weights (Keras 3, whole model)
vocab.txt, tokenizer_config.json 621-token SMILES vocabulary
model_card_finetune.json trainer's model card
finetune_metrics_summary.json run configuration + result summary
history_finetune.csv per-epoch train/val loss
gate_scan_finetuned.json non-finite tensor scan of the weights (0 NaN/Inf)
inference_check.txt output of the post-assembly inference check
infer_finetuned.py scoring / inference check script
generate_finetuned.py conditional generation script
train_llm_foundation.py training/tokenizer code the scripts import
Downloads last month
12
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support