chem_sage_32b_v3

ChemSage 32B v2 — Round 3 fused model

ChemSage is a QLoRA-tuned Qwen2.5-32B chemistry assistant for drug discovery, fine-tuned on Apple Silicon with MLX-LM. It reasons about small molecules, emits verified RDKit and PyMOL tool calls, and is grounded by a RAG layer over a curated cheminformatics corpus.

This is the Round 3 fused model — the current production checkpoint.

Model details

Evaluation (5-round comparative, 2026-06-29)

Evaluated on 100 shared R5 test examples (seed=42) via eval/compare/eval_compare.py. Scores = examples where all instances correct / 100 examples (per-example pass/fail). Full report: eval/compare/results/compare_20260629_1928.html.

Original scorecard (500 examples, eval/eval_chem.py): SMILES 99%, exec 99%, fidelity 68% (overall 89%).

Usage

# Serve the model (OpenAI-compatible endpoint)
mlx_lm.server --model models/chem_sage_32b_v3 --port 8081

# Or use the interactive chat CLI (RAG + RDKit auto-execution)
.venv/bin/python scripts/chat.py --model models/chem_sage_32b_v3

What it does well

• Emits correct, executable RDKit code for every physicochemical property query
• 99% SMILES validity — never produces malformed SMILES
• Drives PyMOL command scripts for structural visualisation
• Interprets PLIP interaction fingerprints in medchem terms
• Grounds answers in the ChemSage corpus (PDB, ChEMBL, PubChem, approved drugs)
• Refuses to invent PDB IDs, measured values, or compound names not in context

Training history

Round 4 training completed 2026-06-27 (1,500 iters, ~43 h). Best val loss 0.041 at iter 950 beats this model's 0.054 by 24%. Fused model saved to models/chem_sage_32b_v4/.

Use eval/eval_chem_original.py to reproduce the original 3-metric scorecard (500 examples).

Compatibility note (2026-06-29)

tokenizer_config.json updated: extra_special_tokens converted from list to dict to fix mlx_lm 0.31.x / transformers compatibility (AttributeError on load in older format).

Built by

Marc C. Deller, D.Phil. · marcdeller.com · [email protected]