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 5 fused model — the current production checkpoint, fused from iter 1600 (best available checkpoint; val loss 0.055).

Model details


Base model	`mlx-community/Qwen2.5-32B-Instruct-4bit`
Fine-tune method	QLoRA (RSLoRA, rank 64, scale 90.0)
Adapted layers	32 of 64 transformer layers
Trainable parameters	268M (0.82% of 32.8B) — double R4
Training examples	16,000 (from 20,000 total; 80/10/10 split)
Behaviour classes	78 (19 new vs R4)
Training iters	2,000 of 3,000 (early stop — val plateau at 0.055 from iter 1600)
Fused checkpoint	iter 1600 (best saved; true best 0.054 at iters 1650/1950 fell between saves)
Best val loss	0.055
Final train loss	0.054
Peak memory	30.3 GB
Fused model size	~17.2 GB
Training duration	~37 h (2026-06-27 to 2026-06-29)

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.

Metric	R1	R2	R3	R4	R5 (this)
SMILES validity	100%	100%	100%	100%	100%
SMARTS validity	N/A	0%	100%	55%	100%
Tool executability	0%	28%	79%	71%	95%
Code attempted	14%	41%	100%	97%	100%
Python extended	36%	34%	78%	69%	99%
Code-then-quote	N/A	0%	47%	19%	61%
Numerical fidelity	N/A	18%	57%	47%	89%
Rounding precision	100%	100%	98%	98%	99%
Refusal accuracy	98%	98%	97%	98%	100%
QED range	100%	100%	100%	100%	100%
PDB ID validity	100%	100%	100%	100%	100%
PyMOL syntax	77%	97%	89%	89%	90%
Degeneration-free	93%	96%	91%	98%	100%
Overall	72%	63%	87%	80%	95%

Usage

bash

# Download from HuggingFace
huggingface-cli download Dellboy/chem_sage_32b_v5 --local-dir models/chem_sage_32b_v5

mlx_lm.server --model models/chem_sage_32b_v5 --port 8081
.venv/bin/python scripts/chat.py --model models/chem_sage_32b_v5

Round 5 additions (78 behaviour classes, 19 new vs R4)

New generators targeting R4 exec/fidelity gaps and breadth:

Category	New classes
Exec drills	`pyexec_drill` (×6), `code_then_quote_v2` (×5), `rounding_explicit` (×5), `fidelity_multistep` (×5)
ADMET / SAR	`herg_liability`, `selectivity_profile`, `prodrug_bcs`, `sar_delta`
Cheminformatics	`mdanalysis`, `conformer_3d`, `mcs_search`, `reaction_smarts`, `recap_fragmentation`
Structural biology	`dssp`, `ppi_interface`, `electron_density`, `biological_assembly`
Drug targets	`drug_target_family` (kinases, GPCRs, NHRs)
External APIs	`uniprot_api`

Training history

Round	Val loss	Improvement
Round 1 (7B)	0.389	—
Round 2 (32B)	0.347	10.8% over R1
Round 3 (32B)	0.054	6.4x over R2
Round 4 (32B)	0.041	24% over R3
Round 5 (32B, this model)	0.055	early stop; harder task (78 classes, 20k examples)

R5's higher val floor (0.055 vs R4's 0.041) reflects the harder task: 78 behaviour classes vs 59, 20k examples vs 6.4k, rank=64 adapter. The model is more capable but the task is larger.

Built by

Marc C. Deller, D.Phil. · marcdeller.com · marc@marcdeller.com