Dellboy/binman-lm-adapter

BINMAN-LM adapter

A LoRA adapter for Qwen/Qwen2.5-3B-Instruct, trained for BINMAN: an atlas of bridging ligands (molecular glues), structural degrons, E3 ligase triage and degradability built from the PDB.

Try it: Dellboy/binman-lm

Current weights: training round 07 — 32 LoRA layers at rank 8, 14,152 iterations. Best Task B macro-F1 measured so far (0.9336) and a perfect Task A set equality (1.000). A further round combining 32 layers with rank 32 is still training; this repository is updated when a round measures better.

What it does, and what it must never do

Three text jobs:

1. Query translation — a question about the atlas into a query object.
2. Evidence triage — a bridging ligand into one of four evidence classes: crystallisation_artefact, molecular_glue, native_cofactor, protac.
3. Abstention — refusing, in structured form, when the atlas schema cannot answer the question.

It never computes, estimates or reports a number. Every figure in BINMAN (ΔSASA, Cβ–Cβ distances, pocket scores, pLDDT) is computed deterministically in Python over the atlas and passed to the interface. A number in this model's output that was not copied verbatim from a retrieved record is a defect, not a feature. That is why abstention is a measured task rather than a disclaimer.

Training

Base	Qwen/Qwen2.5-3B-Instruct
Trained against	mlx-community/Qwen2.5-3B-Instruct-4bit (4-bit)
Method	LoRA, rank 8, scale 20 (PEFT lora_alpha 160), 32 layers (4-35)
Iterations	14,152 at batch 4, two epochs over 28,304 examples
Targets	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Framework	mlx-lm on Apple silicon, converted to PEFT

Measured on a class-balanced 240-sample triage set (60 per class, fixed seed):

Metric	Value
Task B macro-F1	0.9336
molecular_glue F1 / recall	0.958 / 0.950
protac F1	0.975
Task A set equality	1.000
Task C abstention / fabrication	1.00 / 0.00

Metrics are published in FINDINGS.md with their n and their method, including the full Task B confusion matrix rather than the macro-F1 alone: the model's errors concentrate in the glue-against-PROTAC cell, which is the pair worth knowing about.

Known limitation: quantisation transfer

The adapter was trained to correct a 4-bit quantised base and is served against a 16-bit one. Nothing guarantees the correction transfers unchanged. lm/export_hf.py --verify in the repository runs the real held-out test questions through the converted adapter and prints the same metrics the MLX evaluation reports, so the two can be compared rather than assumed equal. Read the published numbers as measured under MLX unless stated otherwise.

Intended use and scope

Research and evaluation on the BINMAN atlas schema. It is a 3B model fine-tuned on one project's query language: it is not a general chemistry assistant, it has no knowledge of binding affinities, assay results or clinical status, and it should abstain when asked for them. Treat any confident-sounding chemistry claim outside the three tasks above as unverified.

Licence

The base model is Qwen/Qwen2.5-3B-Instruct under the Qwen Research License Agreement, which permits use, modification and redistribution for non-commercial purposes only. This adapter is a modification within the meaning of section 3 of that Agreement and is redistributed under the same terms. LICENSE and NOTICE in this repository carry the Agreement and the required attribution.

Training data derives from BioLiP2, MGDB, MolGlueDB, MGTbind and PROTAC-DB, all cited with DOIs in the repository README and in data/validation/MANIFEST.md. No source dataset is redistributed here. PROTAC-DB's own terms permit internal use including derivatives and prohibit redistribution; publishing these weights was a decision the author made explicitly, recorded with its reasoning as DECISIONS.md D-030. Anyone reusing this adapter should read that entry and form their own view.

Citation

@software{deller_binman,
  author = {Deller, Marc C.},
  title  = {BINMAN: Blind-spot INventory of Molecular Adhesives and Neosubstrates},
  url    = {https://github.com/bellcheddar/BINMAN}
}