BINMAN-LM
The small language model behind BINMAN, an atlas of bridging ligands (molecular glues), structural degrons, E3 ligase triage and degradability built from the PDB.
Qwen2.5-3B-Instruct with a LoRA adapter, trained on Apple silicon with
mlx-lm and converted to run here. It does three text jobs and no
arithmetic: every number in BINMAN is computed deterministically in Python
over the atlas and passed to the interface. A figure in this model's output
that was not copied verbatim from a retrieved record is a defect, not a
feature.
The atlas is not served here: it is 142 MB and several source datasets carry licences that do not permit redistribution.
Natural language to a query object against the atlas schema. Parse rate and set equality against held-out questions are reported in the repository's FINDINGS.md.
Classify a bridging ligand into one of four evidence classes. The confusable pair is molecular glue against PROTAC, and the repository reports the full confusion matrix rather than only the macro-F1.
Checks whether ZeroGPU grants this Space a GPU at all.
Questions the atlas schema cannot answer. The model should refuse rather than invent a field: a fabricated number is the failure this task exists to measure.