Qwen3-Coder-30B-A3B — Biology Ontology Edition (Biolink · GO-CAM · OBO)
A QLoRA fine-tune of Qwen3-Coder-30B-A3B-Instruct (8-bit MLX) that emits ontology-conformant
biomedical knowledge structures across three open biology standards:
- Biolink Model — converts grounded biomedical statements into Biolink-conformant RDF/Turtle
knowledge graphs (typed entities, valid predicates, reified associations).
- GO-CAM — constructs Gene Ontology Causal Activity Model YAML (activities, molecular
functions, biological-process context, Relation-Ontology causal edges).
- OBO grounding — normalises biomedical mentions to the correct ontology CURIE (candidate-based
entity linking over GO / MONDO / HP / ChEBI / CL), selecting only from the provided real
candidates rather than emitting a free-form identifier (100% accuracy on held-out mentions).
This is, to our knowledge, the first openly published LLM fine-tuned specifically for
Biolink-Model + GO-CAM conformance. It is the biology counterpart of
fabsssss/qwen3-coder-30b-a3b-ies4.
Why biology ontology conformance
Biomedical knowledge is fragmented across hundreds of incompatible databases. The open
ontologies — Biolink Model (the schema behind the Monarch Initiative and the NCATS Biomedical
Data Translator), the Gene Ontology / GO-CAM, and the OBO Foundry (GO, MONDO, HP, ChEBI,
Cell Ontology) — are the interoperability backbone that makes that knowledge FAIR: findable,
accessible, interoperable, reusable.
LLMs are fast becoming the default interface to that knowledge — but off the shelf they hallucinate
ontology terms. Measured on this model's own held-out set, the base model produced Biolink graphs
with a 41% hallucinated-term rate and 0% conformance: plausible-looking predicates and CURIEs
that do not exist, non-resolvable identifiers, categories that violate the schema. In a knowledge
graph, that is not a cosmetic error — a single fabricated term silently breaks interoperability,
poisons downstream queries, and is indistinguishable from a real edge until something fails.
A model that is conformant by construction flips the LLM from a FAIR liability into a FAIR
asset. Here conformance is not hoped for — it is enforced by a validator in the loop that
rejects 100% of out-of-ontology terms, driving the hallucinated-term rate to 0.000 and term
conformance to 100%, and holding on graph shapes never seen in training. The model proposes;
the validator disposes. That is the only honest way to put a language model near a scientific
knowledge graph.
The three tasks below cover the real work: constructing Biolink knowledge graphs, building GO-CAM
causal activity models, and normalising free-text mentions to the correct ontology identifier —
the structuring, harmonisation, and grounding that every large biomedical data programme needs.
How it was built (correct-by-construction)
Every training target is generated programmatically from real ontology releases and validated
twice — once by the official toolkit, once by an independent membership/conformance validator:
| Slice | Gold built with | Validated by |
|---|
| Biolink | biolink-model pydantic classes + deterministic RDF emitter | bmt (Biolink Model Toolkit): category/predicate membership, subject/object range closure |
| GO-CAM | gocam pydantic data model → LinkML YAML | gocam schema round-trip + real GO(MF/BP/CC)/RO/UniProt/taxon CURIEs |
| OBO grounding | pronto over GO, MONDO, HP, ChEBI, CL releases | term membership + exact label match |
Gene/protein identifiers (NCBIGene, UniProt) were verified against mygene.info. Entities are
real CURIEs from: GO (38,245), MONDO (32,095), HP (19,836), ChEBI (218,253), CL (3,335).
Evaluation (held-out; baseline → fine-tuned)
Same-prompt comparison of the base model vs. this fine-tune on held-out test graphs, scored by the
conformance validators. OOD = multi-association Biolink graphs, a shape never seen in training.
Biolink Turtle (in-distribution)
| Metric | Base | Tuned |
|---|
| term conformance % | 0.0 | 100.0 |
| hallucinated-term rate | 0.418 | 0.0 |
| structural conformance | 0.0 | 1.0 |
| syntactic validity % | 81.7 | 100.0 |
Biolink Turtle (out-of-distribution, multi-association)
| Metric | Base | Tuned |
|---|
| term conformance % | 0.0 | 88.9 |
| hallucinated-term rate | 0.259 | 0.038 |
| structural conformance | 0.0 | 0.846 |
GO-CAM schema validity %: base 0.0 → tuned 100.0
OBO grounding accuracy %: base 0.0 → tuned 100.0
The conformance gate is the validator, not the model
Conformance is enforced by a validator in the loop, which rejects 100% of out-of-ontology
terms and fabricated CURIEs (Biolink Model Toolkit membership + subject/object range closure;
gocam schema; OBO membership). The fine-tuned model is the proposer; the validator is the
gate. Consistent with this design, the model does not reliably self-refuse malformed requests
(e.g. a nonsensical predicate) — refusal rate is low and is a documented limitation. Safety
comes from the symbolic validator that catches every invalid term, not from model self-restraint.
Always run outputs through bmt / gocam / SHACL before downstream use.
Intended use & limits
- Use: structuring given, grounded biomedical statements into Biolink/GO-CAM; entity
normalisation with candidate lists. A structuring aid, not a source of biomedical truth.
- Not for: asserting novel biological facts, clinical decisions, or recalling identifiers
without a candidate set. Associations in training are schema-correct but not curated biological
claims. Always validate outputs with
bmt / gocam / SHACL before downstream use.
Prompt format
System prompt sets the task (Biolink / GO-CAM / normalisation); the user turn supplies the
statement and the grounded entity CURIEs. See the dataset card for exact templates.
