fabsssss/qwen3-coder-30b-a3b-bio

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:

1. Biolink Model — converts grounded biomedical statements into Biolink-conformant RDF/Turtle knowledge graphs (typed entities, valid predicates, reified associations).
2. GO-CAM — constructs Gene Ontology Causal Activity Model YAML (activities, molecular functions, biological-process context, Relation-Ontology causal edges).
3. 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.

Dataset: fabsssss/bio-ontology-instruct