Drug Approvals GLiNER

GLiNER fine-tune for extracting disease and phenotype mentions from FAERS and DailyMed drug indication and contraindication text. Training pipeline and evaluation code: SkyeAv/MedliNER.

Model

• Base: urchade/gliner_small-v2.1
• Labels: disease, phenotype
• Tasks: drug indications and contraindications. Indication vs contraindication is context metadata, not an entity label. The model extracts the condition span in either context.
• Training data: reviewed FAERS and DailyMed-derived text, augmented with weighted synthetic and model-suggested examples

Usage

from gliner import GLiNER

model = GLiNER.from_pretrained("SkyeAv/drug-approvals-gliner-small-v2.1")

text = """Lisinopril is indicated for the treatment of hypertension and chronic heart failure.
It is contraindicated in patients with a history of angioedema or severe renal impairment."""

indication, contraindication = text.splitlines()

for context, sentence in [("indication", indication), ("contraindication", contraindication)]:
    for e in model.predict_entities(sentence, ["disease", "phenotype"], threshold=0.5):
        print(context, e)

Output (real captured output from the published checkpoint):

indication {'start': 45, 'end': 57, 'text': 'hypertension', 'label': 'disease', 'score': 0.9998974800109863}
indication {'start': 62, 'end': 83, 'text': 'chronic heart failure', 'label': 'disease', 'score': 0.99992835521698}
contraindication {'start': 52, 'end': 62, 'text': 'angioedema', 'label': 'disease', 'score': 0.99937903881073}
contraindication {'start': 66, 'end': 89, 'text': 'severe renal impairment', 'label': 'disease', 'score': 0.9927625060081482}

Indication and contraindication are context, not entity labels. The model extracts the condition span in either context, so run it on the relevant sentence or section and keep that context with each returned span. For example, spans from the first sentence are indications, while spans from the second sentence are contraindications. This is how MedliNER and DAKP use the model in production.

Limitations

This model is an information-extraction component, not a clinical decision-support system. Predictions require review and should not be used as medical advice. Training and evaluation data are domain-specific; performance may vary across documents, labels, and clinical contexts.

License and attribution

Apache-2.0. This model is fine-tuned from urchade/gliner_small-v2.1; users must also review and comply with the base model's license and terms. FAERS and DailyMed are external source datasets with their own terms and attribution requirements.