Feline-NER

Named Entity Recognition model for feline veterinary medicine, trained on 1,300 annotated sentences from PubMed literature.

Model Description

Feline-NER is a token classification model fine-tuned for extracting clinical entities from feline veterinary scientific literature.

Model lineage:

• Base: BERT (Devlin et al., 2019)
• Domain-adapted: BioBERT v1.2 (Lee et al., 2019) - biomedical literature
• Further adapted: Feline-BERT - 11,830 feline PubMed articles (MLM fine-tuning)
• Task-specific: Feline-NER - 1,300 manually annotated sentences (token classification)

This model is intended solely for research and educational use.

Task

Span-level named entity recognition using a BIO tagging scheme over five entity types:

• DISEASE
• SYMPTOM
• MEDICATION
• PROCEDURE
• ANATOMY

Training Data

Fine-tuned on a manually annotated dataset of 1,300 sentences extracted from feline-related PubMed Central articles. Annotations were produced by a non-veterinary researcher using an iterative human-in-the-loop workflow with LLM-assisted pre-labeling.

Evaluation

• Macro F1: ~0.65
• Micro F1: ~0.64
  (Evaluated on a 150-sentence held-out test set)

Performance varies by entity type; PROCEDURE and ANATOMY remain challenging due to boundary ambiguity.

Intended Use

• Veterinary NLP research
• Information extraction from feline scientific literature
• Educational demonstrations

Usage

This model can be used with the Hugging Face pipeline API:

from transformers import pipeline
ner = pipeline("ner", model="Statistical-Impossibility/Feline-NER", aggregation_strategy="simple")
print(ner("The cat was diagnosed with FIV and treated with prednisolone."))

⚠️ Limitations & Warnings

• NOT FOR CLINICAL USE
• Not validated for diagnosis or treatment
• Annotation noise and boundary ambiguity are present
• Single-annotator dataset

Author

Statistical-Impossibility

Project repository: Feline-Project