University of Florida / NVIDIA · Healthcare

GatorTron

One of the largest clinical language models at 8.9 billion parameters, trained on over 90 billion words of clinical and biomedical text.

Overview

GatorTron is a large-scale clinical language model developed through a collaboration between the University of Florida and NVIDIA. Trained on over 90 billion words combining clinical notes from the UF Health system, PubMed articles, and Wikipedia, it represents one of the most ambitious efforts to build a comprehensive clinical NLP model. GatorTron achieves state-of-the-art results on five major clinical NLP tasks and demonstrates the benefits of scaling in the healthcare domain.

Parameters

8.9B

Architecture

BERT-style transformer (Megatron-LM)

Training Data

90B+ words (clinical notes, PubMed, Wikipedia)

Context Window

512 tokens

GPU Training

992 NVIDIA A100 GPUs

Capabilities

Clinical concept extraction

Medical relation extraction

Semantic textual similarity in clinical context

Natural language inference for clinical text

Medical question answering

Use Cases

Automating extraction of medical concepts from clinical notes at scale

Building clinical decision support systems

De-identifying patient records for research use

Supporting clinical trial matching from unstructured EHR data

Pros

  • +Massive scale captures nuanced clinical language patterns
  • +Trained on diverse clinical and biomedical corpora
  • +State-of-the-art on five clinical NLP benchmark tasks
  • +Developed with real-world clinical note data

Cons

  • -Very large model requiring significant compute resources
  • -Clinical training data from a single health system may introduce bias
  • -Complex deployment requiring specialized infrastructure
  • -Research-focused; not directly production-ready

Pricing

Open-source for research use. Self-hosting requires substantial GPU infrastructure. NVIDIA NeMo framework recommended for deployment.

Related Models