← Research

Omi-Sum 3B: Open-Source Model for Medical Summaries

Open clinical model

A Phi-3 based 3B model fine-tuned on synthetic medical dialogue to produce structured SOAP notes.

MIT · 3B params
Figure: Omi-Sum converts medical dialogue into SOAP structure using a compact open model. The benchmark table below shows ROUGE-1 results.

We're excited to announce the release of Omi-Sum (3B) Small, a compact yet powerful language model designed to turn medical dialogues into structured SOAP summaries. Omi-Sum is openly available on HuggingFace and has shown higher ROUGE-1 scores than GPT-4 Turbo on our summarization benchmark.

Omi-Sum was fine-tuned on our synthetic medical-dialogue-to-soap-summary dataset (10,000 examples) using Microsoft's Phi-3-mini-4k-instruct as a base. The model, dataset, and training code are released under the MIT licence to encourage adoption and collaboration.

Parameters
3 billion
Base model
Phi-3 Mini 4k Instruct
Training data
10,000 synthetic examples
Licence
MIT

Benchmark results (ROUGE-1 on test set)

ROUGE-1 summarization benchmark on the Omi test set. Higher is better.
Model ROUGE-1
Omi-Sum 3B Small (Omi Health)70
GPT-4 Turbo (OpenAI)69
Llama-3 8B Instruct (Meta)59
Phi-3 Mini 4k Instruct — base (Microsoft)55
GPT-3.5 Turbo (OpenAI)54
Phi-2 — base (Microsoft)41

Omi-Sum is designed for research and development of AI-powered medical documentation tools. While it is not yet approved for clinical use, we believe this open-source release is a step towards safer, more transparent AI for healthcare.

Where to find it

We look forward to seeing how the community uses and improves this model.

Cite this model

APA — Omi Health. (2024). Omi-Sum 3B: Open-Source Model for Medical Summaries. https://omi.health/research/omi-sum

@misc{omi_sum_3b_2024,
  title   = {Omi-Sum 3B: Open-Source Model for Medical Summaries},
  author  = {{Omi Health}},
  year    = {2024},
  url     = {https://omi.health/research/omi-sum},
  note    = {3B parameter clinical SOAP model, fine-tuned from Phi-3, MIT licence}
}

Related research