OpenBioLLM 8B

A biomedical large language model built on Meta’s Llama 3 8B, fine-tuned by Saama AI Labs for medical and clinical question answering. OpenBioLLM is optimized for the biomedical domain and achieves strong results on standard medical benchmarks for its size class.

🚀 Overview

openbiollm packages aaditya/Llama3-OpenBioLLM-8B, an open biomedical language model fine-tuned from Meta-Llama-3-8B. It was developed by Ankit Pal (Aaditya Ura) at Saama AI Labs and trained to specialize in medical, clinical, and biomedical reasoning while retaining the general instruction-following of its Llama 3 base.

Base model: Meta Llama 3 8B (~8B parameters)
Context length: 8K tokens
Domain: Biomedical / clinical question answering
License: Meta Llama 3 Community License

📊 Benchmarks

Scores reported by the authors across nine biomedical datasets (higher is better). OpenBioLLM-8B is compared against larger proprietary and open models.

Dataset	OpenBioLLM-8B	GPT-3.5 Turbo	GPT-4	Med-PaLM-2
Clinical KG	76.10	74.71	86.04	88.3
Medical Genetics	86.10	74.00	91.00	90.0
Anatomy	69.83	72.79	80.00	77.8
Pro Medicine	78.21	72.79	93.01	95.2
College Biology	84.21	72.91	95.14	94.4
College Medicine	68.04	64.73	76.88	80.9
MedQA (4 opts)	58.99	57.71	78.87	79.7
PubMedQA	74.12	72.66	75.20	79.2
MedMCQA	56.91	53.79	69.52	71.3
Average	72.50	66.00	82.85	84.08

Per the authors, OpenBioLLM-8B “demonstrates superior performance compared to larger models, such as GPT-3.5, Meditron-70B across 9 diverse biomedical datasets.” Benchmark numbers are from the official model card; see the HuggingFace page for full methodology.

🎯 Key Features

Biomedical specialization, fine-tuned for medical, clinical, and life-sciences question answering.
Strong for its size, competitive with much larger models on standard medical benchmarks.
Llama 3 foundation, inherits the reasoning and instruction-following of Meta-Llama-3-8B.
8K context, handles typical clinical notes, abstracts, and Q&A prompts.
Drop-in Ollama model, runs locally with a single ollama run command.

🏷️ Available Versions

Tag	Size	Context	Notes
`latest`	8.5GB	8K	Q8_0 quantization, highest quality

BPW Reference Guide

Quant	BPW	Notes
IQ3_M	3.66	Smallest, lowest quality
IQ4_XS	4.25	Compact, good quality
Q4_K_M	4.85	Recommended default
Q5_K_M	5.68	Higher quality
Q6_K	6.56	Near-lossless
Q8_0	8.5	Highest quality

💻 Quick Start

CLI

ollama run richardyoung/openbiollm

cURL

curl http://localhost:11434/api/chat -d '{
  "model": "richardyoung/openbiollm",
  "messages": [
    { "role": "user", "content": "What is the mechanism of action of metformin?" }
  ]
}'

Python

import ollama

response = ollama.chat(
    model="richardyoung/openbiollm",
    messages=[{"role": "user", "content": "What is the mechanism of action of metformin?"}],
)
print(response["message"]["content"])

JavaScript

import ollama from "ollama";

const response = await ollama.chat({
  model: "richardyoung/openbiollm",
  messages: [{ role: "user", content: "What is the mechanism of action of metformin?" }],
});
console.log(response.message.content);

🛠️ Use Cases

Biomedical research, literature exploration, hypothesis generation, and concept lookup.
Medical education, study aid for students reviewing clinical and life-sciences material.
Clinical NLP prototyping, entity extraction, summarization, and Q&A over medical text.
Benchmarking, a strong open baseline for evaluating biomedical LLMs at the 8B scale.

All use cases are for research and educational purposes only, see the Medical Disclaimer below.

📋 System Requirements

VRAM	Performance
10GB	Runs `latest` (Q8_0) with reduced context
12GB	Comfortable for `latest` at moderate context
16GB+	Full 8K context and higher throughput

CPU-only inference is possible but significantly slower; roughly 10GB of free system RAM is recommended for the Q8_0 quant.

🔧 Technical Details

Base Model: aaditya/Llama3-OpenBioLLM-8B (fine-tuned from Meta-Llama-3-8B)
Parameters: ~8B
Context Length: 8K tokens
Architecture: Auto-regressive transformer with Grouped-Query Attention (GQA)
Quantization: Q8_0 (8-bit)
Domain fine-tuning: Biomedical / clinical (Saama AI Labs)

⚠️ Medical Disclaimer

This model is for research and educational use only. It is NOT a medical device and does NOT provide medical advice.

OpenBioLLM is a language model and its outputs may be inaccurate, incomplete, or misleading. It must not be used for diagnosis, treatment, prescription, triage, or any clinical decision-making. Nothing it produces constitutes professional medical advice, and no doctor-patient relationship is created by its use.

Always consult a qualified healthcare professional for any medical concern. Do not rely on this model in emergencies, contact your local emergency services. The maintainers, base-model authors, and Meta accept no liability for any use of this model in healthcare or clinical settings.

🙏 Acknowledgments

Base model: Ankit Pal (Aaditya Ura) / Saama AI Labs, OpenBioLLM, fine-tuned from Meta Llama 3.
Foundation model: Meta, Llama 3 8B.
Quantization: llama.cpp.

Built & maintained by Richard Young · DeepNeuro

openbiollm

Models

Readme