EpiBrief-MMWR-LM

EpiBrief-MMWR-LM: AI Epidemiologist

A specialized language model fine-tuned for CDC-style epidemiological reasoning and public health analysis.

Model Description

EpiBrief-MMWR-LM is a fine-tuned Llama 3.1 8B model trained on 11,632 instruction-response pairs derived from 9 years (2016-2025) of CDC Morbidity and Mortality Weekly Reports (MMWR). The model specializes in:

• 📊 Executive Summaries - CDC-style structured summaries with "What is already known?", "What is added?", and "What are the implications?" sections
• 📈 Data Interpretation - Quantitative analysis of surveillance data, outbreak metrics, and epidemiological trends (35% of training data)
• 🔬 Outbreak Analysis - Reasoning about disease transmission, risk factors, and outbreak dynamics
• 💡 Public Health Recommendations - Evidence-based guidance following CDC communication standards

Training Data

• Source: CDC MMWR Weekly Reports (2016-2025)
• Training Pairs: 11,632 high-quality instruction-response pairs
• Quantitative Focus: 85% of training data involves numerical reasoning (data interpretation, statistics, trends)
• Quality Control: Manually curated to ensure accuracy and CDC-style formatting

Task Distribution

• Executive Summaries: 25%
• Data Interpretation & Analysis: 35%
• Outbreak Investigation: 20%
• Public Health Implications: 20%

Training Details

• Base Model: meta-llama/Llama-3.1-8B
• Method: LoRA (Low-Rank Adaptation)
  • Rank: 32
  • Alpha: 64
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Infrastructure: Tinker AI distributed training platform
• Training Duration: ~8 hours (3 epochs)
• Final Checkpoint: Epoch 3

Usage

Installation

pip install transformers peft torch

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "BryanTegomoh/EpiBrief-MMWR-LM")

# Generate
prompt = """Based on this MMWR article excerpt, generate an executive summary following CDC format with 'What is already known about this topic?', 'What is added by this report?', and 'What are the implications for public health practice?' sections.

Measles cases have increased 300% in the United States during 2024, with 97% occurring among unvaccinated individuals. The outbreak was concentrated in communities with vaccination rates below 85%.

Generate the CDC-style executive summary:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=400, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Example Output

Input: "Explain the measles outbreak scenario with new SARS-CoV-2 variants"

Output:

Measles is a highly contagious vaccine-preventable viral disease. Increasing U.S. measles cases have been driven by unvaccinated persons who are exposed while traveling internationally. U.S. health officials should coordinate response activities to prevent and limit the spread, assess and improve vaccination coverage, and ensure MMR vaccination for all eligible children and adults.

Model Performance

The model demonstrates strong capabilities in:

• ✅ CDC-style formatting and structure
• ✅ Quantitative reasoning with epidemiological data
• ✅ Evidence-based public health recommendations
• ✅ Professional medical terminology usage

Limitations

• Domain-Specific: Optimized for epidemiology and public health; not a general-purpose model
• Training Data: Limited to MMWR reports (2016-2025); may not reflect future disease patterns or guidelines
• Not for Clinical Use: This model is for research and educational purposes only. Not intended for clinical decision-making
• Hallucination Risk: Like all LLMs, may generate plausible but incorrect information. Always verify with authoritative sources

Intended Use

✅ Appropriate Uses:

• Epidemiology education and training
• Draft generation for public health communications
• Research on AI in public health
• Understanding MMWR report structures

❌ Not Intended For:

• Clinical diagnosis or treatment decisions
• Real-time outbreak response without human oversight
• Replacing qualified epidemiologists or public health professionals

Citation

@model{tegomoh2025epibrief,
  author = {Bryan Tegomoh},
  title = {EpiBrief-MMWR-LM: A Specialized Language Model for CDC-Style Epidemiological Reasoning},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/BryanTegomoh/EpiBrief-MMWR-LM}
}

Model Card Authors

Bryan Tegomoh, MD, MPH

Contact

• Email: [email protected]
• GitHub: @BryanTegomoh
• Website: bryantegomoh.com

License

Apache 2.0

Acknowledgments

• CDC MMWR Program: For publishing comprehensive epidemiological reports
• Tinker AI: For distributed training infrastructure
• Meta AI: For the Llama 3.1 base model