sczzz/RaDaR-32B

Overview

RaDaR (Rare Disease navigatoR) is a 32B-parameter reasoning large language model specialized for rare-disease differential diagnosis from free-text clinical narratives.

RaDaR was initialized from DeepSeek-R1-Distill-Qwen-32B and further trained using real-world rare-disease case reports and phenotype-anchored synthetic clinical cases, followed by supervised fine-tuning (SFT) and direct preference optimization (DPO).

RaDaR is designed to process free-text clinical records directly, without requiring users to first convert the clinical narrative into Human Phenotype Ontology (HPO) terms.

One motivation for releasing RaDaR as an open-weight model is to support local deployment, including settings in which sensitive clinical data should remain within an institution's own computing environment.

Resources

• Training and synthetic-data construction code: https://github.com/sczzz3/RaDaR
• Paper: https://arxiv.org/abs/2606.24510
• Web application: https://raredx.datummed.com

Clinical-use notice

RaDaR is a research and clinical decision-support model. It is not intended to replace qualified healthcare professionals or to provide autonomous diagnosis or treatment decisions. Model outputs may be incorrect, incomplete, or overconfident and should always be independently reviewed by qualified clinicians.

Quick start

This section provides step-by-step instructions for downloading RaDaR and running it locally.

1. Hardware and storage

The released RaDaR-32B checkpoint is stored in BF16 format and occupies approximately 65.5 GB of disk space. We recommend reserving at least 70 GB of free disk space for the model files, plus additional space for the Python environment and model cache. Because the BF16 model weights alone occupy approximately 65.5 GB, GPU inference requires additional memory beyond the raw model size for runtime overhead and the key-value cache. Depending on the available hardware, users can run RaDaR on:

• a single high-memory GPU;
• multiple GPUs using automatic model sharding; or
• multiple GPUs using tensor parallelism with an inference engine such as vLLM.

Actual memory requirements depend on the input length, output length, inference backend, and number of concurrent requests. For users with limited GPU memory, shorter context lengths can substantially reduce runtime memory requirements.

2. Create a Python environment

We recommend Python 3.11.

conda create -n radar python=3.11 -y
conda activate radar

Install the required packages:

pip install torch
pip install transformers accelerate safetensors huggingface_hub

The released checkpoint was saved with Transformers 4.45.2. If you encounter version-related compatibility issues, you can use:

pip install transformers==4.45.2

3. Download RaDaR

Install or update the Hugging Face Hub command-line utility:

pip install -U huggingface_hub

Download the complete model to a local directory:

hf download sczzz/RaDaR-32B \
    --local-dir ./RaDaR-32B

After the download completes, the local directory should contain the model weights, tokenizer files, configuration files, and generation configuration. You can also skip this manual download step and use the Hugging Face repository name directly in from_pretrained(). However, explicitly downloading the checkpoint is recommended when preparing a fully offline clinical environment.

4. Local inference with Transformers

Create a file called run_radar.py:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Path to the locally downloaded checkpoint.
MODEL_PATH = "./RaDaR-32B"

tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

model.eval()

# ---------------------------------------------------------------------
# Replace the example text below with your own de-identified or locally
# governed clinical narrative.
# ---------------------------------------------------------------------

case_text = """
A 38-year-old male visited the surgery clinic with a year-long history of upper abdominal pain. The pain was described as dull and aching, with episodes of increased intensity lasting 1 to 2 hours...
"""

prompt = f"""
As a medical expert, enumerate the top 10 most likely diagnoses for the following patient in descending order of likelihood, with the most likely disease listed first. Ensure that each diagnosis is as specific as possible, avoiding vague terms like "rare genetic disease".

Here is the case: 
{case_text}

Only output the diagnosis in numeric order, one per line. For example:
1. Disease A;
2. Disease B;
...

Do not output anything else!
"""

# For the DeepSeek-R1-Distill model family, instructions are placed
# directly in the user message rather than in a separate system prompt.
messages = [
    {"role": "user", "content": prompt}
]

formatted_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(
    formatted_prompt,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=32768,
        do_sample=True,
        temperature=0.6,
        top_p=0.95,
        pad_token_id=tokenizer.eos_token_id,
    )

generated_tokens = outputs[0][inputs["input_ids"].shape[-1]:]

response = tokenizer.decode(
    generated_tokens,
    skip_special_tokens=True,
)

print(response)

Run the script:

python run_radar.py

The first model load may take several minutes depending on disk and GPU speed.

Optional: local API deployment with vLLM

RaDaR can also be served as a local API for integration with an institutional interface or clinical research application.

Install vLLM:

pip install vllm

For example, to serve RaDaR using two GPUs:

vllm serve ./RaDaR-32B \
    --dtype bfloat16 \
    --tensor-parallel-size 2 \
    --max-model-len 32768 \
    --served-model-name RaDaR-32B

Adjust --tensor-parallel-size to match the number of GPUs available. The maximum model length can also be reduced if GPU memory is limited. The server listens locally on port 8000 by default.

An example request is:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "RaDaR-32B",
    "messages": [
      {
        "role": "user",
        "content": "Based on the following clinical information, provide a ranked differential diagnosis of up to five rare diseases.\n\nClinical information:\n[CLINICAL RECORD]"
      }
    ],
    "temperature": 0.6,
    "top_p": 0.95,
    "max_tokens": 32768
  }'

For a fully offline deployment, start the server from the local checkpoint after enabling offline mode:

export HF_HUB_OFFLINE=1

vllm serve ./RaDaR-32B \
    --dtype bfloat16 \
    --tensor-parallel-size 2 \
    --max-model-len 32768 \
    --served-model-name RaDaR-32B

This allows a local application to interact with RaDaR without sending clinical records to an external LLM API.

Input format

RaDaR is primarily designed for free-text clinical narratives.

Useful information may include:

• age, sex, and age of symptom onset;
• major symptoms and physical examination findings;
• relevant past medical and family history;
• laboratory results;
• imaging findings;
• pathology or procedural findings;
• previous diagnostic investigations;
• treatment history and treatment response;
• genetic or genomic findings, when available.

Output

RaDaR generates diagnostic reasoning and candidate rare-disease diagnoses. For clinical decision support, the output should be interpreted as a differential diagnosis list, not as a definitive diagnosis. The primary diagnostic evaluation in our study used a top-5 differential-diagnosis setting, reflecting the intended use of RaDaR as a tool for prioritizing a short list of diseases for clinician review.

Model development

RaDaR was initialized from:

DeepSeek-R1-Distill-Qwen-32B

The training corpus included:

• 49,170 publicly available real-world rare-disease cases; and
• 104,666 phenotype-anchored synthetic rare-disease cases.

Training consisted of two reasoning-enhancement stages:

1. Supervised Fine-Tuning (SFT) using rare-disease clinical narratives paired with diagnostic reasoning trajectories; and
2. Direct Preference Optimization (DPO) using preferred and dispreferred diagnostic reasoning outputs.

The full methodology is described in the accompanying paper. Code for phenotype sampling, synthetic-case construction, SFT, and DPO is available at: https://github.com/sczzz3/RaDaR

Citation

If you use RaDaR in your research, please cite:

@article{chen2026radar,
  title   = {A specialized reasoning large language model for accelerating rare disease diagnosis: a randomized AI physician-assistance trial},
  author  = {Chen, Haichao and
             Zhou, Songchi and
             Zhao, Zhengyun and
             Hu, Shikai and
             Jin, Xianghong and
             Ji, Hongwei and
             He, Li and
             Li, Shuli and
             Qin, Yiming and
             Tan, Xin and
             Shi, Runfeng and
             Tham, Yih Chung and
             Zhu, Jiaye and
             Li, Ye and
             Jin, Ye and
             Cao, Longhao and
             Li, Dawei and
             Wu, Honghan and
             Gu, Hongqiu and
             Li, Guanqiao and
             Groza, Tudor and
             Li, Chunying and
             Zeng, Dian and
             Yu, Weihong and
             Baynam, Gareth and
             Jamuar, Saumya Shekhar and
             Shen, Min and
             Zhang, Shuyang and
             Sheng, Bin and
             Yu, Sheng and
             Wong, Tien Yin},
  journal = {arXiv preprint arXiv:2606.24510},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.24510}
}