EviNurse-32B

EviNurse-32B

EviNurse-32B is a domain-specific large language model developed for evidence-based nursing. It is based on Qwen3-32B and adapted to nursing-domain use through supervised fine-tuning. In the EviNurse study, the model was evaluated with automated multiple-choice testing, expert assessment of short-answer responses, and real-world usability evaluation among nurses.

The goal of this release is to support reproducible research on nursing AI by providing a model artifact that can be evaluated together with the public NursData-MCQ benchmark.

Resources

Model Details

• Base model: Qwen3-32B
• Architecture: Qwen3ForCausalLM
• Parameters: approximately 32.8B
• Precision: bfloat16
• Context length in config: 40,960 tokens
• Training method: supervised fine-tuning for nursing-domain question answering and evidence-based nursing use cases
• Intended interface: chat/text generation, preferably through an OpenAI-compatible serving engine such as vLLM

Intended Use

This model is intended for research on:

• Nursing-domain language models
• Evidence-based nursing question answering
• Retrieval-augmented generation for nursing evidence
• Automated nursing MCQ benchmark evaluation
• Comparative evaluation of general-purpose and domain-specific LLMs

This model is not a medical device and should not be used as a substitute for professional clinical judgment.

Evaluation

The automated evaluation benchmark used with this model is available at:

• https://huggingface.co/datasets/Agnania/NursData-MCQ

The benchmark contains 3,438 Chinese nursing multiple-choice questions. Accuracy is computed by exact match between the extracted final option letter and the standard answer.

Example evaluation code is provided in:

• https://github.com/Creeper12345/EviNurse

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Agnania/EviNurse-32B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are an evidence-based nursing assistant."},
    {"role": "user", "content": "请简要说明预防压疮的核心护理措施。"},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Serving with vLLM

python -m vllm.entrypoints.openai.api_server \
  --model Agnania/EviNurse-32B \
  --served-model-name EviNurse \
  --tensor-parallel-size 4 \
  --trust-remote-code

Then evaluate through the OpenAI-compatible endpoint:

python scripts/evaluate_mcq.py \
  --input data/NursData-MCQ/evinurse_automated_eval_3438.json \
  --output outputs/evinurse_mcq_predictions.json \
  --base-url http://127.0.0.1:8000/v1 \
  --model EviNurse \
  --api-key EMPTY

Release Notes

Before uploading this directory, ensure that all weight shards referenced in model.safetensors.index.json are present:

model-00001-of-00004.safetensors
model-00002-of-00004.safetensors
model-00003-of-00004.safetensors
model-00004-of-00004.safetensors

The non-weight configuration files in this release have been checked for local training and server path leaks.

License

EviNurse-32B is released under the Apache License 2.0.

Base model:

• Qwen3-32B

Please also comply with the license of the original base model and any applicable data/source licenses.

This model is intended for research and educational use in nursing and healthcare AI. It is not a medical device and should not be used as the sole basis for clinical decision-making.