ShizhenGPT-7B-VL

GitHub | Paper

ShizhenGPT is the first multimodal LLM for Traditional Chinese Medicine (TCM). It not only possesses strong expertise in TCM, but also supports TCM multimodal diagnostic capabilities, which involve looking (望), listening/smelling (闻), questioning (问), and pulse-taking (切).

👉 More details on GitHub: ShizhenGPT

Model Info

ShizhenGPT-7B-VL is a variant derived from ShizhenGPT-7B-Omni that includes only the LLM and vision encoder. It is recommended if your use case involves text or vision tasks exclusively. For broader multimodal needs, please select one of the versions below.

	Parameters	Supported Modalities	Link
ShizhenGPT-7B-LLM	7B	Text	HF Link
ShizhenGPT-7B-VL	7B	Text, Image Understanding	HF Link
ShizhenGPT-7B-Omni	7B	Text, Four Diagnostics (望闻问切)	HF Link
ShizhenGPT-32B-LLM	32B	Text	HF Link
ShizhenGPT-32B-VL	32B	Text, Image Understanding	HF Link
ShizhenGPT-32B-Omni	32B	Text, Four Diagnostics (望闻问切)	Available soon

Note: The LLM and VL models are parameter-split variants of ShizhenGPT-7B-Omni. Since their architectures align with Qwen2.5 and Qwen2.5-VL, they are easier to adapt to different environments. In contrast, ShizhenGPT-7B-Omni requires transformers==4.51.0.

Usage

You can use ShizhenGPT-7B-VL in the same way as Qwen2.5-VL-7B-Instruct. You can deploy it with tools like vllm or Sglang, or perform direct inference:

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info


processor = AutoProcessor.from_pretrained("FreedomIntelligence/ShizhenGPT-7B-VL")
model = Qwen2_5_VLForConditionalGeneration.from_pretrained("FreedomIntelligence/ShizhenGPT-7B-VL", torch_dtype="auto", device_map="auto")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "/path/to/your/image.png",
            },
            {"type": "text", "text": "请从中医角度解读这张舌苔。"},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

📖 Citation

@misc{chen2025shizhengptmultimodalllmstraditional,
      title={ShizhenGPT: Towards Multimodal LLMs for Traditional Chinese Medicine}, 
      author={Junying Chen and Zhenyang Cai and Zhiheng Liu and Yunjin Yang and Rongsheng Wang and Qingying Xiao and Xiangyi Feng and Zhan Su and Jing Guo and Xiang Wan and Guangjun Yu and Haizhou Li and Benyou Wang},
      year={2025},
      eprint={2508.14706},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.14706},
}