MKTY-3B-Chat

<p align="center"> <br> <img src="./image/mkty_cn_light_huggingface.svg" style="width:63%;"> </p> <br>

Minh Khoe Tue Y LLM (MKTY-3B-Chat)

🌍 Documentation Language

Chinese Simplified (简体中文) | English | Vietnamese (Tiếng Việt)

Please note that the English and Vietnamese versions of this document are translated from the Chinese version using LLM, with manual proofreading. However, discrepancies may still exist. In case of inconsistencies between the English or Vietnamese versions and the Chinese version, the Chinese version shall prevail.

Full Project Title: Minh Khoe Tue Y (Chinese Simplified: 明康慧医; Vietnamese: Minh Khỏe Tuệ Y; Nom Script: 明劸慧醫 ) — Design and Implementation of a Health Management and Diagnostic Assistance System Based on LLMs and Multimodal Artificial Intelligence. Abbreviation: MKTY Smart Healthcare System

📖 Model Overview

This model is a component of the "Minh Khoe Tue Y - Design and Implementation of a Health Management and Assisted Diagnosis System Based on LLM and Multimodal Artificial Intelligence" project (referred to as the Minh Khoe Tue Y Smart Healthcare System). It was developed as part of my undergraduate graduation project for the Faculty of Computer Science and Technology, Qilu University of Technology (Shandong Academy of Sciences), class of 2025. The project has been open-sourced and is available at: https://github.com/duyu09/MKTY-System.

This model has been fine-tuned and optimized in the fields of medicine, healthcare, and biology, outperforming its base model, Qwen2.5-3B-Instruct. The fine-tuning process employs the LoRA algorithm and is conducted in two stages, focusing solely on the Chinese language. Initially, during the Pretrain phase, the model undergoes incremental training using medical textbooks, medical records, and healthcare-related articles. Subsequently, Supervised Fine-Tuning (SFT) is performed using corpora that include symptoms and corresponding medical records, doctor-patient dialogues (symptom descriptions and diagnoses), medical knowledge Q&A, and dialogue corpora based on the "LLM Discussion Mechanism." The total data volume is approximately 2.88GB.

Notably, the model has been optimized for the "LLM Discussion Mechanism." The specific operation of this mechanism is as follows: when answering each question, the model generates multiple results based on different contexts, simulating a scenario where "multiple individuals express their viewpoints." The system also includes a "moderator" role responsible for summarizing the viewpoints from each round of discussion. Subsequently, all participants engage in the next round of discussion based on the original question, the moderator's summary, and their respective contexts. This process iterates until the discussion results converge (i.e., the semantics become consistent) or the preset maximum number of discussion rounds is reached.

🔧 Hardware Requirements

For GPU inference, a minimum of 7GB of VRAM is required. If the VRAM capacity is insufficient or if no dedicated GPU is available, the MKTY-3B large model can also run using CPU + 7GB RAM.

🚀 Usage Example

Based on the Tongyi Qianwen (Chinese: 通义千问) Qwen2.5-3B-Instruct model, it can be quickly loaded and launched using the transformers library.

Model Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
def load_model_and_tokenizer(model_name):
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype="auto",
        device_map="auto"
    )
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    return model, tokenizer
def generate_response(prompt, messages, model, tokenizer, max_new_tokens=2000):
    messages.append({"role": "user", "content": prompt})
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
    generated_ids = model.generate(
        **model_inputs,
        max_new_tokens=max_new_tokens
    )
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]
    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    messages.append({"role": "assistant", "content": response})
    return response

Standard Q&A Mode

if __name__ == "__main__":
    model_name = r"MKTY-3B-Chat"
    messages = []
    model, tokenizer = load_model_and_tokenizer(model_name)
    while True:
        prompt = input("User> ")
        if prompt == "exit":
            break
        response = generate_response(prompt, messages, model, tokenizer)
        print("MKTY>", response)

LLM Discussion Mode (Example language: Chinese Simplified)

if __name__ == "__main__":
    model_name = "MKTY-3B-Chat"
    discuss_rounds = 3
    agent_number = 3
    model, tokenizer = load_model_and_tokenizer(model_name)
    messages_arr = [[] for _ in range(agent_number)]
    while True:
        prompt = input("User> ")
        if prompt == "exit":
            break
        moderator_opinion = "暂无"
        for i in range(discuss_rounds):
            responses_arr = []
            prompt_per_round = "- 问题：\n" + prompt + "\n - 上轮讨论主持人意见：\n" + moderator_opinion + "\n - 请你结合主持人意见，对上述医疗或医学专业的问题发表详细观点，可以质疑并说明理由。\n"
            for j in range(agent_number):
                messages = messages_arr[j]
                response = generate_response(prompt_per_round, messages, model, tokenizer)
                responses_arr.append(response)
                print(f"第{i + 1}轮讨论，LLM {j + 1}观点>\n", response)
                print("-------------------")
            moderator_prompt = "- 问题：\n" + prompt + "\n\n"
            for res_index in range(len(responses_arr)):
                moderator_prompt = moderator_prompt + f"- LLM {res_index + 1}观点：\n" + responses_arr[res_index] + "\n\n"
            moderator_prompt = moderator_prompt + "对于给定的医疗相关问题，请综合各LLM观点，结合自身知识，得出你自己的判断，尽可能详尽，全部都分析到位，还要充分说明理由。\n"
            moderator_opinion = generate_response(moderator_prompt, [], model, tokenizer)
            print(f"第{i + 1}轮讨论，主持人的意见>\n", moderator_opinion)
            print("-------------------")
            clear_history(messages_arr)
    

🎓 Authors

██\      ██\     ██\   ██\   ████████\  ██\     ██\
███\    ███ |    ██ | ██  |  \__██  __| \██\   ██  |
████\  ████ |    ██ |██  /      ██ |     \██\ ██  /
██\██\██ ██ |    █████  /       ██ |      \████  /
██ \███  ██ |    ██  ██<        ██ |       \██  /
██ |\█  /██ |    ██ |\██\       ██ |        ██ |
██ | \_/ ██ |██\ ██ | \██\ ██\  ██ |██\     ██ |██\
\__|     \__|\__|\__|  \__|\__| \__|\__|    \__|\__|

This model is used for the graduation project of the Faculty of Computer Science and Technology, Qilu University of Technology (Shandong Academy of Sciences) in 2025, and is only for academic exchange. Neither I nor my supervisor teachers are responsible for any consequences arising from the use of the model.

• 🧑‍💻 Project Author:

  • DU Yu (Chinese: 杜宇; Vietnamese: Đỗ Vũ; Email: [email protected]), undergraduate student at Faculty of Computer Science and Technology, Qilu University of Technology (Shandong Academy of Sciences)

• 🏫 Thesis Advisors:

  • Academic Advisor: JIANG Wenfeng (Chinese: 姜文峰; Vietnamese: Khương Văn Phong), Associate professor, Faculty of Computer Science and Technology, Qilu University of Technology (Shandong Academy of Sciences)
  • Industry Advisor: LI Jun (Chinese: 李君; Vietnamese: Lý Quân), Shandong Strong (Shichuang) Software Training College, Ambow Education Group (NYSE: AMBO)

The complete project's open source address: https://github.com/duyu09/MKTY-System. Welcome to download and discuss about it.

🔗 Links

• Qilu University of Technology (Shandong Academy of Sciences): https://www.qlu.edu.cn/

• Shandong Computer Center (National Supercomputing Center in Jinan, NSCCJN): https://www.nsccjn.cn/

• Faculty of Computer Science and Technology, Qilu University of Technology (Shandong Academy of Sciences): http://jsxb.scsc.cn/

• DuYu's GitHub Account: https://github.com/duyu09/

📄 Citation

@software{du_2025_17444889,
  author       = {Du, Yu},
  title        = {Minh Khoe Tue Y Smart Healthcare System},
  month        = oct,
  year         = 2025,
  publisher    = {Zenodo},
  version      = {v1.1.2},
  doi          = {10.5281/zenodo.17444889},
  url          = {https://github.com/duyu09/MKTY-System},
  swhid        = {swh:1:dir:a633243bf04e6ba18e2d5ffcf92ea57f73566f43
                   ;origin=https://doi.org/10.5281/zenodo.17444888;vi
                   sit=swh:1:snp:37dc91d2c166a07c7dc8ebac0b4be97961b0
                   267b;anchor=swh:1:rel:a88f82a5ca10d278bcc10734f5cf
                   a560286a8b47;path=duyu09-MKTY-System-8edd0c9
                  },
}