ZhangQiao123/medgemma-chinese-medical-qa

📚 华佗百科问答数据集 (1,486条):
├── 来源: 复旦大学自由智能实验室 (FreedomIntelligence)
├── 特点: 高质量中文医疗百科问答
├── 覆盖: 疾病诊断、症状分析、治疗方案
├── 格式: 结构化问答对 (Question-Answer)
├── 权重: 53.3% (基础医学知识)
└── 语言: 纯中文

🧠 医学O1推理数据集 (800条):
├── 来源: FreedomIntelligence/medical-o1-reasoning-SFT
├── 特点: 复杂医疗推理思维链训练
├── 覆盖: 诊断推理、病例分析、治疗决策
├── 格式: 包含推理过程的多轮对话
├── 配置: 中文配置 (zh) + 混合配置 (zh_mix)
├── 权重: 28.7% (推理能力训练)
└── 语言: 中文为主

🎯 MedAlpaca医学闪卡 (500条):
├── 来源: MedAlpaca项目医学闪卡数据集
├── 特点: 医学知识点快速问答
├── 覆盖: 基础医学、临床医学概念
├── 格式: 简洁问答对
├── 处理: 英文原始数据，添加中文医疗提示
├── 权重: 18.0% (知识补充)
└── 语言: 英文转中文处理

专科分布:
├── 内科学 (Internal Medicine) - 35%
├── 外科学 (Surgery) - 20%
├── 儿科学 (Pediatrics) - 15%
├── 妇产科学 (Obstetrics & Gynecology) - 10%
├── 心血管科 (Cardiology) - 8%
├── 神经科学 (Neurology) - 7%
└── 急诊医学 (Emergency Medicine) - 5%

# 实际使用的数据集加载代码
datasets = {
    # 华佗百科问答 - 基础医学知识 (50%)
    'huatuo_qa': load_dataset("FreedomIntelligence/huatuo_encyclopedia_qa", split="train"),

    # 医学推理数据集 - 推理能力训练 (30%)
    'medical_reasoning': load_dataset("FreedomIntelligence/medical-o1-reasoning-SFT", "zh", split="train"),

    # 医学推理混合 - 中英混合数据 (15%)
    'medical_reasoning_mix': load_dataset("FreedomIntelligence/medical-o1-reasoning-SFT", "zh_mix", split="train"),

    # 医学闪卡 - 补充知识 (5%)
    'medical_flashcards': load_dataset("medalpaca/medical_meadow_medical_flashcards", split="train")
}

pip install transformers torch accelerate

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# 加载模型和分词器 (验证可用 ✅)
model_name = "ZhangQiao123/medgemma-chinese-medical-qa"
print("🔄 加载模型...")

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    low_cpu_mem_usage=True
)

print("✅ 模型加载成功！")

# 医疗问答函数
def medical_chat(question):
    """生成医疗问答回复"""
    prompt = f"<start_of_turn>user\n{question}<start_of_turn>model\n"
    inputs = tokenizer(prompt, return_tensors="pt")

    # 移动到GPU (如果可用)
    if torch.cuda.is_available():
        inputs = {k: v.cuda() for k, v in inputs.items()}

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.7,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id
        )

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response.split("<start_of_turn>model\n")[-1]

# 测试示例 (验证有效)
test_questions = [
    "高血压的主要症状有哪些？",
    "糖尿病患者应该注意什么？",
    "感冒和流感有什么区别？"
]

print("\n🧠 开始医疗问答测试:")
for i, question in enumerate(test_questions, 1):
    print(f"\n📋 问题 {i}: {question}")
    answer = medical_chat(question)
    print(f"🏥 回答: {answer}")
    print("-" * 60)

# 注意: 需要HuggingFace Token认证
import requests

def call_inference_api(question, hf_token):
    """调用HuggingFace推理API"""
    url = "https://api-inference.huggingface.co/models/ZhangQiao123/medgemma-chinese-medical-qa"

    headers = {
        "Authorization": f"Bearer {hf_token}",
        "Content-Type": "application/json"
    }

    payload = {
        "inputs": f"<start_of_turn>user\n{question}<start_of_turn>model\n",
        "parameters": {
            "max_new_tokens": 512,
            "temperature": 0.7,
            "do_sample": True
        }
    }

    response = requests.post(url, headers=headers, json=payload)

    if response.status_code == 200:
        return response.json()
    else:
        print(f"API调用失败: {response.status_code}")
        return None

# 使用示例 (需要您的HF Token)
# result = call_inference_api("高血压有什么症状？", "your_hf_token_here")