app.py

import gradio as gr
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

# 初始化模型
emotion_analyzer = pipeline("text-classification", model="j-hartmann/emotion-english-distilroberta-base")
safety_checker = pipeline("text-classification", model="meta-llama/Meta-Llama-Guard-2-8B")
tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-medium")
model = AutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-medium")

# 预定义安全回复模板
SAFE_RESPONSES = {
    "crisis": "我注意到您可能需要专业帮助，建议立即联系心理咨询师或拨打心理援助热线。",
    "sadness": "听起来您最近压力很大，要不要试试深呼吸或听舒缓音乐？",
    "anger": "情绪波动很正常，我们可以一起分析问题的根源。"
}

def generate_response(user_input, history):
    # 安全检查
    safety_result = safety_checker(user_input)[0]
    if safety_result["label"] == "UNSAFE":
        return SAFE_RESPONSES["crisis"]
    
    # 情绪分析
    emotion = emotion_analyzer(user_input)[0]["label"]
    
    # 根据情绪选择生成策略
    if emotion in ["sadness", "fear"]:
        return SAFE_RESPONSES.get(emotion, "我理解您的感受，可以多聊聊吗？")
    
    # 生成对话
    inputs = tokenizer.encode(user_input + tokenizer.eos_token, return_tensors="pt")
    reply_ids = model.generate(
        inputs, 
        max_length=1000,
        pad_token_id=tokenizer.eos_token_id,
        no_repeat_ngram_size=3  # 避免重复
    )
    response = tokenizer.decode(reply_ids[:, inputs.shape[-1]:][0], skip_special_tokens=True)
    
    return response

# 创建Gradio界面
demo = gr.ChatInterface(
    fn=generate_response,
    examples=["最近总是失眠", "感觉没有人理解我", "考试成绩让我很焦虑"],
    title="青少年心理健康助手",
    description="请随时倾诉您的感受，我会尽力帮助您调整情绪。"
)

demo.launch()