app.py

import gradio as gr
from huggingface_hub import InferenceClient
import evaluate

# 创建困惑度计算工具
perplexity = evaluate.load("perplexity", module_type="metric")

# 创建推理客户端
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")

def compute_perplexity(message):
    # 制备消息列表，这里只有用户消息
    messages = [{"role": "user", "content": message}]

    # 通过客户端完成聊天生成任务
    response = client.chat_completion(
        messages,
        max_tokens=512,
        stream=False,
        temperature=0.7,
        top_p=0.95
    )
    
    # 获取生成的文本内容
    generated_text = response.choices[0].delta.content

    # 计算困惑度
    perplexity_results = perplexity.compute(model_id='gpt2', add_start_token=False, predictions=[generated_text])
    perplexity_value = perplexity_results['perplexity']

    # 返回困惑度结果
    return f"Perplexity of the response: {perplexity_value}"

# 设置 Gradio 界面
demo = gr.Interface(
    fn=compute_perplexity,
    inputs="text",
    outputs="text",
    title="Compute Perplexity",
    description="Enter a text to compute its perplexity based on the gpt2 model."
)

if __name__ == "__main__":
    demo.launch()