app.py

from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr

# 加载模型
model_name = "SecurityXuanwuLab/HaS-820m"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)  # 默认加载到 CPU，避免 accelerate 依赖

# 定义基本功能函数
def paraphrase(text):
    prompt = f"将以下句子润色为更自然的表达：{text}"
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=50, temperature=0.7)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).replace(prompt, "").strip()

def summarize(text):
    prompt = f"将以下文本总结为简短的摘要：{text}"
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.7)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).replace(prompt, "").strip()

def translate(text, target_lang):
    prompt = f"将以下文本从中文翻译到{target_lang}：{text}"
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.7)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).replace(prompt, "").strip()

def reading_comprehension(context, question):
    prompt = f"根据以下上下文回答问题：\n上下文：{context}\n问题：{question}"
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=50, temperature=0.7)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).replace(prompt, "").strip()

def classify(text):
    prompt = f"判断以下文本的情感是积极还是消极：{text}。情感是"
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=10, temperature=0.7)
    return tokenizer.decode(outputs[0], skip_special_tokens=True).replace(prompt, "").strip()

# 定义 HaS 隐私保护功能
def hide(text):
    hide_template = """<s>Paraphrase the text:%s\n\n"""
    input_text = hide_template % text
    inputs = tokenizer(input_text, return_tensors="pt")
    outputs = model.generate(**inputs, max_length=100)
    pred = outputs[0][len(inputs['input_ids'][0]):]
    return tokenizer.decode(pred, skip_special_tokens=True)

def seek(hide_input, hide_output, original_input):
    seek_template = "Convert the text:\n%s\n\n%s\n\nConvert the text:\n%s\n\n"
    input_text = seek_template % (hide_input, hide_output, original_input)
    inputs = tokenizer(input_text, return_tensors="pt")
    outputs = model.generate(**inputs, max_length=512)
    pred = outputs[0][len(inputs['input_ids'][0]):]
    return tokenizer.decode(pred, skip_special_tokens=True)

# 处理任务
def process_task(task, text, context=None, question=None, target_lang=None, hide_output=None):
    if task == "润色":
        return paraphrase(text), None, None
    elif task == "摘要":
        return summarize(text), None, None
    elif task == "翻译":
        return translate(text, target_lang), None, None
    elif task == "阅读理解":
        return reading_comprehension(context, question), None, None
    elif task == "文本分类":
        return classify(text), None, None
    elif task == "隐私保护 (Hide)":
        hidden_text = hide(text)
        return hidden_text, hidden_text, None
    elif task == "隐私保护 (Seek)":
        if not hide_output:
            return "请提供 Hide 输出后的翻译结果", None, None
        original_output = seek(text, hide_output, context)  # text 为 hide_input, context 为 original_input
        return original_output, None, None

css = """ 
footer {
        visibility: hidden;
    }

"""

# Gradio 界面
with gr.Blocks(title="腾讯玄武产品体验 - HaS-820m", css=css) as demo:
    gr.Markdown("# 腾讯玄武产品体验 - HaS-820m")
    gr.Markdown("支持润色、摘要、翻译、阅读理解、文本分类及隐私保护（Hide 和 Seek）。")
    
    task = gr.Dropdown(
        choices=["润色", "摘要", "翻译", "阅读理解", "文本分类", "隐私保护 (Hide)", "隐私保护 (Seek)"],
        label="选择任务"
    )
    
    # 输入组件
    text_input = gr.Textbox(label="输入文本", lines=5)
    context_input = gr.Textbox(label="上下文（阅读理解或 Seek 的原始输入）", lines=5, visible=False)
    question_input = gr.Textbox(label="问题（阅读理解）", visible=False)
    lang_input = gr.Dropdown(choices=["英语", "法语", "西班牙语"], label="目标语言（翻译）", visible=False)
    hide_output_input = gr.Textbox(label="Hide 输出后的翻译结果（Seek）", lines=3, visible=False)
    
    # 输出组件
    output = gr.Textbox(label="输出结果", lines=5)
    hidden_text_output = gr.Textbox(label="Hide 输出（隐私保护）", lines=3, visible=False)
    note = gr.Markdown("**注意**：隐私保护 (Seek) 需要手动输入 Hide 输出后的翻译结果（当前无外部 API 支持）。")

    # 动态显示输入框
    def update_inputs(task):
        if task == "阅读理解":
            return (gr.update(visible=False), gr.update(visible=True, label="上下文"), 
                    gr.update(visible=True), gr.update(visible=False), gr.update(visible=False),
                    gr.update(visible=False))
        elif task == "翻译":
            return (gr.update(visible=True), gr.update(visible=False), 
                    gr.update(visible=False), gr.update(visible=True), gr.update(visible=False),
                    gr.update(visible=False))
        elif task == "隐私保护 (Hide)":
            return (gr.update(visible=True), gr.update(visible=False), 
                    gr.update(visible=False), gr.update(visible=False), gr.update(visible=False),
                    gr.update(visible=True))
        elif task == "隐私保护 (Seek)":
            return (gr.update(visible=True, label="Hide 输出"), gr.update(visible=True, label="原始输入"), 
                    gr.update(visible=False), gr.update(visible=False), gr.update(visible=True),
                    gr.update(visible=False))
        else:
            return (gr.update(visible=True), gr.update(visible=False), 
                    gr.update(visible=False), gr.update(visible=False), gr.update(visible=False),
                    gr.update(visible=False))
    
    task.change(
        fn=update_inputs,
        inputs=task,
        outputs=[text_input, context_input, question_input, lang_input, hide_output_input, hidden_text_output]
    )

    # 提交按钮
    submit_btn = gr.Button("提交")
    submit_btn.click(
        fn=process_task,
        inputs=[task, text_input, context_input, question_input, lang_input, hide_output_input],
        outputs=[output, hidden_text_output, note]
    )

demo.launch()