app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# 加载模型和 tokenizer
model_name = "LilithHu/mbert-manipulative-detector"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# 设置为评估模式
model.eval()

# 设置运行设备
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# 标签名
labels = ["Non-manipulative / 非操纵性", "Manipulative / 操纵性"]

# 推理函数
def classify(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(device)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=1)[0]
        
        threshold = 0.7  # 自定义阈值（你可以改成别的）
        if probs[1].item() > threshold:
            pred = 1  # 判为操纵性
        else:
            pred = 0  # 判为非操纵性

        confidence = min(probs[pred].item(), 0.95)  # 置信度依然可以控制上限

        percent = round(confidence * 100, 2)
        result = f"Prediction / 预测:\n{labels[pred]}\n"
        return result

#谁大选谁        
#pred = torch.argmax(probs).item()
#confidence = min(probs[pred].item(), 0.95)  # 限制置信度最大为95%

# Gradio 界面
interface = gr.Interface(
    fn=classify,
    inputs=gr.Textbox(
        lines=4, 
        placeholder="Enter text in English or Chinese... ",
        label="📝 Input Text"
    ),
    outputs=gr.Markdown(label="📊 Prediction"),
    title="🔍 Manipulative Language Detector",
    description="""
🧪 Enter a sentence in English or Chinese to detect if it's manipulative.  

📌 **Disclaimer**  
This system is for **research and educational purposes only**.  
It **does not guarantee accuracy** and **should not be used as legal or clinical evidence**.  

🤖 **Model Info**
- Model: `LilithHu/mbert-manipulative-detector`  
- Base: `mDeBERTa-v3` multilingual pre-trained model  
- Fine-tuned using HuggingFace Transformers on 10,000 labeled Chinese data  

⚠️ **About Examples**
The examples provided below are those **cited in the paper**, including implicit moral coercion, polite masking and false positives.


🌐 **Built with Gradio and hosted on HuggingFace Spaces**



""",
    examples=[
        ["A good kid would never act like this"],
        ["hi, your mother is tired, can you just do whatever she says"],
        ["I feel so sad when you don’t listen to me."],
        ["If you are willing to help, it doesn’t matter if you are not... let it go, i know my position"],
        ["If you take that job, don’t call me mom again."],
        ["I’ve sacrificed my whole life for you, how could you treat me this way "]
    ],
    theme="default",
    allow_flagging="never"
)

interface.launch()