DetectAnyLLM / app.py
JiachenFu's picture
update: fix some bug
40b691d
import gradio as gr
import numpy as np
import torch
import os
import json
from core.model import DiscrepancyEstimator
import re
import docx
import spaces
from datasets import load_dataset
def read_file_content(file):
if file is None:
return ""
if file.name.endswith('.txt'):
with open(file.name, 'r', encoding='utf-8') as f:
return f.read()
elif file.name.endswith('.docx'):
doc = docx.Document(file.name)
full_text = []
for para in doc.paragraphs:
full_text.append(para.text)
return '\n'.join(full_text)
return ""
def split_sentences(text):
"""根据句号、句点、分号分割文本成句子,同时保留分句符号。"""
sentences = re.split(r'([。.])', text)
combined_sentences = [sentences[i] + sentences[i+1] for i in range(0, len(sentences)-1, 2)]
if len(sentences) % 2 == 1:
combined_sentences.append(sentences[-1])
return [s.strip() for s in combined_sentences if s.strip()]
def count_words(sentence, language='Chinese'):
"""统计句子的词数。"""
return len(sentence.replace('\n', '').replace('\r', '').split()) if language != 'Chinese' else len(sentence.replace('\n', '').replace('\r', ''))
def segment_text(sentences, language='Chinese'):
"""按照要求拼接句子,确保不忽略第一段并处理最后一句话不足100词的情况。"""
result = []
current_segment = []
current_length = 0
for i, sentence in enumerate(sentences):
word_count = count_words(sentence, language)
if word_count > 100:
# 如果单个句子超过100词,考虑拼接
if i + 1 < len(sentences) and word_count + count_words(sentences[i + 1], language) <= 200:
# 拼接当前和下一个句子
if current_segment: # 先保存当前段
result.append(' '.join(current_segment) if language != 'Chinese' else ''.join(current_segment))
result.append((sentence + ' ' + sentences[i + 1]) if language != 'Chinese' else (sentence + sentences[i + 1]))
current_segment = []
current_length = 0
i += 1 # 跳过下一个句子
continue
else:
# 单独存放
if current_segment: # 先保存当前段
result.append(' '.join(current_segment) if language != 'Chinese' else ''.join(current_segment))
result.append(sentence)
current_segment = []
current_length = 0
else:
if current_length + word_count > 100:
# 当前段超过100词,保存并开始新段
if current_segment:
result.append(' '.join(current_segment) if language != 'Chinese' else ''.join(current_segment))
current_segment = [sentence]
current_length = word_count
else:
# 继续累积
current_segment.append(sentence)
current_length += word_count
# 处理最后一段
if current_segment:
if current_length < 100 and result and current_length + count_words(result[-1], language) <= 200:
# 如果最后一段不足100词,且可以与前一段合并
last_segment = result.pop() if result else ''
current_segment = (last_segment.split() if language != 'Chinese' else list(last_segment)) + current_segment
result.append(' '.join(current_segment) if language != 'Chinese' else ''.join(current_segment))
else:
# 直接添加最后一段
result.append(' '.join(current_segment) if language != 'Chinese' else ''.join(current_segment))
return result
def extract_latex_text(latex_source):
# 提取document环境中的内容
doc_pattern = re.compile(r'\\begin{document}(.*?)\\end{document}', re.DOTALL)
match = doc_pattern.search(latex_source)
content = match.group(1) if match else latex_source
# 删除注释(排除转义后的%)
content = re.sub(r'(?<!\\)%.*', '', content, flags=re.MULTILINE)
# 排除常见非文本环境
excluded_envs = ['figure', 'table', 'equation', 'align\*?', 'verbatim', 'lstlisting']
env_pattern = re.compile(
r'\\begin{(' + '|'.join(excluded_envs) + r')}.*?\\end{\1}',
re.DOTALL
)
content = env_pattern.sub('', content)
# 新增处理:删除所有cite命令及其内容
content = re.sub(r'\\cite(\[[^\]]*\])?\{[^}]*\}', '', content)
# 新增处理:删除行内table/figure命令及其内容
content = re.sub(r'\\(table|figure)\*?(\[[^\]]*\])?\{[^}]*\}', '', content)
# 删除简单命令(无参数)
content = re.sub(r'\\([a-zA-Z]+)\*?\b', '', content)
# 递归处理带参数的命令(最多迭代10次防止死循环)
for _ in range(10):
new_content = re.sub(
r'\\([a-zA-Z]+)\*?(?:\[.*?\])*{((?:[^{}]*|{[^{}]*})*)}',
lambda m: m.group(2),
content,
flags=re.DOTALL
)
if new_content == content:
break
content = new_content
# 处理特殊字符
replacements = {
'~': ' ', '\\&': '&', '\\$': '$', '\\%': '%',
'\\_': '_', '\\#': '#', '\\\\': '\n', '\n': ' ',
'“': '"', '”': '"', '‘': "'", '’': "'"
}
for k, v in replacements.items():
content = content.replace(k, v)
# 清理空白字符
content = re.sub(r'[ \t]+', ' ', content)
content = re.sub(r'\n{2,}', '\n\n', content)
return content.strip()
class ProbEstimator:
def __init__(self, ref_file_dir):
self.tasks = ["polish", "generate", "rewrite"]
self.real_crits = {"polish": [], "generate": [], "rewrite": []}
self.fake_crits = {"polish": [], "generate": [], "rewrite": []}
for task in self.tasks:
task_ref_data = load_dataset(ref_file_dir, data_files=f'{task}.json')['train']
self.real_crits[task].extend(task_ref_data['original_discrepancy'])
self.fake_crits[task].extend(task_ref_data['rewritten_discrepancy'])
print(f'ProbEstimator: total {sum([len(self.real_crits[task]) for task in self.tasks]) * 2} samples.')
def crit_to_prob(self, crit):
probs = {}
for task in self.tasks:
real_crits = self.real_crits[task]
fake_crits = self.fake_crits[task]
total_len = len(real_crits) + len(fake_crits)
offset = np.sort(np.abs(np.array(real_crits + fake_crits) - crit))[int(0.1*total_len)]
cnt_real = np.sum((np.array(real_crits) > crit - offset) & (np.array(real_crits) < crit + offset))
cnt_fake = np.sum((np.array(fake_crits) > crit - offset) & (np.array(fake_crits) < crit + offset))
probs[task] = (cnt_fake / (cnt_real + cnt_fake)) if (cnt_real + cnt_fake) > 0 else 0.5
return probs
device = 'cuda'
zh_prob_estimator = ProbEstimator(ref_file_dir="JiachenFu/Qwen2-0.5B-detectanyllm-detector-ref-zh")
en_prob_estimator = ProbEstimator(ref_file_dir="JiachenFu/Qwen2-0.5B-detectanyllm-detector-ref-en")
@spaces.GPU
def greet(mode, language, input_text):
if mode == "LaTex":
input_text = extract_latex_text(input_text)
split_texts = split_sentences(input_text)
sub_texts = segment_text(split_texts, language=language)
detected = []
if language == "Chinese":
model = DiscrepancyEstimator(pretrained_ckpt="JiachenFu/Qwen2-0.5B-detectanyllm-detector-zh").to(device)
prob_estimator = zh_prob_estimator
else:
model = DiscrepancyEstimator(pretrained_ckpt="JiachenFu/Qwen2-0.5B-detectanyllm-detector-en").to(device)
prob_estimator = en_prob_estimator
model.eval()
for i, sub_text in enumerate(sub_texts):
text_content = sub_text
print(f'processing {sub_text}')
tokens = model.scoring_tokenizer(
text_content, return_tensors='pt', padding=True, truncation=True, return_token_type_ids=False
)
print(f'tokenized')
input_ids = tokens['input_ids'].to(device)
attention_mask = tokens['attention_mask'].to(device)
with torch.no_grad():
output = model.get_discrepancy_of_scoring_and_reference_models(
input_ids_for_scoring_model=input_ids,
attention_mask_for_scoring_model=attention_mask,
input_ids_for_reference_model=None,
attention_mask_for_reference_model=None,
)
discrepancy = output['scoring_discrepancy']
discrepancy = discrepancy.cpu().numpy().item()
print(f'discrepancy: {discrepancy}')
probs = prob_estimator.crit_to_prob(discrepancy)
if discrepancy < 15:
for task in probs.keys():
probs[task] = 0.0
detected.append({
'order': i,
'text': text_content,
'words_count': len(text_content) if language == "Chinese" else len(text_content.split()),
'probs': probs
})
# 添加绝对定位的总概率显示
# 构建动画效果
html_output = '''
<style>
@keyframes reveal {
from { opacity: 0; }
to { opacity: 1; }
}
.reveal-char {
opacity: 0;
animation: reveal 0.2s forwards;
white-space: pre-wrap;
}
</style>
<div style="position: relative; padding-bottom: 60px; min-height: 120px;">
'''
current_delay = 0.0 # 当前动画延迟时间
char_duration = 0.001 # 每个字符的间隔时间
# 处理文本内容
for item in detected:
ai_generate_prob = item['probs']['generate']
ai_revise_prob = max(item['probs']['polish'], item['probs']['rewrite'])
prob = max(ai_generate_prob, ai_revise_prob)
if prob >= 0.75:
if ai_generate_prob >= ai_revise_prob:
color = "red"
item["generate"] = 1
item["revise"] = 0
else:
color = "orange"
item["generate"] = 0
item["revise"] = 1
else:
color = "black"
item["generate"] = 0
item["revise"] = 0
for char in item['text']:
html_output += f'<span class="reveal-char" style="color: {color}; animation-delay: {current_delay:.2f}s;">{char}</span>'
current_delay += char_duration
total_length = sum(item['words_count'] for item in detected)
# total_prob = sum(item['prob'] * item['words_count'] for item in detected) / total_length if total_length > 0 else 0
generate_prob = sum(item["generate"] * item["words_count"] for item in detected) / total_length if total_length > 0 else 0
revise_prob = sum(item["revise"] * item["words_count"] for item in detected) / total_length if total_length > 0 else 0
html_output += f'''
<div style="
position: absolute;
bottom: 0;
right: 0;
background-color: rgba(255, 255, 255, 0.9);
padding: 8px 12px;
border-radius: 4px;
box-shadow: 0 2px 4px rgba(0,0,0,0.1);
border: 1px solid #e0e0e0;
font-size: 14px;
">
🤖 AI Generated Rate: <strong>{generate_prob:.2%}</strong><br>
✍️ AI Revised Rate: <strong>{revise_prob:.2%}</strong>
</div>
'''
html_output += '</div>'
return html_output
# 使用Blocks替代Interface以获得更好的自定义能力
# 修改CSS部分
with gr.Blocks(css="""
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap');
:root {
--accent-color: #6366f1;
--text-color: #374151;
--border-color: #e5e7eb;
--background-light: #f9fafb;
--background-card: #ffffff;
}
body, .gradio-container {
background: var(--background-light);
font-family: 'Inter', sans-serif;
color: var(--text-color);
}
#header {
text-align: center;
padding: 2rem;
margin: 0 auto; /* Use gap for spacing, remove margin-bottom */
background-color: var(--background-card);
background-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='40' height='40' viewBox='0 0 40 40'%3E%3Cg fill-rule='evenodd'%3E%3Cg fill='%23e5e7eb' fill-opacity='0.3'%3E%3Cpath d='M0 38.59l2.83-2.83 1.41 1.41L1.41 40H0v-1.41zM0 1.4l2.83 2.83 1.41-1.41L1.41 0H0v1.41zM38.59 40l-2.83-2.83 1.41-1.41L40 38.59V40h-1.41zM40 1.41l-2.83 2.83-1.41-1.41L38.59 0H40v1.41zM20 18.6l2.83-2.83 1.41 1.41L21.41 20l2.83 2.83-1.41 1.41L20 21.41l-2.83 2.83-1.41-1.41L18.59 20l-2.83-2.83 1.41-1.41L20 18.59z'/%3E%3C/g%3E%3C/g%3E%3C/svg%3E");
border: 1px solid var(--border-color);
border-radius: 16px;
box-shadow: 0 4px 12px rgba(0,0,0,0.05);
}
#title {
font-weight: 800;
font-size: 2.5em;
letter-spacing: -0.02em;
color: var(--text-color);
margin-bottom: 0.25em;
}
.detect-grad {
background: -webkit-linear-gradient(left, #ff8c8c, #ffc89e);
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
font-weight: 800;
}
.anyllm-grad {
background: -webkit-linear-gradient(left, #a0e6ff, #aaffd4);
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
font-weight: 800;
}
#authors {
font-size: 1.1em;
color: #6b7280;
margin: 0;
}
#main-container {
max-width: 1200px;
margin: 0 auto;
padding: 0 1rem;
gap: 2rem; /* Add gap for consistent spacing */
}
#controls-row {
justify-content: center;
gap: 2rem;
}
/* Custom styles for Radio Button Groups */
#controls-row > div {
background-color: var(--background-card);
background-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='40' height='40' viewBox='0 0 40 40'%3E%3Cg fill-rule='evenodd'%3E%3Cg fill='%23e5e7eb' fill-opacity='0.3'%3E%3Cpath d='M0 38.59l2.83-2.83 1.41 1.41L1.41 40H0v-1.41zM0 1.4l2.83 2.83 1.41-1.41L1.41 0H0v1.41zM38.59 40l-2.83-2.83 1.41-1.41L40 38.59V40h-1.41zM40 1.41l-2.83 2.83-1.41-1.41L38.59 0H40v1.41zM20 18.6l2.83-2.83 1.41 1.41L21.41 20l2.83 2.83-1.41 1.41L20 21.41l-2.83 2.83-1.41-1.41L18.59 20l-2.83-2.83 1.41-1.41L20 18.59z'/%3E%3C/g%3E%3C/g%3E%3C/svg%3E");
border: 1px solid var(--border-color);
border-radius: 16px;
padding: 1rem;
box-shadow: 0 4px 12px rgba(0,0,0,0.05);
}
#controls-row .gradio-button {
border-radius: 10px !important;
transition: background-color 0.2s ease, color 0.2s ease;
}
#controls-row .gradio-button.selected {
background: var(--accent-color) !important;
color: white !important;
border-color: var(--accent-color) !important;
}
#content-row {
gap: 1.5rem;
}
.card {
background-color: var(--background-card);
background-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='40' height='40' viewBox='0 0 40 40'%3E%3Cg fill-rule='evenodd'%3E%3Cg fill='%23e5e7eb' fill-opacity='0.3'%3E%3Cpath d='M0 38.59l2.83-2.83 1.41 1.41L1.41 40H0v-1.41zM0 1.4l2.83 2.83 1.41-1.41L1.41 0H0v1.41zM38.59 40l-2.83-2.83 1.41-1.41L40 38.59V40h-1.41zM40 1.41l-2.83 2.83-1.41-1.41L38.59 0H40v1.41zM20 18.6l2.83-2.83 1.41 1.41L21.41 20l2.83 2.83-1.41 1.41L20 21.41l-2.83 2.83-1.41-1.41L18.59 20l-2.83-2.83 1.41-1.41L20 18.59z'/%3E%3C/g%3E%3C/g%3E%3C/svg%3E");
border: 1px solid var(--border-color);
border-radius: 16px;
padding: 1.5rem;
box-shadow: 0 4px 12px rgba(0,0,0,0.05);
height: 100%;
display: flex;
flex-direction: column;
gap: 1rem;
}
.card-title {
font-weight: 600;
font-size: 1.2rem;
color: var(--text-color);
padding-bottom: 0.75rem;
border-bottom: 1px solid var(--border-color);
}
#input-text textarea {
flex-grow: 1;
border: none !important;
box-shadow: none !important;
padding: 0 !important;
font-size: 1.1em;
line-height: 1.7;
}
#result-html {
flex-grow: 1;
font-size: 1.1em;
line-height: 1.7;
overflow-y: auto;
height: 520px;
}
#input-footer {
display: flex;
justify-content: space-between;
align-items: center;
margin-top: auto; /* Push to bottom */
}
#char-counter {
font-size: 0.9em;
color: #9ca3af;
}
#char-counter.error {
color: #ef4444;
}
#submit-btn {
flex-grow: 1;
max-width: 200px;
font-size: 1.05em;
font-weight: 600;
background: var(--accent-color);
color: white;
border-radius: 10px;
}
#submit-btn:hover {
background: #4f46e5;
}
.disclaimer {
text-align: center;
margin: 0 auto; /* Remove vertical margins */
color: #64748b;
font-size: 1.1em;
max-width: 800px;
}
/* Reveal 动画更丝滑 */
@keyframes reveal {
from { opacity: 0; }
to { opacity: 1; }
}
.reveal-char {
opacity: 0;
animation: reveal 0.2s forwards;
white-space: pre-wrap;
}
""") as demo:
with gr.Column(elem_id="main-container"):
gr.Markdown("""
<div id="header">
<h1 id="title"><span class="detect-grad">Detect</span><span class="anyllm-grad">AnyLLM</span>: Towards Generalizable and Robust Detection of Machine-Generated Text Across Domains and Models</h1>
<p id="authors">Jiachen Fu, Chun-Le Guo, Chongyi Li</p>
</div>
""")
with gr.Row(elem_id="controls-row"):
language_radio = gr.Radio(
choices=["English", "Chinese"],
value="English",
label="🌐 Language",
interactive=True
)
mode_radio = gr.Radio(
choices=["Text-Only", "LaTex"],
value="Text-Only",
label="✍️ Input Type",
interactive=True
)
with gr.Row(equal_height=True, elem_id="content-row"):
with gr.Column(scale=1, min_width=500):
with gr.Column(elem_classes="card"):
gr.HTML('<div class="card-title">📝 Input</div>')
upload_btn = gr.File(
label="Upload File (txt, docx)",
file_types=['.txt', '.docx'],
elem_id="upload-btn"
)
input_text = gr.Textbox(
show_label=False,
placeholder="Enter text to detect or upload a file...",
lines=15,
elem_id="input-text",
max_length=100000,
)
with gr.Row(elem_id="input-footer"):
counter_html = gr.HTML("<div id='char-counter'>0/100000</div>")
submit_btn = gr.Button("✨ Detect", variant="primary", elem_id="submit-btn")
with gr.Column(scale=1, min_width=500):
with gr.Column(elem_classes="card"):
gr.HTML('<div class="card-title">🔍 Result</div>')
result = gr.HTML(elem_id="result-html")
gr.HTML("""
<div class="disclaimer">
💡 <i><b style="color: red;">Red fonts</b> indicate a high probability of AI generation. <b style="color: orange;">Orange fonts</b> indicate a high probability of AI revision or polishing. The detection results are for reference only.</i>
</div>
""")
upload_btn.upload(
read_file_content,
inputs=upload_btn,
outputs=input_text
)
input_text.input(
None,
[input_text],
None,
js="""
(text) => {
setTimeout(() => {
const counter = document.getElementById("char-counter");
if (counter) {
const length = text.length;
counter.innerHTML = `${length}/100000`;
counter.classList.toggle("error", length > 100000);
}
}, 0);
return text;
}
"""
)
submit_btn.click(
greet,
inputs=[mode_radio, language_radio, input_text],
outputs=result
)
demo.launch(share=True)