# Copyright (c) Hello-SimpleAI Org. 2023. # Licensed under the Apache License, Version 2.0. import os import pickle import re from typing import Callable, List, Tuple import gradio as gr from nltk.data import load as nltk_load import numpy as np from sklearn.linear_model import LogisticRegression import torch from transformers.utils import cached_file from transformers import GPT2LMHeadModel, GPT2Tokenizer AUTH_TOKEN = os.environ.get("access_token") DET_LING_ID = 'Hello-SimpleAI/chatgpt-detector-ling' def download_file(filename): return cached_file(DET_LING_ID, filename, use_auth_token=AUTH_TOKEN) NLTK = nltk_load(download_file('english.pickle')) sent_cut_en = NLTK.tokenize LR_GLTR_EN, LR_PPL_EN, LR_GLTR_ZH, LR_PPL_ZH = [ pickle.load(open(download_file(f'{lang}-gpt2-{name}.pkl'), 'rb')) for lang, name in [('en', 'gltr'), ('en', 'ppl'), ('zh', 'gltr'), ('zh', 'ppl')] ] NAME_EN = 'gpt2' TOKENIZER_EN = GPT2Tokenizer.from_pretrained(NAME_EN) MODEL_EN = GPT2LMHeadModel.from_pretrained(NAME_EN) NAME_ZH = 'IDEA-CCNL/Wenzhong-GPT2-110M' TOKENIZER_ZH = GPT2Tokenizer.from_pretrained(NAME_ZH) MODEL_ZH = GPT2LMHeadModel.from_pretrained(NAME_ZH) # code borrowed from https://github.com/blmoistawinde/HarvestText def sent_cut_zh(para: str) -> List[str]: para = re.sub('([。!?\?!])([^”’)\])】])', r"\1\n\2", para) # 单字符断句符 para = re.sub('(\.{3,})([^”’)\])】….])', r"\1\n\2", para) # 英文省略号 para = re.sub('(\…+)([^”’)\])】….])', r"\1\n\2", para) # 中文省略号 para = re.sub('([。!?\?!]|\.{3,}|\…+)([”’)\])】])([^,。!?\?….])', r'\1\2\n\3', para) # 如果双引号前有终止符,那么双引号才是句子的终点,把分句符\n放到双引号后,注意前面的几句都小心保留了双引号 para = para.rstrip() # 段尾如果有多余的\n就去掉它 # 很多规则中会考虑分号;,但是这里我把它忽略不计,破折号、英文双引号等同样忽略,需要的再做些简单调整即可。 sentences = para.split("\n") sentences = [sent.strip() for sent in sentences] sentences = [sent for sent in sentences if len(sent.strip()) > 0] return sentences CROSS_ENTROPY = torch.nn.CrossEntropyLoss(reduction='none') def gpt2_features( text: str, tokenizer: GPT2Tokenizer, model: GPT2LMHeadModel, sent_cut: Callable ) -> Tuple[List[int], List[float]]: # Tokenize input_max_length = tokenizer.model_max_length - 2 token_ids, offsets = list(), list() sentences = sent_cut(text) for s in sentences: tokens = tokenizer.tokenize(s) ids = tokenizer.convert_tokens_to_ids(tokens) difference = len(token_ids) + len(ids) - input_max_length if difference > 0: ids = ids[:-difference] offsets.append((len(token_ids), len(token_ids) + len(ids))) # 左开右闭 token_ids.extend(ids) if difference >= 0: break input_ids = torch.tensor([tokenizer.bos_token_id] + token_ids) logits = model(input_ids).logits # Shift so that n-1 predict n shift_logits = logits[:-1].contiguous() shift_target = input_ids[1:].contiguous() loss = CROSS_ENTROPY(shift_logits, shift_target) all_probs = torch.softmax(shift_logits, dim=-1) sorted_ids = torch.argsort(all_probs, dim=-1, descending=True) # stable=True expanded_tokens = shift_target.unsqueeze(-1).expand_as(sorted_ids) indices = torch.where(sorted_ids == expanded_tokens) rank = indices[-1] counter = [ rank < 10, (rank >= 10) & (rank < 100), (rank >= 100) & (rank < 1000), rank >= 1000 ] counter = [c.long().sum(-1).item() for c in counter] # compute different-level ppl text_ppl = loss.mean().exp().item() sent_ppl = list() for start, end in offsets: nll = loss[start: end].sum() / (end - start) sent_ppl.append(nll.exp().item()) max_sent_ppl = max(sent_ppl) sent_ppl_avg = sum(sent_ppl) / len(sent_ppl) if len(sent_ppl) > 1: sent_ppl_std = torch.std(torch.tensor(sent_ppl)).item() else: sent_ppl_std = 0 mask = torch.tensor([1] * loss.size(0)) step_ppl = loss.cumsum(dim=-1).div(mask.cumsum(dim=-1)).exp() max_step_ppl = step_ppl.max(dim=-1)[0].item() step_ppl_avg = step_ppl.sum(dim=-1).div(loss.size(0)).item() if step_ppl.size(0) > 1: step_ppl_std = step_ppl.std().item() else: step_ppl_std = 0 ppls = [ text_ppl, max_sent_ppl, sent_ppl_avg, sent_ppl_std, max_step_ppl, step_ppl_avg, step_ppl_std ] return counter, ppls # type: ignore def lr_predict( f_gltr: List[int], f_ppl: List[float], lr_gltr: LogisticRegression, lr_ppl: LogisticRegression, id_to_label: List[str] ) -> List: x_gltr = np.asarray([f_gltr]) gltr_label = lr_gltr.predict(x_gltr)[0] gltr_prob = lr_gltr.predict_proba(x_gltr)[0, gltr_label] x_ppl = np.asarray([f_ppl]) ppl_label = lr_ppl.predict(x_ppl)[0] ppl_prob = lr_ppl.predict_proba(x_ppl)[0, ppl_label] return [id_to_label[gltr_label], gltr_prob, id_to_label[ppl_label], ppl_prob] def predict_en(text: str) -> List: with torch.no_grad(): feat = gpt2_features(text, TOKENIZER_EN, MODEL_EN, sent_cut_en) out = lr_predict(*feat, LR_GLTR_EN, LR_PPL_EN, ['Human', 'ChatGPT']) return out def predict_zh(text: str) -> List: with torch.no_grad(): feat = gpt2_features(text, TOKENIZER_ZH, MODEL_ZH, sent_cut_zh) out = lr_predict(*feat, LR_GLTR_ZH, LR_PPL_ZH, ['人类', 'ChatGPT']) return out with gr.Blocks() as demo: gr.Markdown( """ ## ChatGPT Detector 🔬 (Linguistic version / 语言学版) Visit our project on Github: [chatgpt-comparison-detection project](https://github.com/Hello-SimpleAI/chatgpt-comparison-detection)
欢迎在 Github 上关注我们的 [ChatGPT 对比与检测项目](https://github.com/Hello-SimpleAI/chatgpt-comparison-detection)
We provide three kinds of detectors, all in Bilingual / 我们提供了三个版本的检测器,且都支持中英文: - [QA version / 问答版](https://huggingface.co/spaces/Hello-SimpleAI/chatgpt-detector-qa)
detect whether an **answer** is generated by ChatGPT for certain **question**, using PLM-based classifiers / 判断某个**问题的回答**是否由ChatGPT生成,使用基于PTM的分类器来开发; - [Sinlge-text version / 独立文本版](https://huggingface.co/spaces/Hello-SimpleAI/chatgpt-detector-single)
detect whether a piece of text is ChatGPT generated, using PLM-based classifiers / 判断**单条文本**是否由ChatGPT生成,使用基于PTM的分类器来开发; - [**Linguistic version / 语言学版** (👈 Current / 当前使用)](https://huggingface.co/spaces/Hello-SimpleAI/chatgpt-detector-ling)
detect whether a piece of text is ChatGPT generated, using linguistic features / 判断**单条文本**是否由ChatGPT生成,使用基于语言学特征的模型来开发; """ ) with gr.Tab("English"): gr.Markdown( """ ## Introduction: Two Logistic regression models trained with two kinds of features: 1. [GLTR](https://aclanthology.org/P19-3019) Test-2, Language model predict token rank top-k buckets, top 10, 10-100, 100-1000, 1000+. 2. PPL-based, text ppl, sentence ppl, etc. English LM is [GPT2-small](https://huggingface.co/gpt2). Note: Providing more text to the `Text` box can make the prediction more accurate! """ ) a1 = gr.Textbox( lines=5, label='Text', value="There are a few things that can help protect your credit card information from being misused when you give it to a restaurant or any other business:\n\nEncryption: Many businesses use encryption to protect your credit card information when it is being transmitted or stored. This means that the information is transformed into a code that is difficult for anyone to read without the right key." ) button1 = gr.Button("🤖 Predict!") gr.Markdown("GLTR") label1_gltr = gr.Textbox(lines=1, label='GLTR Predicted Label 🎃') score1_gltr = gr.Textbox(lines=1, label='GLTR Probability') gr.Markdown("PPL") label1_ppl = gr.Textbox(lines=1, label='PPL Predicted Label 🎃') score1_ppl = gr.Textbox(lines=1, label='PPL Probability') with gr.Tab("中文版"): gr.Markdown( """ ## 介绍: 两个逻辑回归模型, 分别使用以下两种特征: 1. [GLTR](https://aclanthology.org/P19-3019) Test-2, 每个词的语言模型预测排名分桶, top 10, 10-100, 100-1000, 1000+. 2. 基于语言模型困惑度 (PPL), 整个文本的PPL、单个句子的PPL等特征. 中文语言模型使用 闻仲 [Wenzhong-GPT2-110M](https://huggingface.co/IDEA-CCNL/Wenzhong-GPT2-110M). 注意: 在`文本`栏中输入更多的文本,可以让预测更准确哦! """ ) a2 = gr.Textbox( lines=5, label='文本', value="对于OpenAI大力出奇迹的工作,自然每个人都有自己的看点。我自己最欣赏的地方是ChatGPT如何解决 “AI校正(Alignment)“这个问题。这个问题也是我们课题组这两年在探索的学术问题之一。" ) button2 = gr.Button("🤖 预测!") gr.Markdown("GLTR (中文测试集准确率 86.39%)") label2_gltr = gr.Textbox(lines=1, label='预测结果 🎃') score2_gltr = gr.Textbox(lines=1, label='模型概率') gr.Markdown("PPL (中文测试集准确率 59.04%, 持续优化中...)") label2_ppl = gr.Textbox(lines=1, label='PPL 预测结果 🎃') score2_ppl = gr.Textbox(lines=1, label='PPL 模型概率') button1.click(predict_en, inputs=[a1], outputs=[label1_gltr, score1_gltr, label1_ppl, score1_ppl]) button2.click(predict_zh, inputs=[a2], outputs=[label2_gltr, score2_gltr, label2_ppl, score2_ppl]) # Page Count gr.Markdown("""
""") demo.launch() # To create a public link, set `share=True` in `launch()`.