Awesome

Yafu Li^†^‡, Qintong Li^§, Leyang Cui^¶, Wei Bi^¶,Zhilin Wang^$<br>

Longyue Wang^¶, Linyi Yang^‡, Shuming Shi^¶, Yue Zhang^‡<br>

^† Zhejiang University, ^‡ Westlake University, ^§ The University of Hong Kong, ^$ Jilin University, ^¶ Tencent AI Lab

Presenting a comprehensive benchmark dataset designed to assess the proficiency of AI-generation detectors amidst real-world scenarios. Welcome to try detection via our online demo!

📌 Table of Contents

• Introduction
• Activities
• Dataset
• Try Detection
• Data Samples
• Citation

🚀 Introduction

Recent advances in large language models have enabled them to reach a level of text generation comparable to that of humans. These models show powerful capabilities across a wide range of content, including news article writing, story generation, and scientific writing. Such capability further narrows the gap between human-authored and machine-generated texts, highlighting the importance of machine-generated text detection to avoid potential risks such as fake news propagation and plagiarism. In practical scenarios, the detector faces texts from various domains or LLMs without knowing their sources.

To this end, we build a comprehensive testbed for deepfake text detection, by gathering texts from various human writings and deepfake texts generated by different LLMs. This repository contains the data to testify deepfake detection methods described in our paper, MAGE: Machine-generated Text Detection in the Wild. Welcome to test your detection methods on our testbed!

📅 Activities

• 🎉 May 16, 2024: Our paper was accepted by ACL 2024!
• 🎉 June 19, 2023: Update two 'wilder' testbeds! We go one step wilder by constructing an additional testset with texts from unseen domains generated by an unseen model, to testify the detection ability in more practical scenarios. We consider four new datasets: CNN/DailyMail, DialogSum, PubMedQA and IMDb to test the detection of deepfake news, deepfake dialogues, deepfake scientific answers and deepfake movie reviews. We sample 200 instances from each dataset and use a newly developed LLM, i.e., GPT-4, with specially designed prompts to create deepfake texts, establishing an "Unseen Domains & Unseen Model" scenario. Previous work demonstrates that detection methods are vulnerable to being deceived by target texts. Therefore, we also paraphrase each sentence individually for both human-written and machine-generated texts, forming an even more challenging testbed. We adopt gpt-3.5-trubo as the zero-shot paraphraser and consider all paraphrased texts as machine-generated.
• May 25, 2023: Initial dataset release including texts from 10 domains and 27 LLMs, contributing to 6 testbeds with increasing detection difficulty.

📝 Dataset

The dataset consists of 447,674 human-written and machine-generated texts from a wide range of sources in the wild:

• Human-written texts from 10 datasets covering a wide range of writing tasks, e.g., news article writing, story generation, scientific writing, etc.
• Machine-generated texts generated by 27 mainstream LLMs from 7 sources, e.g., OpenAI, LLaMA, and EleutherAI, etc.
• 8 systematic testbeds with increasing wildness and detection difficulty.

📥 How to Get the Data

1. Huggingface

You can access the full dataset, which includes the Cross-domains & Cross-models testbed and two additional wilder test sets, through the Huggingface API:

from datasets import load_dataset
dataset = load_dataset("yaful/MAGE")

which includes traditional splits (train.csv, valid.csv and test.csv) and two wilder test sets (test_ood_set_gpt.csv and test_ood_set_gpt_para.csv). The csv files have three columns: text, label (0 for machine-generated and 1 for human-written) and text source information (e.g., ''cmv_human'' denotes the text is written by humans, whereas ''roct_machine_continuation_flan_t5_large'' denotes the text is generated by ''flan_t5_large'' using continuation prompt).

To obtain the 6 testbeds mentioned in our paper, simply apply the provided script:

python3 deployment/prepare_testbeds.py DATA_PATH

Replace ''DATA_PATH'' with the output data directory where you want to save the 6 testbeds.

2. Cloud Drive

Alternatively, you can access the 6 testbeds by downloading them directly through Google Drive or Tencent Weiyun：

The folder contains 4 packages:

• testbeds_processed.zip: 6 testbeds based on the ''processed'' version, which can be directly used for detecting in-distribution and out-of-distribution detection performance.
• wilder_testsets.zip: 2 wilder test sets with texts processed, aiming for (1) detecting deepfake text generated by GPT-4, and (2) detecting deepfake text in paraphrased versions.
• source.zip: Source texts of human-written texts and corresponding texts generated by LLMs, without filtering.
• processed.zip: This is a refined version of the "source" that filters out low-quality texts and specifies sources as CSV file names. For example, the "cmv_machine_specified_gpt-3.5-trubo.csv" file contains texts from the CMV domain generated by the "gpt-3.5-trubo" model using specific prompts, while "cmv_human" includes human-written CMV texts.

:computer: Try Detection

Python Environment

For deploying the Longformer detector or training your own detector using our data, simply install the following packages:

pip install transformers
pip install datasets
pip install clean-text # for data preprocessing

Or you can run:

pip install -r requirements.txt

Model Access

Our Longformer detector, which has been trained on the entire dataset, is now accessible through Huggingface. Additionally, you can try detection directly using our online demo.

We have refined the decision boundary based on out-of-distribution settings. To ensure optimal performance, we recommend preprocessing texts before sending them to the detector.

import torch
import os
from transformers import AutoModelForSequenceClassification,AutoTokenizer
from deployment import preprocess, detect

# init
device = 'cpu' # use 'cuda:0' if GPU is available
# model_dir = "nealcly/detection-longformer" # model in our paper
model_dir = "yaful/MAGE" # model in the online demo
tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForSequenceClassification.from_pretrained(model_dir).to(device)

text = "Apple's new credit card will begin a preview roll out today and will become available to all iPhone owners in the US later this month. A random selection of people will be allowed to go through the application process, which involves entering personal details which are sent to Goldman Sachs and TransUnion. Applications are approved or declined in less than a minute. The Apple Card is meant to be broadly accessible to every iPhone user, so the approval requirements will not be as strict as other credit cards. Once the application has been approved, users will be able to use the card immediately from the Apple Wallet app. The physical titanium card can be requested during setup for free, and it can be activated with NFC once it arrives."
# preprocess
text = preprocess(text)
# detection
result = detect(text,tokenizer,model,device)

Detection Performance

In-distribution Detection

Out-of-distribution Detection

Wilder Testsets

📄 Data Samples

All instances are stored as rows in a csv format, with each row consiting of 3 columns: Text, Label (0 for machine-generated and 1 for human-written) and Index (indexes in the original data source, used restore alignment after filtering).

News Article

Opinion Statement

Long-form Answer

Story Generation

Scientific Writing

📚 Citation

If you use this dataset in your research, please cite it as follows:

@inproceedings{li-etal-2024-mage,
    title = "{MAGE}: Machine-generated Text Detection in the Wild",
    author = "Li, Yafu  and
      Li, Qintong  and
      Cui, Leyang  and
      Bi, Wei  and
      Wang, Zhilin  and
      Wang, Longyue  and
      Yang, Linyi  and
      Shi, Shuming  and
      Zhang, Yue",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.3",
    doi = "10.18653/v1/2024.acl-long.3",
    pages = "36--53",
}

We welcome contributions to improve this dataset! If you have any questions or feedback, please feel free to reach out at yafuly@gmail.com.