[NeurIPS 2024] DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning

Note: Our latest work, DETree: DEtecting Human-AI Collaborative Texts via Tree-Structured Hierarchical Representation Learning, has also been accepted to NeurIPS 2025. For more details, please refer to the project.

📌 Table of Contents

• Introduction
• Dataset
• Installation
• Testing
• Training
• Citation

🚀 Introduction

Recent advances in large language models (LLMs) have enabled the generation of text that closely resembles human writing, raising challenges for AI-generated text detection. DeTeCtive addresses these challenges with a multi-level contrastive learning framework, distinguishing nuanced writing styles to improve detection accuracy across various domains and models. This repository contains the code and models from our paper, DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning.

📝 Dataset

Before starting, ensure that the required datasets are prepared correctly. This repository uses four primary datasets:

• Deepfake
• SemEval2024-M4
• TuringBench
• OUTFOX

You can download the pre-processed versions of these datasets from Google Drive. Once downloaded, unzip them into a single datasets directory, structured as follows:

$DATA/
|–– Deepfake/
|   |–– cross_domains_cross_models/
|   |–– cross_domains_model_specific/
|   |–– domain_specific_cross_models/
|   |–– domain_specific_model_specific/
|   |–– unseen_domains/
|   |–– unseen_models/
|–– SemEval2024-M4/
|   |–– SubtaskA/
|–– TuringBench/
|   |–– AA/
|–– OUTFOX/
    |–– chatgpt/
    |–– common/
    |–– dipper/
    |–– flan_t5_xxl/
    |–– text_davinci_003/

🔄 Installation

To set up the environment, you can use requirements.txt to install all necessary dependencies. Alternatively, if you already have a PyTorch environment, you can install the core libraries separately:

Using Conda:

conda create -n detective python=3.10
conda activate detective
pip install -r requirements.txt

Using Virtual Environment (venv):

python -m venv env
source env/bin/activate
pip install -r requirements.txt

Core Libraries

If PyTorch is already installed, you can simply install the two core libraries:

pip install lightning faiss-gpu

🔬 Testing

We have released checkpoint models on Hugging Face at heyongxin233/DeTeCtive. Each dataset has an associated checkpoint that enables you to replicate the performance metrics reported in the paper. Download the checkpoint for the dataset you want to evaluate.

Available Checkpoints:

• Deepfake_best.pth
• M4_monolingual_best.pth
• M4_multilingual_best.pth
• OUTFOX_best.pth
• TuringBench_best.pth

After downloading, place the checkpoints in a directory accessible to the testing scripts.

Once the checkpoints are downloaded, use script/test.sh to run the evaluation. Modify the DATA_PATH and Model_PATH in the script to point to the location of your dataset and the downloaded checkpoint model. Choose the dataset you wish to evaluate, then run the script.

Running the Test Script

To initiate testing, execute:

bash script/test.sh

The test.sh script performs KNN classification by first embedding the training set and then running the test set. If the --save_database option is enabled, it will save the embedded database at the specified save_path. For future evaluations, you can skip the training embedding phase by loading the pre-saved database using script/test_from_database.sh.

Running from a Pre-saved Database

To evaluate directly from a saved database, use:

bash script/test_from_database.sh

This script operates similarly to test.sh but loads the training embeddings from the pre-saved database, reducing the computational load by skipping the embedding step for the training data.

🔍 Detect Text

To perform inference on a single text, you first need to generate the corresponding database based on your requirements. You can do this by running script/gen_database.sh to generate the database for each dataset (which is the same as the database generated by script/test.sh). After generating the database, add the database path and the text to be inferred into script/infer.sh, and you can proceed with the single text inference.

bash script/gen_database.sh
bash script/infer.sh

💻 Training

To train the model, simply run the script script/train.sh. Before running, make sure to set the DATA_PATH parameter in the script to point to the location of your dataset. Then, select the dataset you want to use for training.

bash script/train.sh

Configurable Parameters

Some training parameters may require adjustment based on your setup. You can view and modify these parameters by referring to the help in train_classifier.py. Run the following command for more information on available parameters:

python train_classifier.py --help

This will display descriptions of the various configurable options that you can set to optimize training for your specific dataset and hardware configuration.

📚 Citation

If you use our code or findings in your research, please cite us as:

@article{nips2024detective,
  title={DeTeCtive: Detecting AI-generated Text via Multi-Level Contrastive Learning},
  author={Guo, Xun and Zhang, Shan and He, Yongxin and Zhang, Ting and Feng, Wanquan and Huang, Haibin and Ma, Chongyang},
  journal={arXiv preprint arXiv:2410.20964},
  year={2024}
}