This repository contains the official code and data for the paper "GPT-generated Text Detection: Benchmark Dataset and Tensor-based Detection Method" published in the WWW 2024 short paper track.
The Reddit dataset is available in reddit_datasets, where the filtered and unfiltered variants of the dataset can be found.
Each of the files is a Comma Separated Values (CSV) file with two columns:
Data: The text of the snippet.Label: The label of the snippet, where0indicates that the snippet is human-written and1indicates that the snippet is GPT-generated.
The reddit data scrapers are located in the scrapers folder. The dataset construction and cleaning process is facilitated through the Load_Full_Dataset.ipynb notebook.
Please refer to the paper for more details on how the dataset was collected and processed.
The code for our method, GpTen, is primarily available in the pipeline.py script. If the input tensor is too large to be processed by the pipeline.py script, you can use the decomp.m MATLAB script to decompose the tensor into its factors. After decomposition, you can use the norms.m MATLAB script to calculate the reconstruction error. This allows you to handle larger datasets/tensors that might otherwise be un-manageable by the python pipeline script.
The required Python packages and versions can be found in the requirements.txt file. For the MATLAB scripts, tensor_toolbox is required. Please note that MATLAB 2021b or higher is required to run the decomp.m and norms.m scripts.
The data itself is sourced from Reddit and from OpenAI's GPT3.5.
If you use this code or data, please cite the following paper:
TODO