Skip to content

A repository implementing the original detect GPT paper using python and pytorch.

Notifications You must be signed in to change notification settings

sarthakforwet/DetectGPT

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

26 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DetectGPT: Distinguishing between Machine-Generated and Human-Written Text

This repository is an extension of the original work done on the detectGPT model by incorporating three datasets on new writing styles and verifying detectGPT's performance on them.

Original implementation of the experiments in the DetectGPT paper.

An interactive demo of DetectGPT can be found here.

Instructions

First, install the Python dependencies:

    python3 -m venv env
    source env/bin/activate
    pip install -r requirements.txt

Second, execute run.py using python run.py and provide the appropriate command line arguments to the script.

If you have new dataset to include, add its inclusion in custom_datasets.py script and then execute run.py as instructed above.

Please refer to the script for more details on how each function is working and what CLI to give.

Here, we extend the original DetectGPT paper. We apply the method on new datasets and compare the results to the ones achieved by the original authors, verifying the algorithm works well on new datasets (especially ones that have a different style of text). Further, we document the original source code and provide a document outlining how we ran our code (and tuned associated hyperparameters).

Future Work:

  1. Improving GPT-2 model using ensemble methods
  2. Further exploring the relationship between prompting and detection
  3. Determing whether negative log likelihood curvature is present for generative models in other domains: audio, video and images.

About

A repository implementing the original detect GPT paper using python and pytorch.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published