The Impact of Prompts on Zero-Shot Detection of AI-Generated Text

This project is designed with reproducibility in mind and can be used for paper reproduction purposes. Below is a list of available detectors:

• DetectGPT
• FastDetectGPT
• NPR
• LRR
• FastNPR
• Binoculars
• Log-likelihood
• Rank
• Log-Rank
• Entropy
• Fine-tuned RoBERTa-large for GPT2
• RADAR
• Intrinsic-Dimension

How to use

0. First, You need to create directories.

mkdir results plots txtdata

1. Run the following code to obtain results for the Zero-shot Detector without Binoculars:

python detection.py

2. If you specifically need results from the Binoculars detector, execute the following command:

python detection_binoculars_only.py

3. To enable the use of other detectors, set the is_{detector} flag in detection.py. This allows you to experiment with different methods.

For white-box settings, use the following code:

python detection.py --prompt

Acknowledgement

Some code in this projects is derived from or refers to the work of the following repositories:

• https://github.com/eric-mitchell/detect-gpt
• https://github.com/baoguangsheng/fast-detect-gpt
• https://github.com/ArGintum/GPTID
• https://github.com/ahans30/Binoculars

Feel free to explore and enhance the functionality based on your specific needs. If you have any questions or suggestions, please don't hesitate to reach out.