A collection of experiments to assess the quality of explanations for detectors of machine-generated text.

Setup

git clone --recurse-submodules repo_url

cd repo_dir

Models

cd repo_dir

wget https://openaipublic.azureedge.net/gpt-2/detector-models/v1/detector-base.pt models/radford et al/detector-base.pt

Cache

Unzip explanation_cache.zip to explanation_cache. The filenames contain the SHA256 hash of the input string. See fi_explainer.py.

Python

Create a .venv and activate it.

pip install -r requirements.txt

python -m spacy download en_core_web_sm

python -m spacy download en_core_web_lg

You may want to intall pytorch manually.

Make sure to select the .venv in all notebooks.

Dataset

Run dataset_sampling.ipynb to obtain the base dataset from Guo et al. 2023 (CC-BY-SA). The individual notebooks derive additional synthetic datasets (CC-BY-SA).

Detectors

All detectors are extended to support masked input:

• In-domain fine-tuned RoBERTa in Guo et al. 2023: detector_guo.py
• Out-of-domain fine-tuned RoBERTa of Soliman 2019 et al. / Radford et al. 2018: detector_radford.py
• Zero-shot method in Mitchell et al. 2023 with surrogate model as proposed in Mireshghallah et al. 2023: detector_detectgpt.py

Explanation Methods

SHAP is used as-is.

Forks of LIME and Anchor are provided as submodules.

Anchor

• Addition of a budget limiting the number of samples used during search to cap runtime (200 samples per candidate during search, unlimited samples in final "best of each size" round)
• DistillBERT was replaced with DistillRoBERTA and the mask probability adjusted to increase the coherence of perturbations
• Changes to the rendering functions for the user-study (used to share JS and CSS scope with LIME)

LIME

• Cosmetic changes to the bar-charts for the user-study

Experiments

The explanations are provided as a zip file. All experiments are designed so that any subset of the dataset can be processed in parallel by executing the notebooks with different offsets.