DACTYL: Diverse Adversarial Corpus of Texts Yielded from Large language models

---

Overview

This repository contains the code and resources for my master's thesis at Cambridge: DACTYL: Diverse Adversarial Corpus of Texts Yielded from Large language models.

Datasets

• DACTYL dataset for AI-text detection
• DACTYL dataset for continued-pretraining

Models

• DACTYL-trained AI-text detectors

• Domain-Specific Llama 3.2 1B Instruct models

Generating DACTYL

We used the dactyl-generation package to generate texts. This can be installed via pip:

pip install dactyl-generation

Code

• finetuning: Continues pre-training a Llama 3.2 1B Instruct model for a specific domain.
• baselines: Contains code to get predictions from existing AI-text detectors on the DACTYL dataset.
• cpt_generations: Performs a randomized generation parameter sweep to determine which parameters evade detection better for the Llama 3.2 1B Instruct models.
• training: Trains an AI-text classifier.
• evaluation: Evaluates DACTYL-trained classifiers.

Citation

@misc{thorat2025dactyldiverseadversarialcorpus,
      title={DACTYL: Diverse Adversarial Corpus of Texts Yielded from Large Language Models}, 
      author={Shantanu Thorat and Andrew Caines},
      year={2025},
      eprint={2508.00619},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.00619}, 
}