metadata
license: mit
task_categories:
- text-generation
language:
- en
size_categories:
- 100K<n<1M
DACTYL Pretraining Corpus
This corpus contains human texts that have passed a quality check from the textdescriptives library.
These texts have been used to further train various Llama 3.2 1B Instruct models by domain and split. For a domain/split combination, you can use the corresponding model:
ShantanuT01/fine-tuned-Llama-3.2-1B-Instruct-apollo-mini-{domain}-{split}
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},
}