ontocord/auto_redteam_triplets

Auto Redteam Triplets (ART): a synthetic dataset to perform reinforcement learning redteaming for the EU AI Act and Biden-Harris AI Executive Order concerns

Dataset Summary

The ART dataset contains triplets of prompts and responses related to concerns under the proposed EU AI Act and Biden-Harris AI Executive Order concerns, in addition to common redteaming concerns.

Column 'text' is the positive response answer pair, and text2 and text3 are both rejections. These columns can be used to perform reinforcement learning, or simply text can be used to do supervised finetuning.

More details to come.

YOU MUST BE 18 YEARS OR OLDER TO DOWNLOAD OR USE THIS DATASET

This dataset contains sensitive and triggering content.

Limitations

This is a work-in-progress dataset which means there can be issues related to style, quality, length of response, or accuracy. So please be aware and review the data before using.

We are actively improving the dataset and recognize that both the EU AI Act and the Biden-Harris AI Executive Order are still being fully fleshed out - with local legislation in the first case and regulatory guidence development in the second case. Thus, we base our work on the general concerns provided in the text of the respective Act and Order. And we will improve the datasets as the law evolves. Please note that use of this dataset is not given as legal advice nor a gurantee of compliance with law.

Citation

To cite the ART dataset, please use:

@misc{ontocord2024art,
      title={Auto Redteam Triplets (ART): a synthetic dataset to perform reinforcement learning redteaming for the EU AI Act and Biden-Harris AI Executive Order concerns}, 
      author={Ontocord.AI},
      year={2024},
}