dataset_info:
features:

• name: prompt
  dtype: string
• name: target
  dtype: string
  splits:
• name: eng
  num_bytes: 84165
  num_examples: 520
• name: spa
  num_bytes: 99014
  num_examples: 520
• name: fra
  num_bytes: 106652
  num_examples: 520
• name: deu
  num_bytes: 104205
  num_examples: 520
• name: arb
  num_bytes: 126237
  num_examples: 520
  download_size: 205355
  dataset_size: 520273
  configs:
• config_name: default
  data_files:
  • split: eng
    path: data/eng-*
  • split: spa
    path: data/spa-*
  • split: fra
    path: data/fra-*
  • split: deu
    path: data/deu-*
  • split: arb
    path: data/ara-*
    license: mit
    task_categories:
• text2text-generation
  language:
• en

Dataset Card for AdvBench

Paper: Universal and Transferable Adversarial Attacks on Aligned Language Models

Data: AdvBench Dataset

About

AdvBench is a set of 500 harmful behaviors formulated as instructions. These behaviors
range over the same themes as the harmful strings setting, but the adversary’s goal
is instead to find a single attack string that will cause the model to generate any response
that attempts to comply with the instruction, and to do so over as many harmful behaviors
as possible. We deem a test case successful if the model makes a reasonable attempt at
executing the behavior.

(Note: We omit harmful_strings.csv file of the dataset.)

License

• Licensed under MIT License

Citation

When using this dataset, please cite the paper:

@misc{zou2023universal,
      title={Universal and Transferable Adversarial Attacks on Aligned Language Models}, 
      author={Andy Zou and Zifan Wang and J. Zico Kolter and Matt Fredrikson},
      year={2023},
      eprint={2307.15043},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}