AdvBench-Korean

AdvBench-Korean is a Korean translation of the AdvBench dataset for evaluating jailbreak and adversarial attacks on aligned language models. The prompts were translated into Korean using Google/Gemma-3-12B-Instruct (google/gemma-3-12b-it). All samples were successfully translated without any refusals, preserving the full set of 520 examples from the original dataset. For safety reasons, details of the translation generation pipeline and prompting strategy are not publicly released.

Dataset Structure

Each example contains:

column	description
prompt	harmful instruction (Korean translation)
target	target response used in AdvBench

The dataset contains 520 examples in the training split.

Original Dataset

This dataset is derived from AdvBench, introduced in:

Universal and Transferable Adversarial Attacks on Aligned Language Models

Citation

If you use this dataset, please cite the original AdvBench 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}
}

If you specifically use the Korean translation, you may also cite:

@dataset{park2026advbenchkorean,
  title={AdvBench-Korean: Korean Translation of the AdvBench Dataset},
  author={Bumjin Park},
  year={2026},
  url={https://huggingface.co/datasets/leo-bjpark/AdvBench-Korean}
}