dataset_info:
features:

• name: behavior_id
  dtype: string
• name: prompt
  dtype: string
  splits:
• name: single_turn
  num_bytes: 442059
  num_examples: 500
• name: multi_turn
  num_bytes: 2287355
  num_examples: 500
  download_size: 1072566
  dataset_size: 2729414
  configs:
• config_name: default
  data_files:
  • split: single_turn
    path: data/single_turn-*
  • split: multi_turn
    path: data/multi_turn-*

JBDistill-Bench: A Challenging Benchmark for LLM Safety

Paper: Jailbreak Distillation: Renewable Safety Benchmarking, published in EMNLP 2025 (Findings).

JBDistill-Bench is a challenging benchmark for LLM safety. It is produced by running our proposed Jailbreak Distillation method, which transforms seed goals into adversarial prompts using diverse attacks, and then employ prompt selection algorithms to select an effective set of evaluation prompts.
JBDistill-Bench contains 500 single-turn adversarial prompts and 500 multi-turn adversarial prompts.
This version of JBDistill-Bench contains seed goals from the HarmBench dataset. It is very easy to employ our Jailbreak Distillation method to generate new safety benchmarks using other seed goals as well.

Explanation of dataset columns:

• prompt: attack prompts that will be served as the input to target LLM.
• behavior_id: behavior ID of the original harmbench behavior.

Evaluation Results on Popular LLMs

As shown below, our JBDistill-Bench achieves very high attack success rate (ASR) on a wide range of LLMs. Specifically, we successfully attack OpenAI o1 with >60% ASR, evaluated using the HarmBench judge.

Reference

If you find our work useful, please consider citing our paper:

@article{zhang2025jailbreakdistillation,
      title={Jailbreak Distillation: Renewable Safety Benchmarking}, 
      author={Jingyu Zhang and Ahmed Elgohary and Xiawei Wang and A S M Iftekhar and Ahmed Magooda and Benjamin Van Durme and Daniel Khashabi and Kyle Jackson},
      year={2025},
      url={https://arxiv.org/abs/2505.22037},
      journal = {arXiv preprint}
}