pretty_name: JailbreakDB
license: cc-by-4.0
tags:

• llm-safety
• jailbreak
• prompts
• security
• text-classification
  language:
• en

JailbreakDB

Two large-scale, curated text-only datasets for LLM safety research:

• text_jailbreak_unique.csv (~6.6M rows): jailbreak/adversarial prompts
• text_regular_unique.csv (~5.7M rows): benign prompts

Each record includes: system_prompt, user_prompt, jailbreak (0/1), source, tactic.

This release provides clean, deduplicated text data only. No engineered features or experimental results are included.

Paper

• ArXiv: https://arxiv.org/abs/2510.15476

Citation

If you find this resource helpful, please cite our arXiv paper:

@misc{hong2025sok,
  title         = {SoK: Taxonomy and Evaluation of Prompt Security in Large Language Models},
  author        = {Hong, Hanbin and Feng, Shuya and Naderloui, Nima and Yan, Shenao and Zhang, Jingyu and Liu, Biying and Arastehfard, Ali and Huang, Heqing and Hong, Yuan},
  year          = {2025},
  eprint        = {2510.15476},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  url           = {https://arxiv.org/abs/2510.15476}
}

Loading with datasets

from datasets import load_dataset

files = {
    "jailbreak": "hf://datasets/haorandai/JailbreakDB/text_jailbreak_unique.csv",
    "regular":   "hf://datasets/haorandai/JailbreakDB/text_regular_unique.csv",
}
ds = load_dataset("csv", data_files=files)
print(ds)

Safety Notice

This dataset may contain harmful, offensive, or disturbing content and is intended strictly for research on model safety and robustness. Please review your institutional and legal requirements before use.