Skip to content

Dataset · ScaleAI

Multi-turn jailbreaks for testing LLM defense

ScaleAI/mhj

A collection of successful multi-turn attack dialogues from red teamers for testing language model safety.

What's inside

Inside are 2,912 prompts combined into 537 multi-turn dialogues in English. The authors included only successful attacks devised by red teaming specialists, and added metadata with their comments on choosing the hacking tactics. The card metadata indicates the image modality, but by description these are text prompts and dialogues.

What you can do with it

  • Evaluating LLM resistance to multi-turn jailbreaks
  • Training or fine-tuning safety classifier models (guardrails)
  • SFT or RLHF to increase language model defense against attacks
Tasks
Text generation
Languages
English
Size
up to 1K rows
Downloads this month
473
Likes
38
Access
on request on Hugging Face

How to load it

from datasets import load_dataset

ds = load_dataset("ScaleAI/mhj")