license: mit
language:

• en
  task_categories:
• text-classification
  tags:
• safety
• jailbreak
• dialogue
• embeddings
• llm-safety
• multi-turn
• sentence-transformers
  size_categories:
• 1K<n<10K

NBF-LLM Dialogue Embedding Data

Training and validation dialogue embedding data from the paper:

Steering Dialogue Dynamics for Robustness against Multi-turn Jailbreaking Attacks
Hanjiang Hu, Alexander Robey, Changliu Liu
arXiv:2503.00187 | GitHub

Dataset Description

This dataset contains pre-computed sentence-transformer (all-mpnet-base-v2, 768-dim) embeddings of multi-turn jailbreaking dialogue trajectories. Each trajectory is a list of dialogue steps; each step is a dict with:

• u: user turn embedding (torch.Tensor, shape [768])
• y: assistant turn embedding (torch.Tensor, shape [768])
• score: safety score (int, 1–5)

Files

Usage

from huggingface_hub import hf_hub_download
import torch

# Download training data
cb_actorattack = torch.load(
    hf_hub_download("hanjianghu/NBF-LLM-Dialogue-Embedding", "circuit_breakers_actorattack.pt"),
    map_location="cpu", weights_only=False
)
cb_others = torch.load(
    hf_hub_download("hanjianghu/NBF-LLM-Dialogue-Embedding", "circuit_breakers_others.pt"),
    map_location="cpu", weights_only=False
)

# Each is a list of trajectories; each trajectory is a list of step dicts
print(f"Train trajectories: {len(cb_actorattack) + len(cb_others)}")
step = cb_actorattack[0][0]
print(f"Step keys: {step.keys()}")  # dict_keys(['u', 'y', 'score'])
print(f"Embedding dim: {step['u'].shape}")  # torch.Size([768])

Data Collection

Multi-turn attack dialogues were collected from:

• ActorAttack
• Automated-Multi-Turn-Jailbreaks (Crescendo, Acronym, Opposite-Day)

Citation

@article{hu2025steering,
  title={Steering Dialogue Dynamics for Robustness against Multi-turn Jailbreaking Attacks},
  author={Hu, Hanjiang and Robey, Alexander and Liu, Changliu},
  journal={arXiv preprint arXiv:2503.00187},
  year={2025}
}