license: cc-by-4.0
tags:
- cybersecurity
- multi-turn
- adversarial-ai
- tense-reformulation
- jailbreaking
pretty_name: CyMulTenSet
size_categories: - n<1K
CyMulTenSet: Grammatical Mirage Attack
Associated Paper: Jailbreaking LLMs Through Tense Manipulation in Multi-Turn Dialogues
Dataset Summary
This dataset is part of the Grammatical Mirage Attack for multi-turn jailbreaking research on large language models (LLMs).
This dataset is an extension of single-turn jailbreaking attacks from AdvBench and HarmBench to multi-turn jailbreaking attacks, focusing exclusively on cybersecurity contexts.
It contains adversarial prompts in cybersecurity contexts reformulated in both present and past tense, and structured across 1-turn, 2-turn, and 3-turn conversations.
It can be used to benchmark the robustness of LLMs against multi-turn jailbreaking attacks.
Total rows: 762
Base prompts: 127
Turns: 1, 2, 3
Tenses: Present, Past
Column Descriptions
| Column Name | Description |
|---|---|
| base_id | Original prompt index (0–126) |
| conversation_id | Unique conversation identifier (base_id_turn_tense) |
| turn_depth | Number of turns in the conversation (1–3) |
| tense | Tense of the conversation (present or past) |
| semantic_category | Broad cybersecurity category (from AdvBench/HarmBench) |
| subtopic | Specific cybersecurity topic: Hacking, Malware, or Phishing |
| query_type | Instruction (how-to steps) or Script (code/script) |
| source | Data source (AdvBench/HarmBench/Original) |
| turn_1 | First step in the multi-turn conversation |
| turn_2 | Second step (empty if turn_depth < 2) |
| turn_3 | Third step (empty if turn_depth < 3) |
Usage
Option 1: Load directly from Hugging Face repository:
import pandas as pd
# Direct URL to raw CSV from Hugging Face
url = "https://huggingface.co/datasets/Micdejc/cymultenset/resolve/main/cymultenset.csv"
df = pd.read_csv(url)
print(df.head())
Option 2: Use Hugging Face datasets library:
from datasets import load_dataset
# Load dataset as a Hugging Face Dataset object
dataset = load_dataset("Micdejc/cymultenset", data_files="cymultenset.csv")
# Convert to pandas DataFrame
df = dataset["train"].to_pandas()
print(df.head())
Test
Example 1: Filter AdvBench conversations
# Filter AdvBench conversations
df_advbench = df[(df.source=='AdvBench')]
print(df_advbench.shape)
Example 2: Filter 2-turn present conversations
# Filter 2-turn present conversations
df_2turn_present = df[(df.turn_depth==2) & (df.tense=='present')]
print(df_2turn_present.shape)
Citation
If you find this dataset useful in your research, please consider citing:
@inproceedings{tchuindjang2026mirage,
title = {Jailbreaking LLMs Through Tense Manipulation in Multi-turn Dialogues},
author = {Tchuindjang, Michael and Duran, N. and Legg, P. and Medjek, F.},
booktitle = {Advances in Computational Intelligence Systems},
editor = {Hart, E. and Horvath, T. and Tan, Z. and Thomson, S.},
series = {Advances in Intelligent Systems and Computing},
volume = {1468},
year = {2026},
publisher = {Springer},
address = {Cham},
doi = {10.1007/978-3-032-07938-1_24}
}