license: mit
dataset_info:
features:
- name: tags
dtype: string - name: en
dtype: string - name: ko
dtype: string - name: sw
dtype: string
splits: - name: train
num_bytes: 78398
num_examples: 314 - name: test
num_bytes: 19955
num_examples: 82
download_size: 67010
dataset_size: 98353
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: test
path: data/test-*
- split: train
Multi_jailbreak_augmented
The dataset originates from the Multilingual Jailbreak Challenges in Large Language Models paper and is publicly available at DAMO-NLP-SG/MultiJail, which contains samples in multiple languages. For this work, a subset of the data was selected, focusing specifically on three languages categorized by resource level:
- English (high-resource)
- Korean (medium-resource)
- Swahili (low-resource)
8 prohibited scenarios
The tagging scheme follows the categories defined in the "A Cross-Language Investigation into Jailbreak Attacks in Large Language Models" paper. Instead of using the original dataset tags, the data was reclassified into the following scenario-based categories to better match research goals and improve semantic clarity:
- AdultContent (AC)
- Fraudulent Deceptive Activities (FDA)
- Government Decision Making (GDM)
- HarmfulContent (HC)
- IllegalActivity (IA)
- Political Campaigning Lobbying (PCL)
- UnlawfulPractice (UP)
- ViolatingPrivacy (VP)
Methodology
- Classification into these tags was performed using GPT-4o, ensuring consistent semantic labeling aligned with the task of investigating jailbreak attacks.
- To balance the dataset, augmentation was performed using GPT-4o, which generated at least 30 new samples per tag in a one-shot manner, guided by the provided descriptions.
- Semantic similarity between original and augmented samples was measured via sentence embeddings, discarding any augmented samples below a 0.75 similarity threshold to maintain data quality and semantic integrity.