Hugging Face's logo

Datasets:
keithtwesigye
/
wildjailbreak-africa

Tasks:
Text Generation
Languages:
English
Acoli
Lugbara
Size:
100K<n<1M
Tags:
safety
jailbreak
african-languages
instruction-tuning
conversational-ai
harmful-content
License:
Dataset card Files
xet
 Community
wildjailbreak-africa / README.md
keithtwesigye's picture
keithtwesigye
Bronsn's picture
Bronsn
Duplicate from CraneAILabs/wildjailbreak-africa
ec5e680
preview code
|
Raw History Blame Contribute Delete
11.7 kB
metadata 
license: odc-by
task_categories:
  - text-generation
language:
  - en
  - ach
  - lgg
  - lug
  - sw
  - teo
multilinguality:
  - multilingual
size_categories:
  - 100K<n<1M
source_datasets:
  - allenai/wildjailbreak
tags:
  - safety
  - jailbreak
  - african-languages
  - instruction-tuning
  - conversational-ai
  - harmful-content
  - safety-training
pretty_name: WildJailbreak Africa
dataset_info:
  - config_name: en
    features:
      - name: id
        dtype: string
      - name: messages
        sequence:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: source
        dtype: string
    splits:
      - name: train
        num_bytes: 258940000
        num_examples: 50000
    download_size: 258940000
    dataset_size: 258940000
  - config_name: ach
    features:
      - name: id
        dtype: string
      - name: messages
        sequence:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: source
        dtype: string
    splits:
      - name: train
        num_bytes: 272081318
        num_examples: 49819
    download_size: 272081318
    dataset_size: 272081318
  - config_name: lgg
    features:
      - name: id
        dtype: string
      - name: messages
        sequence:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: source
        dtype: string
    splits:
      - name: train
        num_bytes: 290939474
        num_examples: 49854
    download_size: 290939474
    dataset_size: 290939474
  - config_name: lug
    features:
      - name: id
        dtype: string
      - name: messages
        sequence:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: source
        dtype: string
    splits:
      - name: train
        num_bytes: 303608634
        num_examples: 49864
    download_size: 303608634
    dataset_size: 303608634
  - config_name: swa
    features:
      - name: id
        dtype: string
      - name: messages
        sequence:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: source
        dtype: string
    splits:
      - name: train
        num_bytes: 288676344
        num_examples: 49875
    download_size: 288676344
    dataset_size: 288676344
  - config_name: teo
    features:
      - name: id
        dtype: string
      - name: messages
        sequence:
          - name: role
            dtype: string
          - name: content
            dtype: string
      - name: source
        dtype: string
    splits:
      - name: train
        num_bytes: 292089554
        num_examples: 49871
    download_size: 292089554
    dataset_size: 292089554
configs:
  - config_name: en
    data_files:
      - split: train
        path: data/en/train.jsonl
  - config_name: ach
    data_files:
      - split: train
        path: data/ach/train.jsonl
  - config_name: lgg
    data_files:
      - split: train
        path: data/lgg/train.jsonl
  - config_name: lug
    data_files:
      - split: train
        path: data/lug/train.jsonl
  - config_name: swa
    data_files:
      - split: train
        path: data/swa/train.jsonl
  - config_name: teo
    data_files:
      - split: train
        path: data/teo/train.jsonl

WildJailbreak Africa

This dataset contains translations of 50,000 samples from the ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k dataset into 5 African languages. The dataset is designed for instruction tuning and safety training of language models in low-resource African languages.

Dataset Description

The original WildJailbreak dataset is a synthetic safety-training dataset containing both vanilla (direct harmful requests) and adversarial (complex adversarial jailbreaks) prompt-response pairs. This translated version maintains the conversational structure while adapting the content to African languages.

Languages Included

Language Code Language Name Samples Region
en English 50,000 Original dataset
ach Acholi 49,819 Northern Uganda
lgg Lugbara 49,854 Northwestern Uganda/South Sudan
lug Luganda 49,864 Central Uganda
swa Swahili 49,875 East/Central Africa
teo Ateso 49,871 Eastern Uganda

Total samples: ~299,283

Dataset Structure 

├── en/
│   └── train.jsonl          # Original English dataset
├── ach/
│   └── train.jsonl          # Acholi translations
├── lgg/
│   └── train.jsonl          # Lugbara translations  
├── lug/
│   └── train.jsonl          # Luganda translations
├── swa/
│   └── train.jsonl          # Swahili translations
└── teo/
    └── train.jsonl          # Ateso translations

Data Format

Each file contains JSONL format with the following structure:

{
    "id": "unique_identifier",
    "messages": [
        {
            "role": "user",
            "content": "User message in target language"
        },
        {
            "role": "assistant", 
            "content": "Assistant response in target language"
        }
    ],
    "source": "ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k"
}

Usage

Loading with Datasets Library 

from datasets import load_dataset

# Load specific language
dataset = load_dataset("CraneAILabs/wildjailbreak-africa", "swa")

# Load all languages
all_languages = load_dataset("CraneAILabs/wildjailbreak-africa")

Manual Loading 

import json

def load_jsonl(file_path):
    with open(file_path, 'r', encoding='utf-8') as f:
        return [json.loads(line) for line in f]

# Load Swahili samples
swahili_data = load_jsonl("swa/train.jsonl")

Original Dataset Information

  • Source: ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k
  • Original Dataset: WildJailbreak (262K samples)
  • Subset Used: Decontaminated 50K samples from Tulu 3 SFT Mixture
  • Purpose: Safety training and jailbreak resistance
  • License: ODC-BY-1.0

The WildJailbreak dataset was designed to mitigate exaggerated safety behaviors by providing:

  1. Harmful queries (both vanilla and adversarial)
  2. Benign queries that resemble harmful queries but contain no harmful intent

Translation Process

The translations were performed using advanced AI translation systems by Crane AI Labs with the following considerations:

  • Maintenance of conversational context and flow
  • Cultural adaptation where appropriate
  • Preservation of safety-related content structure
  • Quality control through automated consistency checks
  • Native speaker review and validation

Intended Use Cases

Primary Use Cases

  • Safety Training: Training African language models to handle harmful content appropriately
  • Instruction Tuning: Fine-tuning conversational AI models for African languages
  • Jailbreak Research: Studying adversarial prompt behavior in low-resource languages
  • Cross-lingual Safety: Understanding safety patterns across linguistic boundaries

Research Applications

  • Multilingual safety alignment research
  • Low-resource language model development
  • Cultural adaptation of AI safety measures
  • Comparative analysis of harmful content across languages

Limitations and Considerations

Translation Limitations

  • Automated Translation: Content was translated using AI systems, which may introduce:
    • Semantic drift from original meaning
    • Cultural context misalignment
    • Grammatical inconsistencies
    • Loss of nuanced safety-related expressions

Cultural Considerations

  • Western-Centric Content: Original dataset reflects Western cultural contexts that may not align with African cultural norms
  • Harmful Content Relevance: Some harmful scenarios may not be culturally relevant or may carry different implications in African contexts
  • Social Norms: Safety boundaries and social taboos vary significantly across cultures

Technical Limitations

  • Data Quality: Translation quality varies across languages and complexity of content
  • Consistency: Terminology and style may not be consistent within or across languages
  • Coverage: Some nuanced safety concepts may not translate effectively

Ethical Considerations

  • Harmful Content: Dataset contains translated harmful prompts that could be misused
  • Cultural Sensitivity: Some content may be inappropriate or offensive in local cultural contexts
  • Representation: May not adequately represent diverse dialects and regional variations

Safety Warnings

⚠️ Content Warning: This dataset contains potentially harmful, offensive, or inappropriate content translated into African languages. Users should:

  • Implement appropriate safeguards when using this data
  • Consider cultural context when applying safety measures
  • Use only for legitimate research and safety training purposes
  • Avoid deployment without proper safety evaluations

Data Quality Limitations

  • Translation Artifacts: May contain artifacts from automated translation processes
  • Inconsistent Quality: Quality varies significantly between simple and complex prompts
  • Missing Context: Some cultural or contextual nuances may be lost in translation

Evaluation and Validation

Recommended Validation Steps

  1. Human Review: Conduct human evaluation of translated content for cultural appropriateness
  2. Safety Testing: Evaluate model outputs for culturally appropriate safety responses
  3. Quality Assessment: Assess translation quality using native speaker evaluation
  4. Cultural Adaptation: Validate that safety measures align with local cultural norms

Metrics to Consider

  • Translation quality (BLEU, chrF++, human evaluation)
  • Cultural appropriateness scores
  • Safety response effectiveness
  • Cross-lingual consistency

Responsible Use Guidelines

Do's

✅ Use for legitimate AI safety research
✅ Implement proper content filtering and safeguards
✅ Conduct cultural sensitivity reviews
✅ Validate with native speakers before deployment
✅ Credit original dataset creators and translators

Don'ts

❌ Deploy without proper safety evaluation
❌ Use for generating harmful content
❌ Ignore cultural context and local norms
❌ Assume uniform quality across all samples
❌ Use as the sole source for production systems

Citation

If you use this dataset, please cite both this work and the original WildJailbreak dataset:

@dataset{wildjailbreak_africa,
  title={WildJailbreak Africa: Tulu 3.9 WildJailbreak African Languages Dataset},
  author={Crane AI Labs},
  year={2025},
  url={https://huggingface.co/datasets/CraneAILabs/wildjailbreak-africa}
}

@dataset{wildjailbreak,
  title={WildJailbreak: An Open-source Large-scale Synthetic Jailbreak Dataset},
  author={Shen, Liwei and Tao, Zhihong and Cheng, Pengfei and others},
  year={2024},
  url={https://huggingface.co/datasets/allenai/wildjailbreak}
}

Contributing

We welcome contributions to improve translation quality, add more languages, or enhance cultural appropriateness. Please:

  1. Ensure cultural sensitivity in all contributions
  2. Provide proper documentation for changes
  3. Include validation metrics for improvements

License

This dataset follows the licensing terms of the original WildJailbreak dataset (ODC-BY-1.0). Please review the original dataset license before use.

Contact

For questions, concerns, or contributions, please contact:

Dataset Statistics:

  • Total Conversations: 299,283
  • Languages: 6 (English + 5 African languages)
  • Format: JSONL (JSON Lines)
  • Size: ~600MB (estimated)
  • Split: Train only

Contributors: Crane AI Labs

Last Updated: August 2025