configs:
- config_name: default
data_files:- split: train
path: preview.jsonl
- split: train
language:
- en
- fr
- de
- it
- es
license: other
multilinguality: multilingual
task_categories: - text-generation
- question-answering
tags: - red-teaming
- ai-safety
- llm
- multimodal
- adversarial
- benchmark
- alignment
- multilingual
- human-reviewed
- evaluation
- security
annotations_creators: - expert-generated
source_datasets: - original
size_categories: - 1K<n<10K
🛡️ Multimodal Redteaming (EN, FR, DE, IT, ES)
A high-quality multilingual red teaming dataset designed to evaluate the robustness and safety of Large Language Models (LLMs) against adversarial prompts. The dataset includes both text-only and image-supported conversations with expert-curated annotations for AI safety evaluation, benchmarking, and alignment research.
📖 Overview
This dataset contains multilingual red teaming conversations in English, French, German, Italian, and Spanish, spanning both text-only and image-supported prompts. It is designed to evaluate how AI systems respond to adversarial, harmful, or policy-violating requests under realistic attack scenarios.
Each sample has been professionally curated and reviewed by subject-matter experts to ensure annotation consistency and quality. The dataset includes detailed metadata such as attack category, attack strategy, language, use case, and safety labels for multiple model responses, making it suitable for red teaming, safety benchmarking, vulnerability analysis, alignment research, and LLM evaluation.
📋 Data Summary
| Metric | Value |
|---|---|
| Total Samples | 4,500 |
| Languages | English, French, German, Italian, Spanish |
| Media Types | Text & Image |
| Format | JSON |
| Use Cases | Text Understanding, Image Understanding |
📊 Dataset Statistics
The following tables summarize the distribution of attack categories, strategies, and languages across the dataset.
Attack Category
| Category | % | Count |
|---|---|---|
| Crimes | 16.6% | 751 |
| Weapons and Explosives | 13.6% | 614 |
| Harmful Materials | 12.2% | 554 |
| Prohibited Materials | 10.8% | 488 |
| Cyber Attacks | 9.9% | 449 |
| Self-Harm and Suicide | 9.2% | 416 |
| Fraud and Scams | 8.5% | 384 |
| Violence | 6.7% | 303 |
| Hate | 6.0% | 270 |
| Bias | 3.6% | 165 |
Attack Strategy
| Strategy | % | Count |
|---|---|---|
| Step-by-Step Escalation | 31.3% | 2,387 |
| Direct Prompt | 27.1% | 2,066 |
| Hypothetical Testing | 14.7% | 1,125 |
| Storytelling / Role-play | 7.2% | 548 |
| Impersonation or Character | 6.0% | 458 |
| Character Separation | 3.4% | 257 |
| Use of Obscure Science | 1.9% | 148 |
| Use of Informal Language | 1.9% | 145 |
| Language Blending | 1.7% | 128 |
| False Premises | 1.6% | 124 |
Language
| Language | % | Count |
|---|---|---|
| Spanish | 20.2% | 914 |
| Italian | 20.0% | 907 |
| French | 20.0% | 904 |
| German | 19.9% | 903 |
| English | 19.9% | 902 |
🚀 Commercial Version
This repository provides only a public preview of our enterprise-grade red teaming dataset.
The complete commercial version includes:
- Thousands of additional samples
- More attack categories and attack strategies
- Rich metadata and annotation layers
- Custom dataset creation and annotation services
For licensing, custom datasets, or partnership opportunities, please contact us.
📬 Contact
For commercial licensing or collaboration inquiries:
- 📧 Email: [email protected]
- 🌐 Website: https://company.com