Necent/llm-jailbreak-prompt-injection-dataset

---
language:
- en
- ar
- bn
- cs
- de
- es
- fr
- hi
- hu
- id
- it
- ja
- ko
- ms
- nl
- pl
- pt
- ru
- sr
- sv
- sw
- ta
- th
- vi
- zh
- jv
license: mit
size_categories:
- 100K<n<1M
task_categories:
- text-classification
tags:
- safety
- jailbreak
- prompt-injection
- toxicity
- red-teaming
- ai-safety
pretty_name: AI Security Dataset
---

# AI Security Dataset

A unified dataset of **691,331 examples** for AI safety research, combining 15 source datasets across **26 languages**.

## Schema

| Column | Description |
|--------|-------------|
| `prompt` | The attack/test prompt text |
| `response` | Model response (if available) |
| `model_name` | Target model (if specified) |
| `prompt_type` | jailbreak / prompt_injection / obfuscation / linguistic / toxicity / harmful_behavior |
| `category` | Specific sub-category from the source dataset |
| `is_dangerous` | 1 = dangerous/harmful, 0 = safe/benign |
| `source` | Name of the source dataset |
| `language` | ISO 639-1 language code |

## Sources

| Source | Rows | Dangerous | Safe |
|--------|-----:|----------:|-----:|
| PolyglotToxicityPrompts | 412,433 | 209,933 | 202,500 |
| LLMail-Inject | 165,416 | 165,416 | 0 |
| LinguaSafe | 45,422 | 40,167 | 5,255 |
| RedBench | 28,169 | 28,169 | 0 |
| SGToxicGuard | 13,489 | 13,489 | 0 |
| SPML | 16,012 | 12,542 | 3,470 |
| Do-Not-Answer | 3,755 | 3,755 | 0 |
| MultiJail | 3,148 | 3,148 | 0 |
| TensorTrust | 1,346 | 1,346 | 0 |
| RabakBench | 528 | 392 | 136 |
| AdvBench | 520 | 520 | 0 |
| BIPIA | 450 | 250 | 200 |
| HarmBench | 393 | 393 | 0 |
| JBB-Behaviors | 200 | 100 | 100 |
| ArtPrompt | 50 | 50 | 0 |

## Prompt Types

| Type | Count |
|------|------:|
| toxicity | 426,450 |
| prompt_injection | 183,224 |
| linguistic | 48,570 |
| jailbreak | 29,282 |
| harmful_behavior | 3,755 |
| obfuscation | 50 |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Necent/llm-jailbreak-prompt-injection-dataset")
```

## Collection Script

The dataset was collected using the script at: https://github.com/Necentt/ai-security-course

## Disclaimer

This dataset contains harmful, toxic, and offensive content collected for AI safety research purposes only.
It is intended for training safety classifiers and evaluating LLM robustness, not for generating harmful content.