license: cc-by-nc-4.0
task_categories:
- text-generation
language: - en
tags: - safety
- adversarial
- red-teaming
- ai-safety
- multi-turn
- chemical-weapons
- cyber-security
- harm-detection
size_categories: - n<1K
Illicit General Multi-Turn Conversations
Multi-turn adversarial conversations that successfully elicited harmful illicit content from AI models. This sample dataset contains 5 conversations (52 turns) covering chemical weapons, cyber threats, and other safety-critical domains.
Dataset Statistics
| Metric | Value |
|---|---|
| Conversations | 5 |
| Total Turns | 52 |
| Avg Turns/Conv | 10.4 |
| Harm Categories | 3 |
Harm Categories
| Category | Turns | Description |
|---|---|---|
| Chemical Weapons | 25 | Information about chemical weapons and toxic agents |
| Cyber | 11 | Harmful cyber security threats and attack methods |
| General | 16 | Other illicit safety-critical harmful content |
Dataset Structure
Fields
conversation_id(int): Unique conversation identifierharm_category(string): Illicit harm typeturn(int): Turn number (1-indexed)user_prompt(string): User's messagemodel_response(string): AI model's response
Example
{
"conversation_id": 1,
"harm_category": "chemical-weapons",
"turn": 1,
"user_prompt": "[User prompt about chemical threats]",
"model_response": "[Model response with harmful information]"
}
Usage
⚠️ Contains harmful illicit content from successful adversarial attacks.
✅ Intended Use
- AI safety research and evaluation
- Security red team training
- Safety system development
❌ Prohibited Use
- Creating chemical weapons or cyber attacks
- Malicious attack development
- Bypassing safety systems
Full Dataset
This is a sample with 5 conversations. The full dataset contains:
- 581+ conversations
- 3,908+ turns
- Extended harm categories
- Regular updates
Contact [email protected] for full dataset access.
License
CC-BY-NC-4.0 - Research and evaluation only. No commercial use without separate licensing.
Part of the AI Safety Datasets Collection