neuralchemy/prompt-injection-dataset-categorized

Prompt Injection V2 Categorized (Threat Matrix)

Welcome to Prompt Injection dataset Categorized (formerly Threat Matrix) by NeuralAlchemy!

This dataset is the definitive successor to our highly popular Prompt Injection dataset. While most prompt injection datasets only provide a simple safe vs unsafe binary label, this V2 dataset provides a comprehensive, 6-dimensional taxonomy of prompt injections.

We have demerged the multi-label architecture into clean, individual subsets (configs). This allows researchers and engineers to easily download and train specialized multi-class models focused on specific aspects of AI safety!

🚀 How to Use (Subsets)

You can download specific dimensions of the dataset by specifying the subset name in load_dataset().

from datasets import load_dataset

# 1. Train an Intent Classifier
ds_intent = load_dataset("neuralchemy/prompt-injection-v2-categorized", "intent")

# 2. Train a Technique Classifier
ds_technique = load_dataset("neuralchemy/prompt-injection-v2-categorized", "technique")

# 3. Standard Binary Classification
ds_binary = load_dataset("neuralchemy/prompt-injection-v2-categorized", "binary")

📊 Dataset Dimensions (Subsets)

The dataset contains the following dedicated subsets. Each subset retains the text (the prompt) and its associated labels for that specific dimension.

1. intent

Classifies the attacker's underlying goal.

• Classes: System Extraction, Role Hijack, Tool Abuse, Data Exfiltration, Benign, etc.

2. technique

Classifies the specific psychological or structural technique used to deliver the payload.

• Classes: Base64 encoding, Context Ignoring (AutoDAN), Payload Splitting, Dictionary Attacks, None.

3. severity

A numerical rating of how critical a successful bypass would be.

• Classes: Scale from Low (e.g., generating profanity) to Critical (e.g., executing remote code or leaking API keys).

4. surface

Identifies the injection vector.

• Classes: User Input, Document Processing, System Prompt override, etc.

5. source

Indicates the origin of the prompt.

• Classes: Human-crafted, Synthetic (LLM-generated), etc.

6. binary

The standard malicious vs. benign classification for foundational safety filters.

• Classes: 0 (Safe/Benign) and 1 (Malicious/Injection).

💡 Note on ambiguity

The original data also tracked an ambiguity flag (boolean), indicating whether a prompt is borderline or context-dependent (e.g., "delete all files" is benign if it's an IT assistant, but malicious if it's a customer service bot). This flag is incredibly useful for training models to ask for user clarification instead of outright refusing!

🛠 Why Subsets?

Training a single model to output 6 different dimensions simultaneously (Multi-Task Learning) is difficult and often degrades performance. By organizing the dataset into configs, you can train a highly accurate Mixture of Experts (MoE) where one small model specializes in intent, another in technique, and another in severity.