Octavio-Santana/prompt-injection-attack-detection-multilingual

Prompt Injection Attack Detection Multilingual Dataset

📌 Overview

This dataset is a merged and cleaned combination of two publicly available datasets for prompt injection detection:

1. 1.PromptInjectionDataset/Injection-Attack-Detection-Dataset
2. 2.rikka-snow/prompt-injection-multilingual

The goal of this merged dataset is to provide a larger and more diverse benchmark for binary classification of prompt injection attacks.

🎯 Task

Binary classification:

• —0 → Safe prompt
• —1 → Prompt injection attack

📊 Dataset Composition

The dataset was created by:

• —Merging the train split of Injection-Attack-Detection-Dataset
• —Merging both train and test splits of prompt-injection-multilingual
• —Removing duplicates
• —Filtering very short samples
• —Shuffling and performing a stratified train/test split

Final structure:

{
"text": string,
"label": ClassLabel (0 = safe, 1 = injection),
"source": string
}

The source field identifies the original dataset of each example.

⚠️ Disclaimer

This dataset is intended for research and development purposes in LLM security and prompt injection detection. Users should verify licensing terms of the original datasets before commercial use.

🚀 Intended Use

• —Fine-tuning binary classifiers
• —Evaluating LLM guardrails
• —Training safety filters
• —Research in adversarial prompt detection