Dataset · S-Labs
Dataset for prompt injection detection in LLM queries
S-Labs/prompt-injection-dataset
A dataset for binary classification of user prompts into safe and malicious (prompt injection).
What's inside
Includes between 10,000 and 100,000 English text examples divided into two classes: standard legitimate queries (0) and prompt injection attempts (1). The dataset is specifically prepared for training compact encoder models acting as a safety filter before the main language model.
What you can do with it
- Training fast filter classifiers (BERT, RoBERTa, DistilBERT) to protect LLM applications
- Testing security systems and guardrail solutions for robustness against attacks
- Analyzing patterns and formulations of user prompt injections
- Tasks
- Text classification
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 1.4K
- Likes
- 16
Models trained on it
- patronus-studio/wolf-defender-prompt-injectionText classification 10K
- Horizon-Labs/prompt-injection-guard-baseText classification 7.4K
- Horizon-Labs/prompt-injection-guard-smallText classification 5.5K
- patronus-studio/wolf-defender-prompt-injection-smallText classification 3.8K
- mradermacher/Spellguard-1.0-8B-GGUF 934
- Horizon-Labs/prompt-injection-guard-largeText classification 603
- mradermacher/GPT-Distill-Qwen3-8B-Spellguard-SLERP-GGUF 479
- patronus-studio/wolf-defender-prompt-injection-edgeText classification 227
How to load it
from datasets import load_dataset
ds = load_dataset("S-Labs/prompt-injection-dataset")