What problem does it solve? Building a training dataset for prompt injection and jailbreak detection is error-prone: paraphrased attack templates leak across train/test splits, benign classes are underrepresented, and obfuscated attacks like leetspeak or Base64 evade naive data collection. This Skill provides a structured methodology for assembling a leak-free, balanced dataset from public Hugging Face sources. ## Core Features & Use Cases - Attack Taxonomy: Defines nine labeled categories covering direct injection, indirect injection, jailbreaks, adversarial evasion, and benign hard negatives. - Group-Aware Splitting: Uses MD5 prefix clustering with scikit-learn GroupShuffleSplit to prevent paraphrase leakage between train, validation, and test sets. - Provenance Manifest: Generates a JSON manifest tracking dataset sources, sample counts, and split ratios for reproducibility. - Adversarial Augmentation: Provides mutation recipes for leetspeak, Base64 encoding, character spacing, and zero-width character injection. - Use Case: A research team training a DeBERTa-based prompt injection classifier uses this guide to merge six Hugging Face datasets, cluster DAN template variants, and produce a 70/15/15 split without inflated test metrics. ## Quick Start Ask the AI to download the listed Hugging Face prompt injection datasets, cluster them by prompt prefix, and generate a group-aware train/validation/test split with a provenance manifest.
