This dataset contains the processed data used in an experimental evaluation of prompt-injection defenses for instruction-following language models. Repository: leinha/pi-defense-experiment-dataset

Intended use

Academic and experimental evaluation of prompt-injection defenses, including StruQ-like SFT, SecAlign-like DPO and Instruction-Hierarchy-like SFT scenarios.

Content warning

This dataset may contain… See the full description on the dataset page: https://huggingface.co/datasets/leinha/pi-defense-experiment-dataset.