BackdoorVLM data

Processed data for the BackdoorVLM code repository. Copy this folder's clean/, poison/, eval/, and images/ directories into the code repository's data/ directory. dataset_info.json provides the LLaMA-Factory dataset names.

The poisoning datasets contain the selected 1k variants. The distinct 4k and 20k clean training sets are also preserved. The small demo/ package matches the code repository. Supporting trigger resources and saved patches are under images/.

Original source datasets and images may have separate redistribution terms; the uploader should verify these before publication.