EnXan/VIPI-evaluation

VIPI Evaluation

GitHub repository · Python package on PyPI

Reviewed image fixtures for VIPI, an open source visual prompt-injection detector.

The initial collection contains 113 detection images with 223 classification cases and 1,000 OCR reference images. Different contexts for the same detection image share one image entry. Instructions depicted in images are evaluation data.

detection_eval/
  cases.json
  images/
ocr_eval/
  catalog.json
  images/

Image paths in each JSON catalog are relative to the catalog's directory. Files retain their original bytes, including image metadata. Detection labels, contexts and payload transcripts were carried over from the jointly reviewed VIPI inventory. The OCR selection retains upstream reference text, quality, scope and word regions; its silver annotations are not a certification that all visible image text is transcribed. Some references cover only target text.

Use the dataset

Download either suite through the Hugging Face CLI:

hf download EnXan/VIPI-evaluation --repo-type dataset --include 'detection_eval/*' --local-dir vipi-evaluation
hf download EnXan/VIPI-evaluation --repo-type dataset --include 'ocr_eval/*' --local-dir vipi-evaluation

For reproducible evaluations, pass --revision <full-commit-sha>. The VIPI code repository pins its dataset revision in research/dataset.json and provides uv run python -m research.download to retrieve that revision into the expected local directories. Evaluation code remains in the VIPI GitHub repository.

Contribute and review

Contributions are welcome as pull requests containing both image files and their corresponding catalog changes. Use the Hub's Files and versions → Add file → Upload files → Open as a pull request, or follow CONTRIBUTING.md for the CLI workflow.

The owner reviews images, source information, transcripts and labels before merging. Community contributors do not need write access. Only the owner or explicitly authorized maintainers merge contributions into main. Normal dataset downloads use main or a chosen released commit; PR revisions are separate and are not part of the accepted dataset.

Pull-request content in this public repository can be publicly visible before review. Do not submit confidential images or real credentials. Images depicting malicious instructions are intentional benchmark fixtures; the instructions must not be executed as part of reviewing or evaluating them.

Sources and rights

This collection combines third-party source images and synthetic fixtures. Source licenses and attribution obligations remain applicable; this repository does not grant a blanket license over third-party images. OCR source repositories, sample identifiers, hashes, provenance and recorded usage rights are preserved in each ocr_eval/catalog.json record. Detection fixture sources and synthetic provenance are described in the detection suite guide and image names.

The OCR source families are TextOCR, Dagger7164 VisualPromptInjection, CyberSecEval3, MM-SafetyBench and VPI-Bench. Detection also includes FigStep, Phantom, GhostCommit, Anamorpher and synthetic fixtures. See the source references before redistribution or reuse; recorded source rights are provenance carried over from the catalog, not a legal determination.