VL-Compositionality-Benchmarks

This repository contains a collection of evaluation benchmarks used to assess the compositional understanding of Vision-Language Models (VLMs), as presented in the paper Cross-Modal Masked Compositional Concept Modeling for Enhancing Visio-Linguistic Compositionality.

Official GitHub Repository: hiker-lw/MACCO

Dataset Summary

These benchmarks are designed to evaluate how well models like CLIP capture object relations, attribute-object bindings, and word order dependencies, moving beyond "bag-of-words" behavior.

The collection includes the following benchmarks:

• ARO (Attributes, Relations, and Order)
• VL-Checklist (including HAKE, SWiG, and VG)
• Sugar-Crepe
• VALSE
• What's Up

Data Preparation

According to the official repository, datasets should be organized under a datasets/ directory.

Notes for VL-Checklist

The VL-Checklist files are split into multiple parts and must be concatenated before extraction. You can use the following commands:

cat hake.tar.gz.part-* > hake.tar.gz
cat swig.tar.gz.part-* > swig.tar.gz
cat vg.tar.gz.part-* > vg.tar.gz

After extraction, the expected structure for VL-Checklist is:

datasets/
└── VL_checklist/
    ├── VL_checklist_datasets/
    │   ├── data/
    │   ├── hake/
    │   ├── swig/
    │   └── vg/
    └── VL_checklist_json_data/

Citation

@misc{li2026crossmodalmaskedcompositionalconcept,
      title={Cross-Modal Masked Compositional Concept Modeling for Enhancing Visio-Linguistic Compositionality}, 
      author={Wei Li and Zhen Huang and Xinmei Tian},
      year={2026},
      eprint={2606.13288},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.13288}, 
}