CultureMix Dataset Assets

This repository hosts the image assets and metadata that back the official CultureMix benchmark introduced in “World in a Frame: Understanding Culture Mixing as a New Challenge for Vision-Language Models” CultureMix. The dataset probes how LVLMs behave when multiple cultural cues (foods, landmarks, or street scenes) co-exist in a single frame.

Directory Layout

datasets/final/
├── SF/                # Single-food crops rendered on white background
├── MF/                # Two-food compositions (food-only)
├── MFB_landmark/      # Two foods + landmark background
├── MFB_street/        # Two foods + street background
├── SFB_landmark/      # Single food + landmark background
├── SFB_street/        # Single food + street background
└── metadata/          # CSV manifests describing every split
    ├── sf.csv
    ├── mf.csv
    ├── mfb_landmark.csv
    ├── mfb_street.csv
    ├── sfb_landmark.csv
    └── sfb_street.csv

Each filename encodes the food IDs (and background slug for _landmark/_street variants). The CSVs mirror those filenames and append the textual labels needed for evaluation or model training.

Metadata Schema

Every CSV under datasets/final/metadata/ exposes the same column set to keep Hugging Face splits aligned:

Column	Description
image_path	Relative path under datasets/final/
image
food1_id, food1_name, food1_country, food1_continent	Netadata for the left food
food2_id, food2_name, food2_country, food2_continent	etadata for the right food (empty/None for single-food splits)
background_continent, background_country, background_name	Background label (empty when not applicable)

Citation

Please cite the CultureMix paper when using these assets:

@article{kim2025culturemix,
  title={World in a Frame: Understanding Culture Mixing as a New Challenge for Vision-Language Models},
  author={Kim, Eunsu and Park, Junyeong and An, Na Min and Kim, Junseong and Patel, Hitesh Laxmichand and Jin, Jiho and Kruk, Julia and Agarwal, Amit and Panda, Srikant and Ilasariya, Fenal Ashokbhai and Shim, Hyunjung and Oh, Alice},
  year={2025},
  journal={arXiv preprint arXiv:2511.22787}
}

---

For questions about data corrections or additional releases, please reach out via email ([email protected]).