ND400 Derivative Dataset

Description:
ND400 is a derivative evaluation dataset created to study semantic drift in Unified Models (UMs) across multiple image-text generations. It contains 400 samples:

• 200 from DOCCI test set (test_00000–test_00199)
• 200 from NoCaps validation set (first 200 entries)

This dataset does not redistribute original images. It provides a CSV of image IDs, splits, and sources to enable users to fetch the images from the original datasets. A sample script to fetch the original dataset will be provided in the project GitHub repo.

CSV Columns:

source	split	image_id
docci	test	test_00000
nocaps	validation	000123.jpg

Usage:

1. Download the CSV (nd400.csv) from this dataset.
2. Use the provided prepare_dataset.py to organize images and captions for evaluation:

<evaluation_data_root>/
├── text-first/
│ ├── gen-0.csv
├── image-first/
│ └── gen-0/ (images)

License

This dataset combines data from DOCCl (CC BY 4.0) and NoCaps (see original). Please respect the original licenses.

Citation:

@misc{mollah2025telephonegameevaluatingsemantic,
      title={The Telephone Game: Evaluating Semantic Drift in Unified Models}, 
      author={Sabbir Mollah and Rohit Gupta and Sirnam Swetha and Qingyang Liu and Ahnaf Munir and Mubarak Shah},
      year={2025},
      eprint={2509.04438},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.04438}, 
}