apart-global-south-hack/multilingual-counterfactual

Multilingual Counterfactual

A multilingual counterfactual MCQ dataset built from COCO-Counterfactual.

Each row contains an image, two captions (original vs counterfactual), and a multiple-choice question probing the difference between image content and text description.

Languages

Language	Code	Status
English	en	✅ Done
Hindi	hi	✅ Done
Urdu	ur	✅ Done
Telugu	te	✅ Done
Bahasa Indonesia	id	✅ Done

Columns

Column	Type	Description
serial_id	int	Unique row identifier (1–150 per language)
image	image	Original COCO image (512×512)
original_caption	str	Caption describing the original image
counterfactual_caption	str	Caption describing the counterfactual (edited) image
changed_words	dict	The word(s) changed between original and counterfactual
question	str	Multiple-choice question probing the difference
image_answer_bias	str	Correct answer based on image evidence
text_answer_bias	str	Correct answer based on text evidence
plausible_distractor	str	Plausible but incorrect answer
language	str	Language of the row

Dataset Statistics

• Total rows: 750 (150 per language × 5 languages)
• Source: 150 unique COCO-Counterfactual examples
• Image dimensions: 512×512 RGB

Usage

from datasets import load_dataset

ds = load_dataset("feliren/multilingual-counterfactual", split="train")
print(ds[0])

Original Source

This dataset is built from COCO-Counterfactuals by Le et al. (2023).

Tiep Le, Vasudev Lal, and Phillip Howard. 2023. COCO-Counterfactuals: Automatically Constructed Counterfactual Examples for Image-Text Pairs. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track.

@inproceedings{le2023coco,
  title={COCO-Counterfactuals: Automatically Constructed Counterfactual Examples for Image-Text Pairs},
  author={Le, Tiep and Lal, Vasudev and Howard, Phillip},
  booktitle={Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track},
  year={2023}
}