piyushsinghpasi/mscoco-multilingual-30k

MSCOCO Multilingual Image Caption Dataset (Evaluation Subset)

Description

This dataset is a multilingual extension of the MSCOCO 2014 validation set, consisting of a random sample of 30,000 images paired with captions in 10 languages.

Each image is associated with:

• the original English MSCOCO caption
• machine-translated captions in multiple target languages

The dataset is intended primarily for evaluation and analysis of multilingual and cross-lingual image–text models, including vision–language encoders, retrieval systems, and multimodal representation alignment studies.

---

Dataset Structure

.
├── README.md
└── test/
├── images
│   ├── COCO_val2014_000000054123.jpg
│   ├── COCO_val2014_000000012897.jpg
│   └── ...
└── metadata.csv

• Images: JPEG files from the MSCOCO 2014 validation set
• Metadata: CSV file with one row per image
• Split: test only (evaluation-focused)

---

Metadata Schema

Each row in metadata.csv corresponds to one image.

Columns

Column name	Description
file_name	Relative path to the image file
image_name	Original MSCOCO image filename
caption	Original English MSCOCO caption
French	French caption
Greek	Greek caption
Hebrew	Hebrew caption
Indonesian	Indonesian caption
Korean	Korean caption
Persian	Persian caption
Russian	Russian caption
Spanish	Spanish caption
Hindi	Hindi caption

The file_name column is used by the Hugging Face Hub to automatically infer the Image() feature and render image previews in the dataset viewer.

---

Dataset Statistics

• Total images: 30,000
• Captions per image: 1 (per language)
• Languages: 10 (1 English + 9 non-English)
• Source images: MSCOCO 2014 validation set

---

Language Coverage

Language	ISO 639-1 code	Metadata column
English	en	caption
French	fr	French
Greek	el	Greek
Hebrew	he	Hebrew
Indonesian	id	Indonesian
Korean	ko	Korean
Persian	fa	Persian
Russian	ru	Russian
Spanish	es	Spanish
Hindi	hi	Hindi

---

Intended Use

This dataset is designed for:

• Multilingual image–text retrieval evaluation
• Cross-lingual caption alignment analysis
• Multimodal representation learning research
• Zero-shot and transfer evaluation of vision–language models

It is not intended for training large-scale generative models.

---

Usage

from datasets import load_dataset

ds = load_dataset(
    "piyushsinghpasi/mscoco-multilingual-30k",
    split="test"
)

sample = ds[0]

image = sample["image"]
caption_en = sample["caption"]
caption_fr = sample["French"]
caption_hi = sample["Hindi"]

---

Notes

• All non-English captions are machine-generated.
• The dataset preserves the original MSCOCO image identifiers for traceability.
• Only a random subset of the validation set is included.
• No additional image preprocessing has been applied.

---

License & Attribution

• Images and English captions originate from the MSCOCO dataset.
• This dataset follows the CC BY NC-4.0 to comply with translation models used.
• Users must comply with the original MSCOCO terms.

If you use this dataset in academic work, please cite the original MSCOCO dataset and acknowledge this multilingual extension.

---

Citation

This work is accepted in EACL 2026 Findings.

Arxiv:

@article{pasi2026multilingual,
  title={Multilingual-To-Multimodal (M2M): Unlocking New Languages with Monolingual Text},
  author={Pasi, Piyush Singh},
  journal={arXiv preprint arXiv:2601.10096},
  year={2026}
}