Multi-CommonCanvas-captioning

A multilingual extension of CommonCanvas / CommonCatalog for image captioning in five languages: English, German, Spanish, French and Italian.

Dataset summary

Each row is a Creative-Commons Flickr photo from the source dataset together with a detailed description of the image generated in the language of the row (qwen_caption). The images and their metadata come from CommonCatalog; the captions in this dataset are new.

How the data was built

The source dataset is common-canvas/commoncatalog-cc-by-sa. For each image, a caption was generated in each target language with the permissively licensed model Qwen3.5-122B-A10B-FP8; the text is therefore generated automatically. Original fields (Flickr metadata, user_caption, blip2_caption, description) are kept as they are in the source dataset and are not translated.

A sample, not the full dataset

This repository contains a random sample of the full dataset, published for exposure and quick experimentation. It is not designed to be statistically representative. For more information, please contact the organisation through its Hugging Face page.

Languages and size

One configuration per language; every config has a single train split.

Config	lang	Rows	Size
en	English	5,000	10.88 GB
de	German	1,875	3.35 GB
es	Spanish	1,875	3.37 GB
fr	French	1,875	3.44 GB
it	Italian	1,875	3.36 GB
Total		12,500	24.4 GB

Data fields

Column	Description
jpg	The image (decoded as a PIL image by datasets).
qwen_caption	Generated description of the image, written in the language given by lang.
task	Always Captioning.
Column	Description
---	---
photoid, uid, unickname, title, usertags, machinetags, description	Flickr metadata of the source photo (inherited from CommonCatalog, not regenerated).
datetaken, dateuploaded, capturedevice, exif	Capture and upload information of the source photo.
longitude, latitude, accuracy, country, continent	Geolocation of the source photo, when available.
pageurl, downloadurl, url, key, serverid, farmid, secret, secretoriginal, ext, status, error_message, sha256	Source URLs and download information of the image.
licensename, licenseurl	License of the individual image.
width, height, original_width, original_height	Image size.
least_dim_range, aspect_ratio_bucket	Resolution and aspect-ratio buckets of the image, inherited from the organisation of the source dataset.
user_caption	Caption written by the Flickr user (called caption in the source dataset); often empty.
blip2_caption	BLIP-2 caption provided by the source dataset.
lang	Language of the row; equal to the configuration name.

Usage

from datasets import load_dataset

ds = load_dataset("VillanovaAI/Multi-CommonCanvas-captioning", "it", split="train", streaming=True)
row = next(iter(ds))
print(row["lang"], row["qwen_caption"])
row["jpg"].show()

License

The license is the one declared by the source dataset, common-canvas/commoncatalog-cc-by-sa (CC BY-SA 4.0). Each row also keeps the license of its own image (licensename, licenseurl) and its source page (pageurl). Please check those terms before reusing or redistributing the images.

Citation

This dataset extends CommonCanvas / CommonCatalog. If you use it, please cite the original work:

@InProceedings{Gokaslan_2024_CVPR,
    author    = {Gokaslan, Aaron and Cooper, A. Feder and Collins, Jasmine and Seguin, Landan and Jacobson, Austin and Patel, Mihir and Frankle, Jonathan and Stephenson, Cory and Kuleshov, Volodymyr},
    title     = {CommonCanvas: Open Diffusion Models Trained on Creative-Commons Images},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {8250-8260}
}

• Paper: arXiv:2310.16825
• Source dataset: common-canvas/commoncatalog-cc-by-sa