MOSAIC-1M

MOSAIC-1M: A Large-Scale Multi-Modal Dataset with Automated Semantic Annotations for Visual Content Understanding.

1.1M image–text pairs from Flickr (CC-BY 4.0), 427 semantic categories, with BLIP-2 captions, CLIP (ViT-L/14) image and text embeddings (768-d), face counts, NSFW flags, and Flickr title/description. Released as 12 Parquet shards (~100k rows each), shuffled with fixed seed.

Data fields

Column	Type	Description
id	string	Flickr photo ID
category	string	Human-readable (e.g. "Bird watching")
face_count	string	MTCNN face count
img_caption	string	BLIP-2 caption (≤50 tokens)
url	string	Image URL
info	string	Raw Flickr API JSON
img_clip	list[float]	CLIP image embedding, 768-d, L2-normalised
text_clip	list[float]	CLIP caption embedding, 768-d, L2-normalised
flickr_title	string	From photo.title._content
flickr_description	string	From photo.description._content

Image path derivable as {category}/{id}.jpg (category as folder name if needed).

Splits

12 parquet files: mosaic_1m_part_00 … mosaic_1m_part_11 (~100k rows each; last has 1,054).

Load

from datasets import load_dataset
ds = load_dataset("s-emanuilov/mosaic-1m", split="train")

License

CC-BY 4.0 (images from Flickr; annotations and pipeline as per our release).

Citation

If you use this dataset, please cite:

Emanuilov, S., & Dimov, A. MOSAIC-1M: A Large-Scale Multi-Modal Dataset with Automated Semantic Annotations for Visual Content Understanding. Department of Software Technologies, Faculty of Mathematics and Informatics, Sofia University "St. Kliment Ohridski", Bulgaria.

@inproceedings{emanuilov2026mosaic1m,
  title     = {{MOSAIC-1M: A Large-Scale Multi-Modal Dataset with Automated
                Semantic Annotations for Visual Content Understanding}},
  author    = {Emanuilov, Simeon and Dimov, Aleksandar and Makaveev, Ivan},
  booktitle = {Proceedings of the 27th International Conference on
               Computer Systems and Technologies (CompSysTech'26)},
  year      = {2026},
  address   = {Ruse, Bulgaria},
  url       = {https://huggingface.co/datasets/s-emanuilov/mosaic-1m}
}