Flickr8k Captions With Splits

This dataset repackages the local Flickr8k image-caption corpus into a Hugging Face imagefolder layout with per-split metadata.csv files.

Structure

• train/, dev/, and test/ each contain image files and a metadata.csv.
• Each row represents one image with five captions.
• The image feature is inferred from file_name in each metadata.csv.

Columns

• image: image file loaded by the Hugging Face Datasets library
• file_name: image filename stored as a string and used by imagefolder to map rows to image files
• split: one of train, dev, test
• caption_0 to caption_4: the five raw Flickr8k captions for the image

License

• CC0: Public Domain

Counts

• Total included images: 8000
• Train images: 6000
• Dev images: 1000
• Test images: 1000
• Excluded caption entries without split or image file: 92

Source Mapping

• Split assignments come from Flickr_8k.trainImages.txt, Flickr_8k.devImages.txt, and Flickr_8k.testImages.txt.
• Captions come from Flickr8k.token.txt, using #0 through #4 as caption_0 through caption_4.

Notes

• The original caption file contains 8,092 image-caption groups.
• Only the 8,000 images with explicit split assignments are included in this Hugging Face-ready dataset.
• Excluded rows are recorded in excluded_rows.csv.
• Missing local image files are recorded in missing_image_rows.csv.

Loading Example

from datasets import load_dataset

dataset = load_dataset("intro/flickr8k")
print(dataset["train"][0])

Citation

If you use this dataset, please cite:

@article{hodosh2013framing,
  title={Framing Image Description as a Ranking Task: Data, Models and Evaluation Metrics},
  author={Hodosh, Micah and Young, Peter and Hockenmaier, Julia},
  journal={Journal of Artificial Intelligence Research},
  volume={47},
  pages={853--899},
  year={2013},
  url={http://www.jair.org/papers/paper3994.html}
}