metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: caption
      dtype: string
  splits:
    - name: train
      num_bytes: 4993980142
      num_examples: 30000
  download_size: 4898811398
  dataset_size: 4993980142

Dataset Card for "coco-30-val-2014"

This is 30k randomly sampled image-captioned pairs from the COCO 2014 val split. This is useful for image generation benchmarks (FID, CLIPScore, etc.).

Refer to the gist to know how the dataset was created: https://gist.github.com/sayakpaul/0c4435a1df6eb6193f824f9198cabaa5.