metadata
dataset_info:
features:
- name: image
dtype: image
- name: text
dtype: string
splits:
- name: train
num_bytes: 14980756553.95
num_examples: 99350
- name: test
num_bytes: 3764134638.04
num_examples: 24838
download_size: 18699893085
dataset_size: 18744891191.99
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
Translated and filtered COCO 2017 and Flickr30K created by Sajjad Ayoubi & Navid Kanaani . (Original link)
The original dataset was split by a 80/20 train/test ratio with no further processing.