Datasets:
dataset_info:
features:
- name: images
list:
image:
decode: false
- name: image_group_id
dtype: string
- name: qa
list:
- name: view
dtype: string
- name: inputs
dtype: string
- name: choices_list
list: string
- name: labels
dtype: int64
- name: metadata
dtype: string
splits:
- name: test
num_bytes: 146964377
num_examples: 977
- name: train
num_bytes: 1329129783
num_examples: 8780
- name: validation
num_bytes: 148820981
num_examples: 984
download_size: 1624905594
dataset_size: 1624915141
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
- split: train
path: data/train-*
- split: validation
path: data/validation-*
language:
- en
license: other
source_datasets:
- nlphuji/flickr30k
task_categories:
- multiple-choice
- visual-question-answering
tags:
- decision-model
- grounding
license_name: flickr30k-original-noncommercial-image-terms
license_link: https://github.com/BryanPlummer/flickr30k_entities#flickr30k-entities-dataset
flickr30k-grounding
Human phrase grounding from Flickr30k Entities, with native image-level splits.
Each question retains the complete caption and marks the queried phrase in text. Only chains with one native target box are included; other annotated image regions supply choices, never invented negatives. Known unrelated captions are excluded. Images are stored once with canonical grouped QAs, unchanged from the image source. The training cap counts questions, not image rows. Original images are provided for noncommercial research/education under Flickr terms; annotation/code access does not grant ownership of those images.
Original data: nlphuji/flickr30k, original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/flickr_grounding.py.
Images are stored once with grouped qa records: train: 100,000 questions / 8,780 image groups, validation: 11,232 questions / 984 image groups, test: 11,560 questions / 977 image groups.
Conversion provenance: source revisions, hashes, configuration and exclusions.