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: 5096142
num_examples: 828
- name: train
num_bytes: 510314876
num_examples: 83190
- name: validation
num_bytes: 5089681
num_examples: 834
download_size: 520708749
dataset_size: 520500699
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
- split: train
path: data/train-*
- split: validation
path: data/validation-*
language:
- en
license: mit
source_datasets: []
task_categories:
- multiple-choice
- visual-question-answering
tags:
- decision-model
- grounding
- procedural
shapeworld-grounding
Procedurally verified color/shape and spatial grounding using ShapeWorld's renderer.
Native scene objects provide every candidate box. Queries are retained only when exactly one object satisfies their attributes and, where used, spatial relation. Non-overlapping, fully visible scenes have three to six shapes. Independent split/scene seeds and image hashes prevent train/evaluation image overlap. Images are generated as PNG, stored once per scene with up to two grouped QAs. ShapeWorld's MIT license applies; this is a derived grounding task, not its original caption-agreement benchmark. No language model labels or distractors.
Original data: original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/procedural/shapeworld_grounding.py.
Images are stored once with grouped qa records: test: 1,000 questions / 828 image groups, validation: 1,000 questions / 834 image groups, train: 100,000 questions / 83,190 image groups.
Conversion provenance: source revisions, hashes, configuration and exclusions.