dataset_info:
features:
- name: images
list:
image:
decode: false
- name: image_group_id
dtype: string
- name: qa
list:
- name: inputs
dtype: string
- name: choices_list
list: string
- name: labels
dtype: int64
- name: metadata
dtype: string
splits:
- name: test
num_bytes: 520056715
num_examples: 1065
- name: train
num_bytes: 1160523015
num_examples: 2410
- name: validation
num_bytes: 393670940
num_examples: 793
download_size: 2074393682
dataset_size: 2074250670
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
- split: train
path: data/train-*
- split: validation
path: data/validation-*
license: cc-by-nc-sa-4.0
source_datasets:
- TACPS-liv/Spatial-DISE
tags:
- decision-model
- spatial-reasoning
language:
- en
task_categories:
- multiple-choice
- visual-question-answering
spatial-dise
Spatial-DISE native spatial MC questions, with unchanged composite image bytes.
Includes rotation, combination and hidden-face tasks. Projection is excluded for missing viewing cues; folding for a sampled ambiguous set of answer panels. Original train/validation/test partitions are retained, removing generator-ID families shared with a held-out split or the separate 559-question benchmark. Answer choices name the image's panels; Jev may reorder these references without changing panel identity. No generated distractors. Native procedural or curated wild labels are identified in metadata. CC BY-NC-SA 4.0; embedded Icons8 assets and collected aptitude-test images retain their original third-party terms.
Original data: TACPS-liv/Spatial-DISE. Repackaged as parquet for tasksource by scripts/repackage_dataset/spatial_dise.py.
Conversion provenance: source revisions, hashes, configuration and exclusions.