---
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: train
    num_bytes: 1073898773
    num_examples: 92114
  - name: validation
    num_bytes: 1087431
    num_examples: 100
  download_size: 1074000270
  dataset_size: 1074986204
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
language:
- en
license: apache-2.0
source_datasets: []
task_categories:
- multiple-choice
- visual-question-answering
tags:
- decision-model
- procedural
---

# visgym-mcq

Balanced static VisGym: mazes, equation repair, stick rotation and block legality.

One mirror, with 25,000 training and 25 held-out examples per family at the
default limits. Every choice is verified against its instance. Fresh native
procedural images only; no photographic assets or published benchmark rows.
Metadata records family, seed, geometry and verification. Apache-2.0.

Original data: [original distribution](https://github.com/visgym/visgym). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/procedural/balanced_visgym.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/procedural/balanced_visgym.py#L43).

[Conversion provenance](provenance.json): source revisions, hashes, configuration and exclusions.

[Upstream asset credits and licenses](asset-notices.json) accompany the original rendered images.

The training split contains **100,000 QAs**, 25,000 per family, stored in 92,114 image groups. Validation has 100 QAs, 25 per family. Images shared by several questions are stored once; `load_task("visgym/mcq", vision=True)` exposes the flattened examples. Both splits are shuffled with seed 42.