---
dataset_info:
  features:
  - name: images
    list:
      image:
        decode: false
  - name: image_group_id
    dtype: string
  - name: qa
    list:
    - name: inputs
      dtype: string
    - name: labels
      dtype: int64
    - name: metadata
      dtype: string
    - name: choices_list
      list: string
  splits:
  - name: test
    num_bytes: 256660459
    num_examples: 6186
  - name: train
    num_bytes: 2286302065
    num_examples: 49820
  download_size: 2543457015
  dataset_size: 2542962524
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
  - split: train
    path: data/train-*
license: mit
source_datasets:
- OpenMOSS-Team/GameQA-140K
---

# gameqa

GameQA native multiple-choice questions, with original screenshot bytes.

Full native train/test JSON and image archives are joined once. Open-ended
questions are omitted; no distractors are generated. Options retain source
order, equivalent duplicates collapse, and gold indices are remapped.
Embedded option blocks are removed before Jev permutation. Screenshots are
grouped across QAs and test images take precedence over training copies.
Image groups are shuffled within each native split with seed 0; QA order is retained.
Procedural game labels are used; generated rationales are excluded from inputs.
Original source declares MIT. No validation split is fabricated.

Original data: [OpenMOSS-Team/GameQA-140K](https://huggingface.co/datasets/OpenMOSS-Team/GameQA-140K). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/games.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/games.py#L110).

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