---
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: validation
    num_bytes: 14099553
    num_examples: 189
  - name: test
    num_bytes: 14468951
    num_examples: 181
  - name: train
    num_bytes: 38882149
    num_examples: 550
  download_size: 67444148
  dataset_size: 67450653
configs:
- config_name: default
  data_files:
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
  - split: train
    path: data/train-*
license: other
source_datasets: []
---

# memeqa

MemeQA's native single-answer MCQs, preserving meme-level train/dev/test.

Original annotations: https://github.com/npnkhoi/memeqa. Human interpretations,
automatically constructed distractors and human answer verification (ACL 2025).
Uses none_minus only; native multiselect QAs are omitted, never forced to one
answer. Joins the original SemEval images without decoding/re-encoding them.
Annotation repository: MIT; meme images retain their original rights, which
are not covered by the annotation code license. No synthetic splits/options.

Original data: [original distribution](https://github.com/npnkhoi/memeqa). Repackaged as parquet for [tasksource](https://github.com/sileod/tasksource) by [scripts/repackage_dataset/memes.py](https://github.com/sileod/tasksource/blob/main/scripts/repackage_dataset/memes.py#L28).

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