Datasets:

biecho
/

hateful_memes

license: other
license_name: facebook-hateful-memes-research
license_link: https://ai.meta.com/tools/hatefulmemes/
task_categories:
  - image-classification
  - text-classification
tags:
  - hateful-memes
  - multimodal
  - hate-speech
  - memes
  - clip
size_categories:
  - 10K<n<100K

Facebook Hateful Memes Dataset

Complete version of the Hateful Memes Challenge dataset (Kiela et al., 2020) with all images included.

Dataset Description

Hateful memes combine individually benign images and text to produce hateful content. The hate lives in the interaction between modalities, making this one of the hardest content moderation benchmarks.

The dataset includes confounders: meme pairs that share the same text (or image) but carry opposite labels, forcing models to reason about both modalities.

Splits

Split	Examples	Images	Description
train	8,500	8,500 (100%)	Training set
dev_seen	500	398 (80%)	Dev set — memes share text/images with train
dev_unseen	540	433 (80%)	Dev set — novel content not in train
test_seen	1,000	1,000 (100%)	Test set — memes share text/images with train
test_unseen	2,000	1,593 (80%)	Test set — novel content (primary evaluation)

Note: 535 images (IDs ending in 8 or 9) are unavailable from any public source. The original DrivenData competition page is no longer accessible. Train and test_seen are fully complete.

Columns

id (string): 5-digit meme identifier, matches image filename
image (PIL Image): the meme image
text (string): overlaid text extracted from the meme
label (int): 0 = not hateful, 1 = hateful

Citation

@inproceedings{kiela2020hateful,
  title={The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes},
  author={Kiela, Douwe and Firooz, Hamed and Mober, Aravind and Goswami,
          Vedanuj and Misra, Amanpreet and Testuggine, Davide and Mahajan,
          Sachin},
  booktitle={NeurIPS},
  year={2020}
}

License

This dataset was originally released by Meta AI under a research-only license. It is provided here for academic research purposes. See the original license for terms.