QCRI/Arabic-Hateful-Memes

Arabic Hateful Memes (ArHateMeme) — Public Sample

This repository hosts a 100-example diversity-sampled preview drawn from the training split of the ArHateMeme dataset: 5,000 Arabic memes manually annotated for hatefulness and fine-grained sub-types. The full dataset will be released alongside the associated shared task.

⚠️ This preview is intended for format inspection, tooling validation, and schema alignment only. It is not a benchmark and should not be used for model evaluation.

---

About the full dataset

ArHateMeme is a multimodal (image + Arabic text) meme dataset annotated for hate speech in Arabic. It contains 5,000 memes with a binary hatefulness label and a multi-label set of fine-grained sub-types.

Annotation

• 500 memes are triple-annotated (calibration / gold test set).
• 4,500 memes are single-annotated by trained annotators.
• Binary labels use majority voting on the triple-annotated subset.
• Fine-grained sub-types are the union of sub-types from annotators whose binary label matches the majority label.

Label Taxonomy

Aspect	Values
Binary	Hateful, Not Hateful
Hateful sub-types	Mocking, Incitement, Dehumanization, Slurs, Contempt, Inferiority, Exclusion, Stereotyping, Extremism, Threat, Insults, Historical, Other
Non-hateful sub-types	Humor, Sarcasm, Other

A meme is never assigned both hateful and non-hateful sub-types simultaneously.

Official splits (full dataset)

Split	Records	%	Hateful	Not Hateful
train	3,500	70%	1,324	2,176
dev	500	10%	189	311
test	1,000	20%	337	663
Total	5,000	100%	1,850	3,150

All 500 triple-annotated gold memes are in the test split. Splits are stratified by binary label (seed 42) and there is no meme overlap between splits.

---

About this preview sample

• Source split: train (single-annotated bulk memes)
• Size: 100 memes
• Sampling: stratified to cover every fine-grained sub-type present in the training data and preserve a realistic hateful / non-hateful ratio.
• Images: embedded as bytes via the datasets.Image feature — no external files required.
• Arrow/Parquet: stored as a Hugging Face Dataset (Arrow) and uploaded as parquet shards so the Hub viewer renders images inline.

Sample distribution

Binary	Count
Not Hateful	60
Hateful	40

Fine-grained sub-type	Count
Sarcasm	27
Humor	23
Mocking	19
Incitement	15
Other	10
Contempt	8
Slurs	8
Dehumanization	8
Exclusion	5
Inferiority	5

(Fine-grained counts sum to more than 100 because the label is multi-label.)

---

Record schema

{
    "id": "102396787_870863910087838_...jpg",   # string, unique meme id
    "image": <PIL.Image>,                       # embedded bytes, decoded on load
    "text": "…",                                # OCR-extracted meme text (Arabic)
    "label": "Hateful" | "Not Hateful",         # binary label
    "fine_grained_label": ["Mocking", "Incitement"],  # multi-label sub-types
}

Usage

from datasets import load_dataset

ds = load_dataset("QCRI/Arabic-Hateful-Memes", split="train")
print(ds)

example = ds[0]
example["image"].show()
print(example["text"], example["label"], example["fine_grained_label"])

---

Intended use and limitations

• Intended use: research on Arabic multimodal hate speech detection, including binary classification, fine-grained sub-type classification, and vision-language modeling evaluation.
• Limitations: memes reflect online discourse and contain offensive and harmful content. The preview is not balanced and is too small for training or evaluation. Annotations are partially single-annotator and may contain noise.
• Content warning: this dataset contains text and imagery that is offensive, discriminatory, or otherwise harmful by design. Handle with care.

License

Released under CC BY-NC 4.0 for research use only. Not to be used for commercial purposes or for training systems that generate harmful content.

Citation

A citation will be provided when the full dataset is released. Until then, please cite this repository URL.