AHA-Memes

A Fine-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes

Hateful memes carry their meaning in the interaction between an image and the text laid over it, and often through cultural references that neither modality states outright. Arabic has been badly served here: the meme resources that exist annotate propaganda or coarse "harmful content", not who is being attacked or how.

AHA-Memes is a benchmark of 5,000 Arabic memes, each annotated by trained native speakers for binary hatefulness and, where a meme is hateful, for the attack strategy it uses. A further 66,413 memes ship with labels and rich descriptive metadata generated by Gemini 3.1 Pro, for weakly supervised and semi-supervised work.

⚠️ Content warning. This dataset contains hateful, offensive and disturbing material, including slurs and dehumanising imagery targeting protected groups. It is released to enable research on detection and moderation, not to endorse any of it.

• Paper: arXiv:2607.27393
• Code, baselines and reproduction recipes: github.com/MohamedBayan/AHA-Memes
• Licence: CC BY-NC 4.0 (non-commercial research)

Loading it

from datasets import load_dataset

test = load_dataset("QCRI/AHA-MEMES", split="test")
print(test[0]["label"], test[0]["fine_grained_label"])

# The silver corpus is ~5 GB; stream it rather than downloading it whole.
silver = load_dataset("QCRI/AHA-MEMES", split="silver", streaming=True)

If you want files on disk instead — which is what the image and fusion baselines in the code repository expect — use the downloader there:

python -m aha.download                  # train + dev + test
python -m aha.download --splits silver

Splits

Split	Memes	Hateful	Labels	Size
train	3,500	1,324 (37.8%)	human	0.27 GB
dev	500	189 (37.8%)	human	0.04 GB
test	1,000	337 (33.7%)	human	0.08 GB
silver	66,413	580 (0.9%)	Gemini 3.1 Pro, not verified	4.95 GB

The 5,000 human-annotated memes are the benchmark. They are stratified by the binary label, no meme appears in more than one split, and silver is disjoint from all three. All results in the paper come from the human splits.

Fields

Field	Type	Notes
id	string	original filename, e.g. F9yEY9jXIAAJUTg.jpg
image	image	the meme
text	string	text overlaid on the meme, extracted with EasyOCR
label	string	Hateful / Not Hateful; null for 116 silver rows
fine_grained_label	list[string]	subset of the ten categories below
annotation_source	string	human, or gemini-3.1-pro-preview on silver
meta	struct	21 descriptive attributes, all model-generated

Check annotation_source before treating a label as ground truth. It is the only thing distinguishing the human benchmark from the silver corpus once the splits are concatenated.

The taxonomy

Binary hatefulness sits at the top. A meme is Hateful if it attacks people, directly or indirectly, on the basis of a protected characteristic. Two boundaries are deliberate: attacks on groups that themselves perpetrate hate are not counted, and content that is merely rude or offensive without targeting a protected category is Not Hateful.

The ten fine-grained categories are multi-label and conditional on that decision:

Hateful — attack strategy	Not hateful — pragmatic function
Mocking, Incitement, Dehumanization, Slurs, Contempt, Inferiority, Exclusion	Humor, Sarcasm
Other	Other

Other is reachable from either side, which is why the paper's split table lists it twice. Full definitions and the bilingual guidelines the annotators worked from are in docs/annotation-guidelines.md.

Label distribution

Hate	Fine-grained	Train	Dev	Test	Total
Hateful	Mocking	706	90	211	1,007
Hateful	Incitement	320	51	85	456
Hateful	Dehumanization	247	42	58	347
Hateful	Slurs	252	42	47	341
Hateful	Contempt	107	18	50	175
Hateful	Inferiority	57	14	32	103
Hateful	Exclusion	10	4	3	17
Hateful	Other	18	2	7	27
Not Hateful	Other	380	50	96	526
Not Hateful	Sarcasm	934	126	333	1,393
Not Hateful	Humor	863	136	332	1,331
	Memes	3,500	500	1,000	5,000

Counts exceed the number of memes because the label is multi-label: 9.9% of train and 21.4% of test carry more than one category. The taxonomy is long-tailed — Exclusion has three test instances — which is why macro-F1 reads far below micro-F1 for every system.

The meta struct

Generated by Gemini 3.1 Pro for every row. On the human splits it was produced conditioned on the human label, so the model described the meme but never relabelled it; on silver the same call produced both label and description.

topic, mentioned_categories, dialect, visual_manipulation, ocr_text, ocr_english_translation, intent, stance, sentiment, emotion, text_image_relationship, meaning_type, cultural_references, context_scope, requires_current_event_knowledge, propaganda, propaganda_techniques, english_rationale, arabic_rationale, confidence, subtype_raw.

Two are easy to misread. meta.ocr_text is not the same as the top-level text: the former is what Gemini read off the image, the latter is EasyOCR's output, and they disagree often enough to be useful if you care about OCR quality. And mentioned_categories is a model guess, not a human target annotation — the human target labels are not part of this release.

Known quirks in the silver split

• 116 rows have no label. For 115 the model's reply could not be parsed; one parsed but omitted the label. They still carry image and OCR text, so they are usable as unlabelled data. Filter with ds.filter(lambda x: x["label"] is not None).
• Three rows got an out-of-taxonomy category (Criticism ×2, Satire ×1). fine_grained_label is left empty for those so the label space stays closed; the literal reply is preserved in meta.subtype_raw.
• Silver fine-grained labels are single-label. The prompt asked for one subtype, so unlike the human splits these never have more than one element. Do not mix the two when computing multi-label statistics.
• The class balance differs sharply. 0.9% of silver is hateful against 37% of the human splits. The human set was built by pre-selecting candidates with Gemma-3-12B to raise the positive rate; silver covers the broad pool.
• It skews Egyptian. 56,697 of the 66,297 labelled silver rows are tagged egyptian dialect and daily_life is the dominant topic. Large, but not evenly spread across the Arabic-speaking world.

How it was built

Memes were collected from public pages and groups on Facebook, Instagram, Pinterest and Twitter/X, focused on public figures, politics and social commentary. Exact and near-duplicate images were removed using embeddings from a model fine-tuned on social-media imagery, treating pairs within Euclidean distance 3.6 as duplicates. Overlaid text was extracted with EasyOCR; memes with no detectable text were dropped, so every meme here has both modalities.

Because hateful content is rare in the wild, Gemma-3-12B assigned provisional binary labels to 71K memes and the 5,000-meme annotation set was sampled from that pool. Those provisional labels were hidden from annotators and discarded — all gold labels are human.

Annotation was done by a third-party company: three trained native Arabic speakers working from bilingual guidelines, after several rounds of training and guideline refinement, at a cost of roughly $4K. Agreement (Cohen's κ, macro-averaged over subtypes and annotator pairs) was 0.91 for binary hatefulness, 0.75 for hate type and 0.67 for the non-hateful subtypes.

Benchmark results

From the paper's Table 2, on the 1,000-meme test split. Regenerate any of these from the committed predictions in the code repository — no GPU needed.

System	Binary macro-F1	Binary Rec(Hate)	Fine-grained macro-F1
Majority baseline	0.399	0.000	0.050
MARBERTv2 (text, fine-tuned)	0.709	0.596	0.263
MARBERTv2+BEiT (late fusion)	0.724	0.656	0.318
Qwen3-VL-8B-Instruct (zero-shot)	0.643	0.318	0.176
Qwen3-VL-8B (LoRA fine-tuned)	0.768	0.656	0.334
Gemini-2.5-pro (zero-shot)	0.711	0.457	0.340
GPT-5 (zero-shot)	0.628	0.282	0.301

Fine-tuning the open 8B VLM wins the binary task. Zero-shot models are badly under-sensitive — InternVL3.5-8B recalls 6% of hateful memes while posting 0.676 accuracy, barely above the majority baseline. And nothing clears 0.35 macro-F1 on the fine-grained task, where the best system is a prompted closed model: 3,500 training memes are not enough for the long tail.

Intended use and limitations

Intended for research on Arabic multimodal hate detection and safer content moderation. Not for commercial use, and not for targeting, profiling or otherwise harming individuals or groups.

The data comes from four public platforms and does not cover every dialect, region or platform community, nor private or ephemeral content. Some annotation boundaries are genuinely hard — offensive humour, satire, political criticism and protected-group hate shade into one another, which is what the detailed guidelines and the calibration rounds were for. The silver corpus broadens the resource but its labels are model-generated and unverified. Models trained here can both miss harmful content and over-flag legitimate speech, so they belong behind human oversight rather than in front of it.

Citation

@article{kmainasi2026aha,
  title={AHA-Memes: A Fine-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes},
  author={Kmainasi, Mohamed Bayan and Shahroor, Ali Ezzat and Hasnat, Abul and Biswas, Md Rafiul and Zaghouani, Wajdi and Alam, Firoj},
  journal={arXiv preprint arXiv:2607.27393},
  year={2026}
}

Acknowledgments

Supported by NPRP grant 14C-0916-210015 from the Qatar National Research Fund, part of the Qatar Research Development and Innovation Council (QRDI). The findings reported here are solely the responsibility of the authors.