Datasets:
Labels or Input? Augmentations of FHM and MAMI
Content warning: this dataset contains hateful, misogynistic and offensive text and images.
This dataset holds two kinds of augmentation for the train and validation splits of the Facebook Hateful Memes (FHM) and MAMI (SemEval-2022 Task 5) datasets. The test splits are not augmented.
- Inputs: new non-hateful counterfactual memes. They are made from hateful memes whose caption is hateful but whose image is not. gpt-5-nano describes the image and rewrites the caption into a non-hateful one, and gemini-3.1-flash-lite-image generates a new meme from the description and the new caption. Every new meme has label 0.
- Labels: a 0-9 hatefulness score from a teacher model (gpt-5-nano) for every original meme. Each meme gets two scores, one from a Simple prompt and one from a Category prompt.
The original memes are not included here. Download FHM and MAMI from the links above.
Structure
FHM/ex/ and MAMI/ex/
βββ img/ex_<original file name> # new memes, in the same format as the original image (FHM: .png, MAMI: .jpg)
βββ ex_train.jsonl # Inputs
βββ ex_val.jsonl
βββ ex_train_scale.jsonl # Labels
βββ ex_val_scale.jsonl
| File | FHM | MAMI | Fields |
|---|---|---|---|
ex_train.jsonl |
1,275 | 820 | id, img, text, label |
ex_val.jsonl |
40 | 98 | id, img, text, label |
ex_train_scale.jsonl |
8,500 | 9,000 | id, img, text, label, simple_scale, simple_scale_text, category_scale, category_scale_text |
ex_val_scale.jsonl |
640 | 1,000 | same as above |
id: the id of the original meme, kept as a string because FHM ids have leading zeros (e.g.,"08291"). Inex_*.jsonlit is the meme the new one was made from.img: inex_*.jsonl, the new image, relative toex/. Inex_*_scale.jsonl, the original image asimg/<original file name>; for MAMI, this includes the split suffix described below.text: the new non-hateful caption inex_*.jsonl, and the original caption inex_*_scale.jsonl.label: the original binary label: hateful for FHM, misogynous for MAMI. It is always 0 for the new memes.simple_scale/category_scale: the teacher's 0-9 score; 5 or more means hateful.simple_scale_text/category_scale_text: the teacher's raw answer.
The scale files keep every meme. For training on the scores, keep only memes whose score agrees with the binary label (0-4 for label 0, 5-9 for label 1).
Splits of the original datasets
- FHM: train =
train(8,500); validation =dev_seenβͺdev_unseen(640). - MAMI: train =
train.tsv(9,000); validation =validation.tsv(1,000). Five file names appear in both the training and test image folders, so they carry a split suffix:_trfor train,_valfor validation and_tfor test (e.g.,15001_val.jpg).
Usage
from datasets import load_dataset
inputs = load_dataset("sahajps/LoI-Meme", "FHM_inputs") # captions and image paths; also FHM_labels, MAMI_inputs, MAMI_labels
# images: download everything, e.g., next to the original datasets in Data/<DATASET>/ori/
from huggingface_hub import snapshot_download
snapshot_download(repo_id="sahajps/LoI-Meme", repo_type="dataset", local_dir="Data")
The original memes and captions remain under the terms of the FHM and MAMI datasets.
π Cite Us
If you find this work helpful and use our dataset or methodology, please cite:
@article{singh2025labels,
title={Labels or Input? Rethinking Augmentation in Multimodal Hate Detection},
author={Singh, Sahajpreet and Jaidka, Kokil and Mukerjee, Subhayan},
journal={arXiv preprint arXiv:2508.11808},
year={2025}
}
- Downloads last month
- 235
Paper for sahajps/LoI-Meme
Paper β’ 2508.11808 β’ Published