license: mit
gated: true
extra_gated_heading: Request access to ToxicTags
extra_gated_description: >
This dataset contains potentially toxic and offensive memes. It is released
for non-commercial academic research only.
extra_gated_button_content: Request Access
extra_gated_prompt: >
By requesting access you agree to: - use the dataset only for academic
research - not redistribute the raw data - cite the STEMTOX paper if you use
this dataset
extra_gated_fields:
Full name:
type: text
Affiliation / Institution:
type: text
Country:
type: country
Intended research use:
type: text
I will use this dataset only for non-commercial research:
type: checkbox
I will not redistribute the raw dataset:
type: checkbox
I will cite the STEMTOX paper:
type: checkbox
tags:
- memes
- toxicity
- content-moderation
- multimodal
- image-text
- hate-speech
- vision-language
- social-media
- nlp
- computer-vision
- toxictags
- stemtox
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
dataset_info:
features:
- name: img
dtype: image
- name: title
dtype: string
- name: tags
dtype: string
- name: ocr
dtype: string
- name: binary
dtype: string
- name: finegrained
dtype: string
splits:
- name: train
num_bytes: 462434004
num_examples: 5300
- name: test
num_bytes: 87975021
num_examples: 1000
download_size: 547097409
dataset_size: 550409025
language:
- en
STEMTOX: From Collaborative Tags to Fine-Grained Toxic Meme Detection via Entropy-Guided Multi-Task Learning
🎉 Accepted at Transactions of the Association for Computational Linguistics (TACL 2026).
Figure 1. Overview of the STEMTOX framework.
🔎 TL;DR
TOXICTAGS is a large-scale real-world meme dataset comprising 6,300 manually annotated memes collected from publicly available online sources.
Unlike existing datasets, ToxicTags incorporates collaboratively generated social tags associated with the original posts, providing valuable contextual information that complements the visual and textual content.
The dataset is annotated using a two-stage human annotation pipeline. First, each meme is classified as Toxic or Normal. Toxic memes are then further categorized into one of three fine-grained classes: Hateful, Dangerous, or Offensive, resulting in a four-class taxonomy (Normal, Offensive, Dangerous, Hateful).
The dataset is designed to support research on fine-grained toxic meme detection, multimodal content moderation, and vision-language models.
⚠️ Warning: Contains potentially toxic contents
🧮 Dataset Structure
images/
train.csv
test.csv
📋 Fields
| Column | Description |
|---|---|
img |
Relative path to the meme image |
title |
Original meme title |
tags |
Collaborative tags |
ocr |
OCR-extracted text |
binary |
Binary toxicity label |
finegrained |
Fine-grained toxicity label |
🤗 Loading the Dataset
The dataset can be loaded directly from the Hugging Face Hub using the 🤗 Datasets library.
⚙️ Installation
pip install datasets
📦 Load the Dataset
from datasets import load_dataset
dataset = load_dataset("swainsubhankar/ToxicTags")
The dataset consists of two predefined splits:
print(dataset)
Output:
DatasetDict({
train: Dataset({
features: ['img', 'title', 'tags', 'ocr', 'binary', 'finegrained'],
num_rows: 5300
})
test: Dataset({
features: ['img', 'title', 'tags', 'ocr', 'binary', 'finegrained'],
num_rows: 1000
})
})
📑 Access Individual Splits
train = dataset["train"]
test = dataset["test"]
Citation
@misc{swain2026stemtoxsocialtagsfinegrained,
title={STEMTOX: From Social Tags to Fine-Grained Toxic Meme Detection via Entropy-Guided Multi-Task Learning},
author={Subhankar Swain and Naquee Rizwan and Vishwa Gangadhar S and Nayandeep Deb and Animesh Mukherjee},
year={2026},
eprint={2508.04166},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2508.04166}
}