Datasets:

nrizwan
/

FBHM

---
license: mit
modalities:
- image
- text
task_categories:
- image-classification
- text-classification
language:
- en
tags:
- memes
- hate-speech
- toxicity
- content-moderation
- multimodal
- image-text
- vision-language
- social-media
- nlp
- computer-vision
- functional-testing
- fbhm
- lsv
- evaluation
- classification
pretty_name: FBHM
dataset_info:
  features:
  - name: img_id
    dtype: int64
  - name: img
    dtype: image
  - name: text
    dtype: large_string
  - name: label
    dtype: int64
  - name: functionality
    dtype: large_string
  splits:
  - name: train
    num_bytes: 278281238
    num_examples: 500
  - name: test
    num_bytes: 1806290298
    num_examples: 4500
  download_size: 2878166508
  dataset_size: 2084571536
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---
# FBHM: Functional Benchmarking and Steering of VLMs for Hateful Meme Detection

Accepted at **EMNLP 2026 Main** 🎉

<p align="left">
<b>Authors:</b> Paramananda Bhaskar*, Naquee Rizwan*, Daksh Jogchand, Saurabh Kumar Pandey, Animesh Mukherjee<br>(*) denotes equal contribution
</p>
<p align="center">
  <img src="teaser.png" alt="FBHM Dataset">
</p>
<p align="center">
Left: suite of 5,000 FBHM memes spread across 25 functionalities. Each tile presents the functionality number, its description and the corresponding number of memes in that functionality. Right: examples of constructing ten memes for ten target communities using one base image.
</p>
<div align="center">
[![arXiv](https://img.shields.io/badge/arXiv-Paper-B31B1B)](https://arxiv.org/abs/2605.31349v2)
[![GitHub](https://img.shields.io/badge/GitHub-Code-181717?logo=github)](https://github.com/hate-alert/fbhm)
</div>
------------------------------------------
```markdown
**Content Warning** ⚠️
This dataset contains hateful, offensive, and potentially disturbing multimodal content, including derogatory language and harmful stereotypes targeting protected groups.
The content is provided solely for research purposes. Please use the dataset responsibly and with appropriate care when displaying or sharing examples.
```
------------------------------------------
## Abstract
Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational-confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce FBHM, a systematically curated benchmark of **F**unctionality **B**ased **H**ateful **M**emes constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on FBHM, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose LSV (**l**earnable **s**teering **v**ectors), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting FBHM performance by ~30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance.
------------------------------------------
## Usage
FBHM can be loaded directly using the Hugging Face `datasets` library.
1. Installation
```bash
pip install datasets pillow
```
2. Load the Dataset
```python
from datasets import load_dataset
# Load the FBHM dataset
dataset = load_dataset("nrizwan/FBHM")
# Access the train and test splits
train_data = dataset["train"]
test_data = dataset["test"]
print(dataset)
```
3. Inspect a Sample
```python
# Select a sample from the training split
sample = train_data[0]
print("Image ID:", sample["img_id"])
print("Text:", sample["text"])
print("Label:", sample["label"])
print("Functionality:", sample["functionality"])
```
4. Dataset Fields
| Field           | Description                                     |
| --------------- | ----------------------------------------------- |
| `img_id`        | Unique identifier of the meme                   |
| `img`           | Meme image, automatically loaded as a PIL image |
| `text`          | Text associated with the meme                   |
| `label`         | Ground-truth hateful (1) /non-hateful(0) label  |
| `functionality` | Functionality category associated with the meme |
------------------------------------------
## Please cite our paper
~~~bibtex
@inproceedings{bhaskar-etal-2026-fbhm,
    title = "{FBHM}: Functional Benchmarking and Steering of {VLM}s for Hateful Meme Detection",
    author = "Bhaskar, Paramananda  and
      Rizwan, Naquee  and
      Jogchand, Daksh  and
      Pandey, Saurabh Kumar  and
      Mukherjee, Animesh",
    editor = "Lee, Hung-Yi  and
      Dev, Sunipa  and
      Li, Jessy  and
      Elliott, Desmond",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.emnlp-main.1596/",
    pages = "36273--36297",
    ISBN = "979-8-89176-459-0",
    abstract = "Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational{---}confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce **FBHM**, a systematically curated benchmark of ***F****unctionality* ***B****ased* ***H****ateful* ***M****emes* constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on **FBHM**, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose **LSV** (***l****earnable* ***s****t*eering ***v****ectors*), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting **FBHM** performance by {\textasciitilde}30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance."
}
~~~
~~~bibtex
@misc{bhaskar2026fbhmfunctionalbenchmarkingsteering,
      title={FBHM: Functional Benchmarking and Steering of VLMs for Hateful Meme Detection}, 
      author={Paramananda Bhaskar and Naquee Rizwan and Daksh Jogchand and Saurabh Kumar Pandey and Animesh Mukherjee},
      year={2026},
      eprint={2605.31349},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.31349}, 
}
~~~
------------------------------------------
## Contact
For any questions or issues, please contact: pbhaskar@kgpian.iitkgp.ac.in, nrizwan@kgpian.iitkgp.ac.in

	---
	license: mit
	modalities:
	- image
	- text
	task_categories:
	- image-classification
	- text-classification
	language:
	- en
	tags:
	- memes
	- hate-speech
	- toxicity
	- content-moderation
	- multimodal
	- image-text
	- vision-language
	- social-media
	- nlp
	- computer-vision
	- functional-testing
	- fbhm
	- lsv
	- evaluation
	- classification
	pretty_name: FBHM
	dataset_info:
	features:
	- name: img_id
	dtype: int64
	- name: img
	dtype: image
	- name: text
	dtype: large_string
	- name: label
	dtype: int64
	- name: functionality
	dtype: large_string
	splits:
	- name: train
	num_bytes: 278281238
	num_examples: 500
	- name: test
	num_bytes: 1806290298
	num_examples: 4500
	download_size: 2878166508
	dataset_size: 2084571536
	configs:
	- config_name: default
	data_files:
	- split: train
	path: data/train-*
	- split: test
	path: data/test-*
	---

	# FBHM: Functional Benchmarking and Steering of VLMs for Hateful Meme Detection

	Accepted at EMNLP 2026 Main 🎉

	<p align="left">
	<b>Authors:</b> Paramananda Bhaskar, Naquee Rizwan, Daksh Jogchand, Saurabh Kumar Pandey, Animesh Mukherjee<br>(*) denotes equal contribution
	</p>

	<p align="center">
	<img src="teaser.png" alt="FBHM Dataset">
	</p>

	<p align="center">
	Left: suite of 5,000 FBHM memes spread across 25 functionalities. Each tile presents the functionality number, its description and the corresponding number of memes in that functionality. Right: examples of constructing ten memes for ten target communities using one base image.
	</p>

	<div align="center">

	[![arXiv](https://img.shields.io/badge/arXiv-Paper-B31B1B)](https://arxiv.org/abs/2605.31349v2)
	[![GitHub](https://img.shields.io/badge/GitHub-Code-181717?logo=github)](https://github.com/hate-alert/fbhm)

	</div>

	------------------------------------------
	```markdown
	Content Warning ⚠️

	This dataset contains hateful, offensive, and potentially disturbing multimodal content, including derogatory language and harmful stereotypes targeting protected groups.
	The content is provided solely for research purposes. Please use the dataset responsibly and with appropriate care when displaying or sharing examples.
	```

	------------------------------------------
	## Abstract
	Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational-confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce FBHM, a systematically curated benchmark of Functionality Based Hateful Memes constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on FBHM, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose LSV (learnable steering vectors), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting FBHM performance by ~30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance.

	------------------------------------------
	## Usage

	FBHM can be loaded directly using the Hugging Face `datasets` library.

	1. Installation

	```bash
	pip install datasets pillow
	```

	2. Load the Dataset

	```python
	from datasets import load_dataset

	# Load the FBHM dataset
	dataset = load_dataset("nrizwan/FBHM")

	# Access the train and test splits
	train_data = dataset["train"]
	test_data = dataset["test"]

	print(dataset)
	```

	3. Inspect a Sample

	```python
	# Select a sample from the training split
	sample = train_data[0]

	print("Image ID:", sample["img_id"])
	print("Text:", sample["text"])
	print("Label:", sample["label"])
	print("Functionality:", sample["functionality"])
	```

	4. Dataset Fields

	\| Field \| Description \|
	\| --------------- \| ----------------------------------------------- \|
	\| `img_id` \| Unique identifier of the meme \|
	\| `img` \| Meme image, automatically loaded as a PIL image \|
	\| `text` \| Text associated with the meme \|
	\| `label` \| Ground-truth hateful (1) /non-hateful(0) label \|
	\| `functionality` \| Functionality category associated with the meme \|

	------------------------------------------
	## Please cite our paper

	~~~bibtex
	@inproceedings{bhaskar-etal-2026-fbhm,
	title = "{FBHM}: Functional Benchmarking and Steering of {VLM}s for Hateful Meme Detection",
	author = "Bhaskar, Paramananda and
	Rizwan, Naquee and
	Jogchand, Daksh and
	Pandey, Saurabh Kumar and
	Mukherjee, Animesh",
	editor = "Lee, Hung-Yi and
	Dev, Sunipa and
	Li, Jessy and
	Elliott, Desmond",
	booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
	month = oct,
	year = "2026",
	address = "Budapest, Hungary",
	publisher = "Association for Computational Linguistics",
	url = "https://aclanthology.org/2026.emnlp-main.1596/",
	pages = "36273--36297",
	ISBN = "979-8-89176-459-0",
	abstract = "Hateful meme detection remains a formidable challenge for vision-language models, as existing benchmarks are structurally observational{---}confounding rhetorical hate mechanisms with target community features and preventing causal evaluation of model vulnerabilities. To address this, we introduce FBHM, a systematically curated benchmark of *F*unctionality *B*ased *H*ateful *M*emes constructed along two orthogonal axes: 25 distinct rhetorical functionalities and 10 target communities (5,000 memes total). Benchmarking state-of-the-art VLMs reveals a severe generalization gap: models highly accurate on standard datasets catastrophically drop to near-random performance on FBHM, proving they exploit dataset-specific heuristics rather than robust multimodal reasoning. To efficiently close this gap, we propose LSV (*l*earnable *s*teering *v*ectors), an ultra-low data regime strategy that applies a causal intervention objective on as few as 500 steering samples (50 unique base memes), boosting FBHM performance by {\textasciitilde}30 Macro-F1 points while outperforming in-context learning and PEFT without degrading source-domain performance."
	}
	~~~
	~~~bibtex
	@misc{bhaskar2026fbhmfunctionalbenchmarkingsteering,
	title={FBHM: Functional Benchmarking and Steering of VLMs for Hateful Meme Detection},
	author={Paramananda Bhaskar and Naquee Rizwan and Daksh Jogchand and Saurabh Kumar Pandey and Animesh Mukherjee},
	year={2026},
	eprint={2605.31349},
	archivePrefix={arXiv},
	primaryClass={cs.CL},
	url={https://arxiv.org/abs/2605.31349},
	}
	~~~

	------------------------------------------
	## Contact
	For any questions or issues, please contact: pbhaskar@kgpian.iitkgp.ac.in, nrizwan@kgpian.iitkgp.ac.in