AddressHate/AH-Qwen3-VL-8B-Antisemitism
Image-Text-to-Text • Updated • 2
Request access to the AddressHate dataset
Provide your full name, organization, and email, and accept the AddressHate Model and Dataset License v1.0.
Your agreement is with AddressHate, Inc. Noncommercial use only; redistribution of the original model or dataset is prohibited. Read the AddressHate Model and Dataset License v1.0: https://huggingface.co/spaces/AddressHate/README/blob/be3d796893893a53fba3b39fef8148e4666e52d7/LICENSE.md
Log in or Sign Up to review the conditions and access this dataset content.
AH Antisemitism Multimodal
AH Antisemitism Multimodal is an AddressHate dataset for studying and classifying antisemitism in online content. It contains 5,877 human-annotated posts, combining focal text, captured conversational context, associated images or video where available, and multilabel annotations.
All data in this release was annotated between August 12 and October 17, 2026, using the AddressHate Annotator.
Version: V2
Format: Parquet
Modalities: Text, image, video
Dataset overview
Each row represents one post: a social media post or comment, a web article, or an AI-chat response. The dataset is meant for classification and guardrail evaluation, containing labels on if something is antisemitic with a detailed multilabel class of which antisemitic concepts a specific post has. The dataset also includes necessary contextual parent and sibling posts in threads and conversations.
The release contains data from 16 platform/source categories:
| Source | Posts |
|---|---|
| TikTok | 1,419 |
| X | 1,213 |
| YouTube | 741 |
| 727 | |
| 4chan | 605 |
| 415 | |
| Web articles | 315 |
| Rumble | 223 |
| Bluesky | 199 |
| 5 | |
| 8chan | 4 |
| ChatGPT | 3 |
| DeepSeek | 3 |
| Claude | 2 |
| Mastodon | 2 |
| Gemini | 1 |
“Web articles” includes articles from multiple websites.
Media availability
| Associated media | Posts |
|---|---|
| Text only | 5,478 |
| Images, without video | 295 |
| Video, without images | 101 |
| Both images and video | 3 |
Images are stored as PNGs and videos as MP4s. GIFs are represented as videos. Associated media can belong to captured context rather than to the focal post itself.
Quickstart
Accept the dataset’s access terms on Hugging Face, then install the dependencies and authenticate:
pip install "datasets==4.8.5" "Pillow==12.2.0" ipython
hf auth login
Load and display a post with associated images, video, and conversational context in a notebook:
from datasets import load_dataset
from IPython.display import display, Video
dataset = load_dataset(
"AddressHate/AH-Antisemitism-Multimodal",
revision="v2",
token=True,
)
example = next(
split[i]
for split in dataset.values()
for i, modality in enumerate(split["associated_modality"])
if modality == "image+video"
)
print("Post:\n", example["text"])
print("\nContext:\n", example["context_text"])
for image in example["images"]:
display(image)
display(Video(
data=example["video"]["bytes"],
embed=True,
mimetype="video/mp4",
))
Images and video are embedded in the dataset. Associated media may belong to the focal post or its captured context.
Labels and annotation
The dataset provides three annotation fields:
| Field | Meaning |
|---|---|
ideation_codes |
Whether and how the post expresses antisemitic meaning |
leaf_codes |
Specific concepts selected from the antisemitism taxonomy |
ideology_labels |
Supplementary ideological or hate-related selections |
Ideation labels
| Code | Meaning |
|---|---|
I0 — Not antisemitic |
No antisemitic meaning is identified |
I0c — Counter-speech |
Actively challenges or refutes antisemitic claims |
I1 — Antisemitic |
Antisemitic meaning is apparent from the focal content itself |
I1a — Co-textual antisemitism |
Interpretation requires the immediate linguistic context, such as the parent post or surrounding replies |
I1a+ — Confirmation |
Affirms or endorses an antisemitic statement without adding further substantive content |
I1b — Contextual antisemitism |
Interpretation requires external knowledge, such as historical references, tropes, or coded language |
I2 — Unclear ideation |
Content cannot be reliably interpreted, including some language, technical, or accessibility barriers |
Counter-speech requires active rejection of antisemitism; simply mocking or insulting the person making an antisemitic statement does not necessarily qualify.
Conceptual labels
The release contains 40 observed conceptual labels across six categories:
| Category | Focus |
|---|---|
| C1 | Classic Antisemitism |
| C2 | Concepts of Power |
| C3 | Secondary Antisemitism |
| C4 | Further Post-Holocaust Categories |
| C5 | Israel-Related Antisemitism |
| C6 | Aggressive Speech Acts |
See the AddressHate taxonomy and definitions of all 40 leaf concepts for the conceptual label reference. The page also links to the underlying Decoding Antisemitism lexicon.
Labels are stored as lists. Preserve their multilabel structure rather than assuming one class per post. An empty list means no recorded selection; it does not establish that every possible label was independently judged negative. Explicit I0 is distinct from an empty selection.
Dataset structure
The dataset contains 31 typed columns, with the same schema across all splits.
Splits
| Split | Posts | Share |
|---|---|---|
| Train | 4,022 | 68.44% |
| Validation | 904 | 15.38% |
| Test | 951 | 16.18% |
| Total | 5,877 | 100% |
Use training data for fitting and prompt development, validation for model and threshold selection, and test for final evaluation. Preserve the supplied splits when comparing results. These splits are not chronological.
Main fields
| Fields | Description |
|---|---|
post_id |
Stable AddressHate post identifier |
platform, platform_post_id, permalink |
Source platform, source identifier, and link |
text |
Focal post text |
context_text |
Captured context, with role and capture boundaries preserved |
author_id, author_handle |
Nullable captured author identifiers |
like_count, reply_count, repost_count, quote_count, bookmark_count, view_count |
Nullable captured engagement counts |
project |
One associated annotation project, where recorded |
posted_at |
Nullable source publication timestamp |
annotated_at |
Earliest annotation timestamp available in retained history |
ideation_codes, leaf_codes, ideology_labels |
Post-level annotation targets |
annotation_count |
Number of current annotation captures |
annotator_count |
Number of distinct annotators |
annotator_ids |
Stable pseudonymous annotator identifiers |
text_character_count, text_word_count |
Focal text length in characters and whitespace-separated words |
ancestor_count |
Maximum recorded ancestor count within a capture |
associated_modality |
text_only, image, video, or image+video |
images |
List of native Hugging Face images with embedded bytes |
video |
Nullable native Hugging Face Video(decode=False) with embedded bytes |
media_links |
Attachment references containing media kind, feature index, and capture roles |
Images load as Pillow images. Video decoding is disabled by default: video is either None or a dictionary containing embedded bytes and an informational filename. Media loading does not require external URLs or access to the original archive.
Timestamps are UTC. annotated_at records annotation time, not content publication time. Missing engagement counts are null, distinct from stored zeroes, and engagement meanings vary across platforms.
project records one associated annotation project where available. It does not establish the post’s original collection source.
Collection and quality control
The AddressHate annotation team annotated this release from August 12 through October 17, 2026, using the AddressHate Annotator. The tool captures posts alongside their available context and media.
Collection included event-focused projects and targeted searches for less frequently observed concepts. This affects the distribution of platforms, topics, and labels.
Disagreements are reviewed collaboratively, usually during team meetings. Annotators compare interpretations and may revise their selections after discussion, with additional team members contributing when needed.
For inclusion in this release, multiple current captures of the same post must have matching normalized labels and annotation schemas. Posts with unresolved disagreement are excluded rather than assigned labels through majority voting. Posts with unresolved required archived media are also excluded.
Single-annotator posts remain eligible:
| Distinct annotators | Posts |
|---|---|
| One | 2,813 |
| Two | 3,060 |
| Three | 4 |
Overall, 52.1% of posts have at least two annotators. We suggest for social science applications that only posts with at least two annotators are used, but release all data to give researchers as much as possible.
Limitations and sensitive information
The dataset contains antisemitic and other hateful material.
It is a curated collection rather than a representative sample of online speech. Platform coverage is uneven, and event selection, targeted searches, annotation agreement, and media availability affect which posts are included.
Some labels have few evaluation examples. All 40 observed conceptual labels occur in training and validation; 39 occur in test. AS-C6-2 has no test examples. Report per-label support alongside evaluation results.
Captured context may be incomplete. The presence of context does not itself establish a contextual-antisemitism label. Associated images and video should not automatically be attributed to the focal author.
Annotator identifiers are pseudonymized. Source authors, handles, text, and media are not anonymized.
License and attribution
The dataset is governed by the AddressHate Model and Dataset License v1.0, issued by AddressHate, Inc. The complete license controls permitted use, redistribution, and derivatives. Third-party materials retain their applicable rights and terms.
Dataset attribution: Adam Lehavi, AddressHate. Annotations were produced by the AddressHate annotation team.
Include the following attribution in publications or documentation describing substantial use:
Uses resources from AddressHate, Inc.: https://huggingface.co/datasets/AddressHate/AH-Antisemitism-Multimodal.
Citation
@misc{lehavi2026ahantisemitismv2,
author = {Lehavi, Adam},
title = {{AH Antisemitism Multimodal}},
year = {2026},
version = {2},
publisher = {AddressHate, Inc.},
howpublished = {Hugging Face},
url = {https://huggingface.co/datasets/AddressHate/AH-Antisemitism-Multimodal/tree/v2},
note = {Dataset}
}
- Downloads last month
- 74
Models trained or fine-tuned on AddressHate/AH-Antisemitism-Multimodal