EpiReal-Bench

EpiReal-Bench is a benchmark for evaluating visual misinformation in image-generation models. It contains 1,000 entity-grounded false claims and 10,000 corresponding images, organized into 10 topics × 10 visual formats × 100 images.

Dataset Composition

Each topic contains 100 false claims. Each claim is rendered in all ten visual formats, yielding 100 images per topic–format pair and 1,000 images per topic.

Topics: Government, Election, Military, Disaster, Crime, Health, Science, Economy, Education, and Environment.

Visual Formats: Newspaper page, Broadcast news, Social platform post, Platform screenshot, Documentary photo, Red-headed official document, Mobile alert, Bulletin notice, Textbook page, and Chat record.

Directory Structure

Images are stored as <topic>/<visual_format>/<claim_id>.png, with lowercase topic names and six-digit claim IDs. For example:

EpiReal-Bench/
├── EpiReal-Bench.json
├── prefixes.json
├── README.md
├── government/
│   ├── broadcast_news/
│   │   ├── 000001.png
│   │   └── ...
│   └── ...
├── election/
└── ...

Text Data and Templates

EpiReal-Bench.json contains the claim-level text records:

Field	Description
id	Claim ID matching the image filename
category	Topic category
real_entity	Entity associated with the claim
false_claim	Fabricated assertion represented in the images
reason	Risk rationale for the claim

The ten images sharing a claim ID correspond to the same text record. For example, government/broadcast_news/000001.png corresponds to record 000001 in the JSON file.

prefixes.json maps the ten visual-format names to their generation prompt prefixes. Combine a selected prefix with a record's false_claim to construct a generation prompt.

Download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YYY-001/EpiReal-Bench",
    repo_type="dataset",
    local_dir="EpiReal-Bench",
)