Dataset · anson-huang
AI fake news detection (text + image)
anson-huang/mirage-news
15,000 headline-image pairs for training and evaluating neural-generated news detectors.
What's inside
The dataset contains 15,000 English image-caption pairs: some are real news, some AI-generated. The data is split into a training set (10,000), validation set (2,500), and five test sets of 500 examples each. Four test sets are gathered from publications and image generators unfamiliar to the model to test generalization capability.
What you can do with it
- Training multimodal classifiers for deepfakes and fake news
- Evaluating detector robustness against unseen image generators
- Tasks
- Image classification, Text classification
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 442
- Likes
- 13
Models trained on it
How to load it
from datasets import load_dataset
ds = load_dataset("anson-huang/mirage-news")