MMM-Fact: A Multimodal, Multi-Domain Fact-Checking Dataset with Multi-Level Retrieval Difficulty
Overview
MMM-Fact is a real-world, multimodal benchmark for fact checking spanning 1995–2025. Each claim is paired with the full fact-check article and cross-modal evidence (text, images, videos, tables) to evaluate multi-step, cross-modal retrieval and reasoning.
Scale & Coverage
| Item | Details |
|---|---|
| Instances | 125,449 fact-checked claims |
| Time range | 1995–2025 |
| Sources | 4 fact-checking sites + 1 news outlet (each sample includes the complete fact-check article and associated evidence) |
Annotations & Tasks
- Veracity labels: True / False / Not Enough Information
- Supported tasks:
- Veracity prediction
- Explainable fact-checking
- Complex evidence aggregation
- Longitudinal analysis
Retrieval-Difficulty Tiers
To reflect verification effort, each sample is tagged with a retrieval difficulty based on the number of retrievable evidence sources:
| Tier | Evidence Sources |
|---|---|
| Basic | 1–5 sources |
| Intermediate | 6–10 sources |
| Advanced | >10 sources |
Citation
If you find this dataset useful, please cite:
@article{xu2025mmm, title={MMM-Fact: A Multimodal, Multi-Domain Fact-Checking Dataset with Multi-Level Retrieval Difficulty}, author={Xu, Wenyan and Xiang, Dawei and Ding, Tianqi and Lu, Weihai}, journal={arXiv preprint arXiv:2510.25120}, year={2025} }