README.md

metadata

language:
  - zh
  - en
datasets:
  - MMFact/MMFact
tags:
  - mmfact
  - multimodal
  - fake-news-detection

MMFact-CKPT

Model Checkpoints for Multimodal Fake News Detection

English | 简体中文

MMFact-CKPT is the model checkpoint repository for MMFact, supporting reproducible experiments on Weibo-17, Weibo-21^†, and GossipCop.

This repository contains checkpoints for 47 methods, including 24 methods with separate ResNet and ViT variants, covering three datasets and three random seeds for a total of 639 checkpoints. Checkpoints are organized by dataset, method, variant, and seed to support evaluation and comparison using the MMFact toolbox.

Checkpoint Coverage

Method group	Methods	Variants	Checkpoints per method	Total
Methods with a single variant	23	default	9	207
Methods with two visual variants	24	resnet, vit	18	432
Total	47	—	—	639

Each variant has one checkpoint for each combination of the following datasets and random seeds:

Setting	Values
Datasets	weibo17, weibo21, gossipcop
Random seeds	0, 42, 2026
Checkpoint filename	checkpoint.pt

The default variant identifies methods without separate ResNet and ViT versions in this collection.

^† Weibo-21 refers to the reconstructed version provided by MMFact, based on the original paper, with a fixed 8:1:1 train/validation/test split. See the dataset repository for details.

Organization and Usage

All checkpoints follow a consistent directory structure:

<dataset>/<method_slug>/<variant>/seed-<seed>/checkpoint.pt

For example, the EANN directory layout is as follows:

weibo17/
└── eann/
    ├── resnet/
    │   ├── seed-0/
    │   │   └── checkpoint.pt
    │   ├── seed-42/
    │   │   └── checkpoint.pt
    │   └── seed-2026/
    │       └── checkpoint.pt
    └── vit/
        ├── seed-0/
        │   └── checkpoint.pt
        ├── seed-42/
        │   └── checkpoint.pt
        └── seed-2026/
            └── checkpoint.pt

Download an individual checkpoint using huggingface_hub:

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="MMFact/MMFact-CKPT",
    repo_type="model",
    filename="weibo17/eann/resnet/seed-0/checkpoint.pt",
)

print(checkpoint_path)

Change the filename to select a checkpoint for another dataset, method, variant, or random seed.

Use the corresponding model implementation and configuration in the MMFact toolbox to load and evaluate the checkpoint. Dataset definitions, preprocessing resources, and precomputed features are documented in the MMFact dataset repository.

Citation

If you use these checkpoints or other MMFact resources in your research, please cite our paper:

@misc{miao2026mmfactsimplifyingmultimodalmisinformation,
  title={MMFact: Simplifying Multimodal Misinformation Detection with Shared Resources and Tools},
  author={Yuchen Miao and Zijun Wang and Mingtai Zhang and Xiaoyi Chen and Sihan Deng and Xianglin Ding},
  year={2026},
  eprint={2610.34162},
  archivePrefix={arXiv},
  primaryClass={cs.MM},
  url={https://arxiv.org/abs/2610.34162},
}