Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning

FakeVV dataset

Please note that the training sets for the third stage, which include fakett and fakesv, are not displayed due to access restrictions.

FakeVV dataset test set

The complete test dataset is defined in:

• data_config/fakevv_test_data.json.json

Includes webpage url

• data_config/fakevv_test_data_with_urls.json

The associated visual resources are hosted on Hugging Face:

• Video frames:
  https://huggingface.co/datasets/fanrui00/FakeVV_testset

• Downloaded (crawled) videos:
  https://huggingface.co/datasets/fanrui00/FakeVV_testset_video

FakeVV Three-Stage Training Dataset

The annotation files for the three-stage FakeVV training data are hosted at:

• Annotations & Metadata:
  https://huggingface.co/datasets/fanrui00/fact-r1-train-jsons
  This repository also includes metadata for each video. Further details can be found in its included README.md.

• Corresponding Video Files (.mp4):
  https://huggingface.co/datasets/fanrui00/FakeVV_trainset_video

• Corresponding Video Frames (optional):
  https://huggingface.co/datasets/fanrui00/FakeVV_trainset

Requirements

Software Requirements

• Python 3.9+
• transformers>=4.49.0
• flash-attn>=2.4.3
• vllm>=0.7.3

Hardware Requirements

* estimated

Method	Bits	1.5B	3B	7B
GRPO Full Fine-Tuning	AMP	2*24GB	4*40GB	8*40GB

Installation

cd fact-r1
pip install -e .

GRPO Training

bash examples/qwen2_5_vl_7b_fact_r1_grpo.sh

Merge Checkpoint in Hugging Face Format

python3 scripts/model_merger.py --local_dir path_to_your_last_actor_checkpoint

Thanks

We would like to thank the following repos for their great work:

• This work is built upon the EasyR1 and veRL.