A Multi-Modal Fact Verification dataset released as part of the De-Factify workshop in AAAI-22. We won the fifth place and the technical report can be viewed here.
We propose the improved version, Pre-CoFactv2, winning the first prize at the FACTIFY 2 Challenge. Please refer to the new repo of Pre-CoFactv2 for more details.
The task is to find out support, insufficient-evidence and refute between given claims.
- Train model
bash single_model.sh - Evaluate model
python evaluate.py ${model_path} - Ensemble models
python ensemble.py
- Train set: 35,000, 7,000 for each class.
- Validation set: 7,500, 1,500 for each class.
- Test set: 7,500, 1,500 for each class. For more details, please refer to FACTIFY: A Multi-Modal Fact Verification Dataset.
F1 averaged across the 5 categories. The final ranking would be based on the weighted average F1 score.
