A multimodal dataset pairs photographs with captions, but some captions describe different images. What should happen?

1. Check and correct pair alignment before training
2. Evaluate only caption spelling
3. Increase training epochs without checking labels
4. Treat all mismatches as beneficial augmentation