LAION-Beyond is the first multi-domain benchmark specifically designed to evaluate the Out-of-Pre-training (OOP) generalization of vision-language models (e.g., CLIP, OpenCLIP, EVA-CLIP).
We distinguish two types of visual concepts:
IP (In-Pre-training): concepts that appear in the pre-training data (e.g.… See the full description on the dataset page:
https://huggingface.co/datasets/MHuangX/LAION-Beyond.
