- We need to train Hyperbolic Adapters - We should try to do this with CLIP embeddings first - Likely requires a dataset of Hierarchical Data