Large-scale dataset of latent embeddings from ~1700 pretrained vision models across 8 image benchmarks. Enables studying latent geometry, benchmarking semantic alignment methods, and regressing how pretraining strategies shape embedding structure.