CLIP ViT-B/32 embeddings of OSV5M (407k images, 225 countries)

Precomputed 512-d L2-normalized CLIP ViT-B/32 embeddings so you don't need ~10h of CPU to rebuild them. Derived from OSV5M (CC-BY-4.0, https://huggingface.co/datasets/osv5m/osv5m) — images were embedded and discarded; no pixels here.

• embeddings_train.tar: 31 parquet shards, 386,828 rows (id, country, lat, lon, quadtree cell, emb[512])
• embeddings_test.tar: 2 parquet shards, 20,512 rows, same schema
• country_text_emb.parquet: CLIP text embeddings of 'a photo taken in X' for all 222 countries (the zero-shot baseline)

Unpack into data/emb/ then run make insert.