Catalog Update

Replace previous catalog (213,937,319 rows; original used for BioCLIP 2 training) with latest catalog (233,055,986 rows), used for BioCLIP 2.5 Huge training.

Provenance Update

Schema is preserved to stay consistent with the current provenance.

uuid, source_id, data_source, source_url, license_name, copyright_owner, license_link, title, bibliographicCitation 

Row-count assertion passed: matches catalog 233,055,986 exactly.

Construction

• 211,152,120 rows preserved as-is from HF main snapshot (214M rows - 2.8M dropped UUIDs no longer in new catalog)
• 21,903,866 net-new rows derived:
  • 21,901,107 gbif: license_link from per-HDF5 metadata, license_name via HF lookup -> CC URL regex -> "other" fallback, copyright_owner from occurrence.rightsHolder (fallback "not provided"), title = "not provided", bibliographicCitation from GBIF occurrence (else NULL)
  • 2,759 eol: license_name from eol per-HDF5 metadata, license_link via HF reverse-lookup, copyright_owner set to "not provided", title set to "not provided", bibliographicCitation = NULL

title and copyright_owner: 0 null everywhere, matching current provenance literal "not provided" convention.

Worth Noticing

title is 100% NULL "not provided" in the existing and new provenance...

In a new PR, we will add the embeddings:

• another copy of existing text embeddings, with species replaced with bioclip-2
• the BioCLIP 2.5 Huge text embeddings with bioclip-2.5-vith14

In the end, we'll have embeddings/txt_emb_{model_name}.npy and embeddings/txt_emb_{model_name}.json for each model, as described here.