Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

lcsh-db-ft — LCSH authority DB embedded with fine-tuned EmbeddingGemma

lcsh.db — a SQLite + sqlite-vec database of Library of Congress authorities, with the LCSH + LCGFT vector index built from the LCSH-fine-tuned EmbeddingGemma-300M (256-d). The companion to the embedder kltng/embeddinggemma-300m-lcsh-onnx; intended for an on-device / Chrome-extension subject-cataloging workflow.

Parallel to the stock-embedded kltng/lcsh-db; only the auth_embeddings vectors differ (fine-tuned instead of stock EmbeddingGemma).

Contents

Table Rows Notes
auth LCSH 512,644 · LCGFT 2,637 · LCNAF 12,255,750 authority records
alt_label UF / variant labels
auth_embeddings (vec0) 916,142 LCSH+LCGFT labels + alt-labels, 256-d float32
auth_fts, alt_label_fts FTS5 for exact/fuzzy string lookup

db_meta.embedding_model = kltng/embeddinggemma-300m-lcsh-onnx. LCNAF names are present for exact/FTS lookup but not embedded (same policy as the stock db).

Query (must use the matching embedder)

Embed the query with kltng/embeddinggemma-300m-lcsh-onnx, truncate to 256-d, L2-normalize, then KNN:

SELECT auth_id, distance FROM auth_embeddings WHERE embedding MATCH :qvec AND k = 10;
-- then join auth_id -> auth.label

Use fp16 for the closest parity to the build vectors (cos ≈ 0.999); q8 (smallest, what these vectors were spot-checked against) is also fine (cos ≈ 0.97).

License

Vectors derived from google/embeddinggemma-300mGemma Terms of Use. LCSH/LCGFT/LCNAF authority data is from the Library of Congress (public domain).

Downloads last month
31