jimfhahn/lc-bib-embeddings

Library of Congress bibliographic embeddings — Parquet edition

A Parquet repackaging of the Library of Congress's experimental vector embeddings of its bibliographic records, published on https://id.loc.gov/download/ as "Experimental Embeddings Data by LCC". Each record carries a 1024‑dimensional vector generated by LC with Amazon Titan Text Embeddings V2 (amazon.titan-embed-text-v2:0), plus a few descriptive fields.

LC distributes the data as ~190 GB of gzipped JSON Lines split by Library of Congress Classification (LCC) class letter. This edition stores the same values as Parquet so they can be queried with DuckDB, Polars, pandas or the datasets library without downloading and parsing JSON.

Experimental, and incomplete for now. LC labels these embeddings experimental. As of 2026‑10‑08 three of LC's files — P (language & literature), T (technology) and no_LCC (records without a class) — return HTTP 403 from LC's download bucket even though they are linked from the download page, so they are not included yet. They will be added if LC makes them readable. The 19 classes here cover roughly 12 million records.

Contents

12,226,727 records, 47.1 GB of Parquet (54 files), converted 2026‑10‑08.

class	records		class	records
A General works	117,340		L Education	400,688
B Philosophy, psychology, religion	1,287,475		M Music	722,727
C Auxiliary sciences of history	184,235		N Fine arts	628,142
D World history	1,500,614		Q Science	911,724
E History of the Americas	200,142		R Medicine	486,685
F History of the Americas (local)	380,594		S Agriculture	275,019
G Geography, anthropology	829,701		U Military science	136,717
H Social sciences	2,205,841		V Naval science	48,833
J Political science	414,639		Z Bibliography, library science	281,385
K Law	1,214,226		P, T, no_LCC	not yet available

Columns

column	type	notes
lc_001	string	MARC 001 control number of the LC bibliographic record
lccn	string	Library of Congress Control Number as given by LC (may carry prefixes such as unk)
title	string	title as given in LC's meta.Title
creator	string	main creator as given in meta.Creator (null for ~2%)
lcc_code	string	full LCC call-number class, e.g. VA454
lcc_class	string	the source file / class letter (A…Z); also the config name
embedding	list<float32>, length 1024	Titan v2 vector, unit-normalized (cosine similarity = dot product)

Values are lossless: every vector in LC's JSON is exactly representable as float32 (verified on samples), and that is the stored type.

Querying

-- DuckDB: nearest neighbours within one class, straight from the Hub
WITH v AS (SELECT * FROM 'hf://datasets/jimfhahn/lc-bib-embeddings/data/V/*.parquet'),
     q AS (SELECT embedding::FLOAT[1024] AS e FROM v WHERE title ILIKE '%submarine at war%' LIMIT 1)
SELECT array_inner_product(v.embedding::FLOAT[1024], q.e) AS sim, v.title, v.lcc_code
FROM v, q
ORDER BY sim DESC
LIMIT 10;

from datasets import load_dataset
ds = load_dataset("jimfhahn/lc-bib-embeddings", "V", split="train")

Brute-force search over a single class is fast (≈0.3 s for class V's ~49K records on a laptop). For the full collection, build an ANN index (FAISS, DuckDB's vss extension, etc.).

Searching with free text

Query text must be embedded with the same model, amazon.titan-embed-text-v2:0 (AWS Bedrock), at 1024 dimensions with normalization on. Vectors from any other model are not comparable. Without Bedrock access the data is still useful for record‑to‑record similarity, clustering and visualization.

Provenance

• Source: https://id.loc.gov/download/embeddings/<CLASS>_embeddings_2025.jsonl.gz (served from lds-downloads.s3.amazonaws.com; files last modified 2026‑09‑22).
• Conversion: tools/lc-embeddings/lc_embeddings_to_parquet.py, run as a Hugging Face Job (CPU Upgrade, 27 minutes end to end). Each class file is streamed through DuckDB (read_json → CAST(embedding AS FLOAT[1024]) → Parquet, zstd, ~1 GB shards). Any record whose vector is not exactly 1024‑dimensional would fail the run; none did.
• No records were filtered or modified; row order within a class follows LC's files.

Relation to other datasets

jimfhahn/lc-bibframe-hubs holds LC's BIBFRAME Hub (work-level) graph. Both use LCCNs, from the same numbering scheme, but an LCCN identifies a single record. A Hub carries the LCCN of the name/title authority record it was derived from (n…, no…, nr…); rows here carry the LCCN of a bibliographic record. So lccn is a key within each dataset, not a foreign key between them; an exact join on it matches nothing (checked against class D: 0 of 1.5M).

The working link is lc_001. LC's BIBFRAME Work for a bibliographic record lives at https://id.loc.gov/resources/works/{lc_001}, and that Work points to Hubs through bf:relation when the record has a uniform title, a name/title access point or a series. In a 6‑record spot check, all 6 Works resolved and 1 linked to a Hub (a series Hub). Expect Hub links for a minority of records.

License

The bibliographic data is published by the Library of Congress, a U.S. Government agency, and is in the public domain in the United States. This repackaging is released under CC0 1.0. The embeddings were produced by LC using Amazon Titan; consult LC's download page for any usage notes.