lcshbench / README.md
kltng's picture
LCSHBench v1.0 (private staging)
882fc85 verified
|
Raw
History Blame Contribute Delete
2.13 kB
metadata
license: cc0-1.0
language:
  - en
pretty_name: LCSHBench v1.0
task_categories:
  - text-classification
  - text-retrieval
tags:
  - library-science
  - subject-indexing
  - LCSH
  - cataloging
  - multilingual
configs:
  - config_name: dev
    data_files: dev/dev.jsonl
  - config_name: dev_2k
    data_files: dev_2k/dev_2k.jsonl
  - config_name: test
    data_files: test/test.jsonl
  - config_name: consensus_matched
    data_files: consensus_matched/consensus_matched.parquet
  - config_name: vocab
    data_files: vocab/vocab.jsonl

LCSHBench v1.0

A multilingual, consensus-grounded benchmark for Library of Congress Subject Heading (LCSH) assignment. Ground truth is the agreement of independent research libraries (Columbia, Harvard, Princeton) cataloging the same work.

Configs

config rows description
dev 18,993 development split (full per-catalog + merged + unanimous GT)
dev_2k 2,002 language-balanced 2K evaluation subset
test 3,353 held-out test inputs; answers in gt_test.hashed.json (SHA-256)
consensus_matched 1,286,609 every work cataloged by >=2 libraries: per-library headings (hashed IDs) — the provenance/concordance population
vocab 515,281 LCSH+LCGFT retrieval vocabulary

Inter-cataloger concordance (why consensus)

Over 465,187 works cataloged by all three libraries: 93.3% share a concept-level heading, only 0.2% share none (median concept Jaccard 0.86) — yet just 39.4% assign identical exact sets and 35.6% of assertions are single-source. Subject identification is objective; expression is subjective. Human ceiling (a library vs the consensus of the other two): 86.9% exact / 93.0% concept recall.

Not included

The fine-tuned embedder and te3 embedding indices are not released (model not published; te3 vectors are derived from a hosted API). Raw MARC is redistributable from the source libraries; rebuild instructions in the repo.

License & citation

Data: CC0-1.0 (derived from CC0 bulk MARC). Identifiers are hashed. Built by the v2.2 extraction pipeline. Citation: TBD.