biglam/bnl-monographs

BnL Monographs

33,456 sections from 504 monographs digitised by the Bibliothèque nationale du Luxembourg, published between 1690 and 1918. Each row is one section — a chapter, preface, index, illustration and so on — with its OCR text and the Dublin Core metadata the BnL records for it.

This is the BnL's Monograph Text Pack. It is books rather than newspapers, so it does not overlap the BnL newspaper corpora already on the Hub (biglam/bnl_newspapers1841-1879, biglam/bnl_ground_truth_newspapers_before_1878).

section type	rows
CHAPTER	29,983
ILLUSTRATION	1,043
TABLE_OF_CONTENTS	581
VOLUME	470
COVER_SECTION	468
INDEX	384
APPENDIX	201
PREFACE	142
ADVERTISEMENT	97
CORRECTIONS	44
MAP	30
BIBLIOGRAPHY	13

Structure

One train split. 79.2M characters of text in total; no row has empty text, and lengths run from 1 character to 496,745 (median 567).

field	notes
text	the section's OCR text
language	ISO codes, a list — 1,632 sections carry two and 61 carry three
title	section title as recorded
section_type	CHAPTER, PREFACE, … (see above)
monograph_ark	groups sections into books — 504 distinct values
date	year of publication, as a string — see below
publisher	publisher as recorded
extent	page extent as recorded
url	persistent persist.lu link to the section
section_id	identifier from the source filename
relation	the BnL record id
is_referenced_by · has_version	the source record's own cross-references
oai_identifier · oai_datestamp	OAI-PMH header fields
raw_xml	the complete original record, verbatim

from datasets import load_dataset

ds = load_dataset("biglam/bnl-monographs", split="train")

# sections belong to books; group by monograph_ark to reassemble one
book = ds.filter(lambda r: r["monograph_ark"] == ds[0]["monograph_ark"])

Caveats

date is not a clean year. It is kept as a string because the source uses 9999 as a placeholder and uncertain-digit forms such as 187u (meaning "some year in the 1870s"). Casting to int will fail or silently mislead. Filter defensively: ds.filter(lambda r: r["date"].isdigit() and r["date"] != "9999").

468 rows have a blank or Sans titre title. That is exactly the number of COVER_SECTION rows, so it is structural rather than missing data.

The text is raw OCR, uncorrected. Expect the usual errors on historical typefaces, and note that a section's text quality varies with the condition of the original.

Language is recorded, and the collection is genuinely multilingual. French 16,238 · German 15,115 · Luxembourgish 1,816 · Latin 1,240 · English 167 · Dutch 68. It is a list because 1,632 sections are tagged with two languages and 61 with three; 566 sections carry no language tag at all, so filter rather than assume.

No authorship. The deposit ships a dc:creator element on 802 records, but it is empty in every one of them, so no author information exists to carry over. The field is not included.

The original XML is kept. raw_xml holds each OAI-PMH record verbatim (126 MB in total), so nothing in the source is lost to the parse and any field not broken out into a column can still be recovered without re-downloading from the BnL.

Section counts differ slightly from the BnL's own figures. The BnL describes the pack as 51,709 pages and 33,477 extracted chapters; this build yields 33,456 sections, of which 29,983 are typed CHAPTER. The source deposit does not explain the difference.

Licence and credit

Public domain (CC0-1.0), as released by the BnL.

Source data created and digitised by the Bibliothèque nationale du Luxembourg; mirrored and repackaged by Daniel van Strien (Machine Learning Librarian, Hugging Face). Contact for the source data: opendata@bnl.etat.lu.

Maintenance: limited — this is a repackaging, and corrections belong upstream with the BnL.