Encyclopaedia Britannica (1771–1860)

An OCR text corpus of the first eight editions of the Encyclopaedia Britannica, together with the 1801 and 1815–1824 supplements, drawn from National Library of Scotland scans.

The dataset provides two complementary views:

• records, the default training view: 41,594 text chunks, 126,547,014 whitespace-separated words, grouped into 241 edition/letter combinations.
• pages, the source-oriented view: 114,951 page readings from 161 volumes, including material excluded from the training text.

The Parquet files are the authoritative dataset. A matching data/export.jsonl is provided for pipelines that prefer JSONL.

Coverage

Years are the publication-year labels stored in the records. Page counts include front matter, plates, indexes and other pages not necessarily represented in training chunks. Special groups include Appendix and Dissertations.

Edition	Years	Letter groups	Chunks	Page rows	Record words
First edition	1771	24	933	2,371	2,438,731
Second edition	1778–1783	25	3,269	8,606	9,573,262
Third edition	1788–1797	24	5,325	14,555	16,270,489
Supplement 1801	1801	22	585	1,620	1,784,928
Fourth edition	1801–1810	24	5,832	16,041	17,244,139
Fifth edition	1815–1817	24	5,730	16,004	17,206,188
Sixth edition	1820–1823	24	5,764	16,022	17,180,812
Supplement 1824	1815–1824	24	1,280	4,955	4,343,785
Seventh edition	1830–1842	25	6,345	17,064	19,894,555
Eighth edition	1853–1860	25	6,531	17,713	20,610,125
Total	1771–1860	241	41,594	114,951	126,547,014

The selected scans comprise one set of the volumes for these editions and supplements, rather than every copy/reissue in the NLS collection. Separate index volumes are not included. The combined alphabetic groups I-J and U-V reflect the editions' alphabet conventions. Not every supplement has every letter group.

JSONL export

data/export.jsonl contains all 41,594 record objects and all 19 record columns, in sorted Parquet-shard order. Text and metadata values match the written record Parquet exactly; no additional export filtering is applied.

• UTF-8, one JSON object per line.
• 752,156,707 bytes.
• This is the record table only, not a mixture of records and raw pages.

Data files

There are 15 Zstandard-compressed Parquet files, totalling 576,906,413 bytes: five record shards and ten edition-level page files. Exact sizes, row counts and SHA-256 hashes are in manifest.json.

Record file	Rows
data/records/part-0000.parquet	8,879
data/records/part-0001.parquet	8,753
data/records/part-0002.parquet	8,709
data/records/part-0003.parquet	8,113
data/records/part-0004.parquet	7,140

Page files are data/pages/first_edition.parquet through eighth_edition.parquet, plus supplement_1801.parquet and supplement_1824.parquet; their row counts are given in the coverage table.

Supporting files:

• manifest.json: complete file inventory, measured totals and source/model information.
• verification_report.json: independent checks of the written Parquet tables and complete JSONL export.

Record schema - 19 columns

Column(s)	Type	Meaning
doc_id	string	Unique chunk id, e.g. EB4__H__0215. Supplements use EB3S and EB6S; combined-letter ids use IJ/UV.
edition	string	Edition/supplement display name.
edition_no	float64	Edition number; supplements are 3.5 and 6.5.
years	string	Publication-year label or range.
letter	string	Alphabetic group, I-J, U-V, Appendix or Dissertations.
chunk	int64	Contiguous, zero-based chunk index within an edition/letter group.
text	string	Cleaned training text, retaining paragraph breaks, quotations and mathematical content.
n_words, n_chars	int64	len(text.split()) and len(text); characters are Unicode characters, not bytes.
table_words	int64	Visible whitespace-separated words/cells removed from HTML tables on contributing pages; charged once to the first chunk referencing each page.
n_pages	int64	Number of contributing page references in this chunk.
first_page, last_page	string	NLS labels of the first and last contributing pages.
volumes	list of string	Distinct contributing volume ids, in traversal order.
page_orders	list of int64	Ordered page positions within those volumes; length equals n_pages.
capped_pages, looped_pages, latex_pages, redone_pages	int64	Counts of contributing pages carrying the corresponding processing flags.

A page can contribute to multiple chunks. Consequently, summing n_pages or per-chunk flag counters does not count unique pages.

volumes and page_orders are not parallel arrays. Resolve page orders against the first volume until the order resets/decreases, then advance to the next volume. A page identity is (edition, vol_id, page_order).

Chunk sizes are 977–6,469 words, median 2,590; the 5th–95th percentile range is 2,021–5,128. Chunks are intended as manageable text units, not guaranteed complete articles. Joining texts with blank lines in chunk order reconstructs each training letter group.

Page schema - 23 columns

Column(s)	Type	Meaning
vol_id, volume, vol_date	string	NLS volume id, volume title and publication-date label.
edition	string	Edition/supplement display name.
page_order	int64	Stored position within the volume, not necessarily the printed page number.
label	string	NLS page label.
letter	string	Page-level alphabetic/special-group assignment; individual I/J and U/V pages can contribute to combined record groups.
text	string	Page reading, including HTML tables and material excluded from training chunks.
running_head	nullable string	Recognized running-head metadata, where available.
non_prose, index, duplicate	bool	Page classifications used to exclude material from the record view. Pages remain in this table.
capped	bool	Reading was marked as potentially output-limited.
looped	bool	Repetition was identified/processed; not a guarantee that no repetition remains.
latex	bool	Reading contains mathematical LaTeX representation.
redone	bool	Page received additional transcription processing.
out_tokens	nullable int64	Model-response token count where available; not the stored text's whitespace word count. Assembly/cleanup can change the final text.
nls_tokens	nullable int64	Witness-word count used in an agreement calculation.
nls_recall, nls_support	nullable float64	Approximate lexical agreement with an NLS OCR witness: witness coverage and transcription support. Not transcription accuracy.
witness_offset	nullable int64	Witness displacement in stored page-order positions, where alignment was established; null when unconfirmed.
nls_image, nls_alto	string	Relative image/ALTO paths in the NLS source archive. An ALTO pointer alone does not certify alignment.

For the source-reviewed subset, agreement uses NFC/casefold/long-s normalization and multiset overlap of Unicode word/number tokens. Unreliable witness matches are left null rather than scored against unrelated text. Other agreement metadata originates in the OCR process and can use different normalization/crop scope; scores should not be treated as a uniform accuracy benchmark. NLS OCR itself contains historical-font and alignment errors.

Assembly and source review

The OCR models used is allenai/olmOCR-2-7B-1025-FP8. Targeted source review covers 107 pages, with 86 selected source transcriptions. Whole-page images and separate column/region crops were used to establish reading order and retain body text, notes and mathematical material. Some numerical material was transcribed as structured cells, and source-image inspection was used where necessary. Review decisions were checked against the scan rather than accepted solely from word-agreement scores.

The text policy aims to retain historical spelling, punctuation, quotation languages and mathematical notation, with long s (ſ) normalized to s. [?] marks unreadable material. Greek and Latin quotations are retained rather than translated. Where archaic numeral glyphs cannot be represented reliably in plain text, uncertainties are recorded explicitly.

The default record view omits flagged non-prose/index/duplicate pages and HTML table blocks. Mixed prose/table pages can contribute their prose. Mathematical inequalities are retained as text, not mistaken for HTML. Visible table-cell words removed from contributing pages total 313,014 under the documented accounting rule.

110,581 distinct pages contribute to records, through 140,405 page references. The other 4,370 page rows are:

• 4,353 classified non-prose pages, including all 180 index pages;
• 9 duplicate pages, corroborated by comparing their source images with retained counterparts;
• 8 additional table-only pages with no training prose.

All remain available in the page configuration. Current processing flags additionally cover 298 capped, 1,451 looped, 4,131 LaTeX and 3,551 redone pages. Flags overlap.

Verification and limitations

Independent validation reads the written Parquet files and every exported JSON object. It checks schemas, unique identities, contiguous chunks, page references/endpoints, flag counters, stored word/character lengths, file totals and exact Parquet/JSONL equality. Record text contains no detected YAML frontmatter, common HTML tags, invalid control characters or degenerate single-character runs of 100 or more.

The record view contains 733,344,275 Unicode characters and 734,179,720 UTF-8 text bytes, excluding metadata and JSON escaping. These are direct text-size counts, not results from an external tokenizer or statistics scanner.

This is an OCR corpus, not a fully proofread diplomatic edition. Source-image review is targeted and partly model-assisted, not manual validation of every page. Residual spelling/numerical errors, imperfect Greek/transliteration, reading-order mistakes, running heads and page-end hyphenation can remain. Page classifications and letter assignments are approximate; consult the page view when exhaustive source content is required. High lexical agreement does not prove correct reading order, and a low score can reflect faulty NLS OCR rather than a faulty transcription.

Source, rights and attribution

National Library of Scotland, Encyclopaedia Britannica, NLS Data Foundry:

• https://data.nls.uk/data/digitised-collections/encyclopaedia-britannica/
• DOI: 10.34812/cg4r-dn40

NLS describes the scans as public domain through 1853 and no known copyright for 1854–1860. Consult the source collection for the applicable rights statement. Please credit the National Library of Scotland as the source of the scans and source OCR.