Vintage-LLM/MCLModESP

Málaga Corpus of Late Modern English Scientific Prose (clean Markdown)

The 55 texts of the Málaga Corpus of Late Modern English Scientific Prose (MCLModESP): English medical writing, 1703-1896, transcribed from the printed books by the University of Málaga. The corpus covers theoretical treatises, surgical treatises and remedies (recipes). This version is the official plain-text corpus with its layout repaired: tags and page furniture removed, lines joined into paragraphs, and line-end hyphens rejoined. The words, spelling and punctuation of the transcription are not changed.

55 texts, 2,876,826 words, 33,505 paragraphs. 1,554,316 words date from 1700-1799 and 1,322,510 from 1800-1899.

Source

• Corpus page: https://latemodernmss.uma.es/thecorpus/
• Download used: the plain-text corpus, https://latemodernmss.uma.es/wp-content/uploads/2024/11/PlainTextCorpus.zip (dated 2024-11-25). This is the only complete official version: the normalised version has 37 of the 55 texts and the POS-tagged version has 52.
• The corpus can also be searched online in CQPweb: https://latemodernmss.uma.es/cqpweb

The corpus page describes the corpus as "hitherto unedited medical texts produced in the period 1700-1900", with "over 2.5 million words, of which 1.5 million belong to the 18th century and the other million to the 19th".

Licence

The corpus page gives no licence, copyright notice or terms of use. The files are a free public download, and the page asks users to cite the corpus (below). The licence of this dataset is therefore marked unknown. The underlying printed books are from 1703-1896 and are in the public domain. The transcriptions are the work of the Málaga team. Check with them before commercial use.

Citation

Please cite the original corpus, as its authors ask:

Calle-Martín, Javier, Miriam Criado-Peña, Verónica Hernández, Sinéad Linehan-Gómez and Juan Lorente-Sánchez. 2016. The Málaga Corpus of Late Modern English Scientific Prose (MCLModESP). Málaga: University of Málaga. Available from https://latemodernmss.uma.es.

@misc{mclmodesp2016,
  author       = {Calle-Mart{\'i}n, Javier and Criado-Pe{\~n}a, Miriam and Hern{\'a}ndez, Ver{\'o}nica and Linehan-G{\'o}mez, Sin{\'e}ad and Lorente-S{\'a}nchez, Juan},
  title        = {The M{\'a}laga Corpus of Late Modern English Scientific Prose (MCLModESP)},
  year         = {2016},
  publisher    = {University of M{\'a}laga},
  howpublished = {\url{https://latemodernmss.uma.es}}
}

Files

Path	What
data/train-00000-of-00001.parquet	One row per text: the metadata as columns and the Markdown body in text. 5.8 MB, zstd.
<period>/<Author, Title (year)>.md	The same 55 texts as Markdown files with YAML front matter, one folder per quarter century (1700-1724 ... 1875-1899). May be shipped as a compressed archive.

Texts per period: 1700-1724: 7, 1725-1749: 13, 1750-1774: 9, 1775-1799: 6, 1800-1824: 4, 1825-1849: 5, 1850-1874: 4, 1875-1899: 7.

Columns

The parquet columns are the keys of each Markdown file's YAML front matter, plus id, text and file. text is the body only, without the front matter.

Column	Type	Content
id	string	<period>/<Author, Title (year)>
title	string	Title, from the source file's <Title:> header
author	string	Author, from <Author:>. "Unknown author" or "Unknown Author" (as spelled in the source) where the corpus has none.
year	int32	Year of the edition transcribed, from <Year:>
genre	string	From <Genre:>: Theoretical (23), Recipes (17), Surgical (13), Recipe book (1, Joanna Stephens 1740). Null for Bryant 1884.
period	string	Quarter century, as in the corpus folders
source_tokens	int64	The corpus's own token count, from <Tokens:>. Null for Bryant 1884.
words	int64	Letter-and-digit words in text
paragraphs	int64	Paragraphs in text
corpus	string	Corpus name
corpus_url	string	Corpus page
source_zip	string	URL of the official zip
source_file	string	Path of the source .txt inside the zip
source_sha256	string	SHA-256 of that source .txt
text	string	The text as Markdown: one paragraph per line, blank line between paragraphs
file	string	Path of the matching .md file

The Bryant 1884 file has no header in the source, so its title, author and year come from its file name.

from datasets import load_dataset
ds = load_dataset("parquet", data_files="data/*.parquet", split="train")
print(ds[0]["title"], ds[0]["year"], ds[0]["text"][:200])

How the text was made

From each official plain .txt file:

1. Decoding: 54 files are UTF-8 with a byte-order mark, and Bryant 1884 is cp1252. Line ends are normalised to LF.
2. Tags removed: the <Author: ...> <Title: ...> <Year: ...> <Genre: ...> <Tokens: ...> header (kept as metadata) and 59 <p.N> page markers in Mead 1720.
3. Page furniture removed at page breaks: 5,030 page numbers, 1,464 running heads, 1,414 catchwords and 379 signature marks. Eighteenth-century books print the next page's first word and a gathering mark such as B 2 at the foot of each page.
4. Lines joined into paragraphs.
5. Line-end hyphens rejoined ("complica- / tion" → "complication") or kept for compounds ("Rose- / water" → "Rose-water"). The decision uses the Málaga team's POS-tagged version of the same text first, where it exists, since it already resolves these hyphens. Next it uses word counts over the whole corpus. Last, a fallback: two known words keep the hyphen, otherwise the parts are joined if the second starts lowercase.
6. Figure captions and footnotes that interrupt a sentence are moved after the paragraph (1,500 lines).

Check: for every text, the letters and digits of the output, in order, equal those of the source minus the removed furniture, with moved captions and notes checked separately. All 55 pass.

Known limits

• Spelling and transcription errors are kept as in the source ("preparpared", "Inflarnation").
• Many 18th-century books mark paragraphs only by indentation, which the transcription does not keep, so their paragraphs are long.
• Footnotes that do not interrupt a sentence stay where the transcriber put them.
• A few running heads that occur once, next to a finished sentence, are kept. A few captions stay inline where the source joined them to the text.
• The text is not escaped for Markdown. A paragraph starting with a footnote * or a number renders as a list item.

Relation to other copies

The Hugging Face dataset CowFeather/EEBO-Malaga_Coporus contains 54 of these 55 plain texts, byte-identical to the official files. It lacks Richard Mead, Medical Precepts and Cautions (1751). Its other 7 files are EEBO texts, not part of this corpus.