pre1600: Western texts with a 1600 knowledge cutoff

A pretraining corpus in which, as far as can be checked, every word was written before 1601. It covers Latin, Ancient/Byzantine Greek, Old/Middle/Early Modern English and European vernaculars. It was built to train a small language model from scratch and study how its reasoning holds up on problems from after its world ended.

Most of the work went into removing the modern layer around old texts: editors' introductions, footnotes, variant readings, modern translations, section numbers, catalogue records and OCR banners. Every document records how it was dated, and every removal is logged (provenance/).

from datasets import load_dataset

ds = load_dataset("jaevibin/pre1600")                      # strict tier, all languages
latin = load_dataset("jaevibin/pre1600", "strict_latin")
reasoning = load_dataset("jaevibin/pre1600", "reasoning_passages")

Contents

language group	strict docs	strict words	strict tokens*	extended-only docs	extended-only words
English (Old, Middle, Early Modern, Scots)	3,226	132.7M	181.9M	1,910	58.7M
Latin	4,411	50.5M	84.8M	305	2.8M
Greek (Ancient, Byzantine)	2,499	39.1M	71.1M	1,267	25.2M
German (Early New High German)	56	1.25M	2.5M	2	<0.01M
French (Old French to 16th c.)	52	1.13M	2.2M	–	–
Italian	50	0.62M	1.3M	–	–
Welsh	8	0.13M	0.3M	–	–
Gothic	16	0.06M	0.2M	–	–
Old Norse / Icelandic	3	0.04M	0.1M	29	0.46M
total	10,321	225.6M	344.4M	3,513	87.1M (~133M tokens)

*Tokens are counted with the included 32k byte-level BPE tokenizer (tokenizer/tokenizer.json), which was trained only on this corpus.

By date of composition (strict): BC 4%, 1–500 9%, 501–1000 8%, 1001–1300 14%, 1301–1500 7%, 1501–1600 59%. The 16th century dominates because of English printed books (EEBO-TCP).

Configs

config	contents
strict (default)	text transcribed by people or conventional OCR, in original or edition spelling. Use this one for contamination-sensitive work.
extended	only the extra documents that didn't make strict (see Tiers)
all	strict + extended
strict_english, strict_latin, strict_greek, strict_vernaculars	strict, by language group
reasoning_passages	8,052 passages with explicit argument structure, extracted from the corpus
reasoning_curated_works	952 whole works on logic, geometry, arithmetic, dialectic and disputation (these also appear in the main configs)

Tiers

extended holds documents that are genuinely pre-1600 but carry a modern layer that strict avoids:

• EEBO-TCP Phase II English books: only available here in the Visualizing English Print project's partly standardised spelling (begynneth → begins). The original-spelling XML exists (EarlyPrint / Oxford Text Archive) and would move ~59M words into strict.
• Greek transcribed or post-corrected by a modern vision-language model (the Open Greek Corpus's own OCR of Migne and Teubner scans). A model can quietly invent plausible text.
• Apparatus-dense editions: editions where the filters had to cut editorial matter more than once per 4,000 words, so uncaught notes are likely.
• IcePaHC Icelandic: in modern standard orthography.

Schema

strict / extended / all

field	type	description
id	string	stable id, source:native-id (e.g. tcp:A00429, pl:tmp953.tmp953.opp-lat1)
text	string	paragraphs separated by \n\n; verse keeps line breaks
language	string	ISO 639-3 (eng, enm, ang, sco, lat, grc, gkm, deu, fra, fro, ita, isl, non, cym, got)
language_group	string	english, latin, greek, german, french, italian, norse, welsh, gothic
tier	string	strict or extended
date_hi	int	latest plausible year of composition (negative = BC). Always ≤ 1600.
date_lo	int?	earliest year, where known
date_basis	string	how the date was established: imprint, manual, collection, canon_century, author_death, first_publication, written/print, century
title, author	string	from the source catalogue
source	string	source collection (see Sources)
edition	string	source edition / identifier
spelling	string	original (diplomatic, e.g. EEBO-TCP), edition_normalised, standardised
transcription	string?	Greek only: digital_edition, ocr, vlm_ocr, llm_corrected
license	string	licence of this document's source
flags	list[string]	spelling_standardised, dense_apparatus
meta	string (JSON)	source-specific extras (EEBO subjects, detected language, TLG id, apparatus density…)

reasoning_passages: kind (scholastic_article, euclidean_proof, catechism_qa, disputation, worked_problem), doc_id (the corpus document it came from), language, date_hi, tier, title, text.

reasoning_curated_works: id, category (geometry, arithmetic, logic, applied_math, dialectic, disputation), title, author, language, date_hi, tier, text.

How it was built

In short (full detail in DATASHEET.md; code in code/pre1600_pipeline.zip):

1. Dating. A document is kept only if its latest plausible date of composition is ≤ 1600, read conservatively:

   • EEBO imprint ranges use their upper bound;
   • a play written in 1587 but first printed in 1618 is out;
   • all 619 Latin Library entries were dated by hand (Bacon, Kepler, Galileo, Descartes, Newton and modern Latin translations are excluded);
   • editors' prefaces bound as "works" in Migne are dropped, as is the pseudo-Dexter forgery of 1619.

2. Original language only. No modern translations (including 19th/20th-c. German translations in DiBiLit, and the English column of bilingual Aquinas pages). Pre-1600 translations (Tyndale, Golding's Caesar, Fischart) stay.

3. Edition layer removed:

   • TEI headers, notes, apparatus, variant readings and figure descriptions;
   • editors' Arabic line/section numbers and critical signs in Latin/Greek editions;
   • editors' diacritics in Latin;
   • Latin-script apparatus inside Greek passages.

4. Paragraph tripwires, checked with a line/sentence fallback so that one note doesn't cost a book. They catch:

   • web/digitisation boilerplate;
   • post-1600 concepts (telescope, oxygen, medieval, Elizabethan…);
   • apparatus vocabulary (MS., codd., coni., ibid., vide notam…);
   • names of 17th–19th-c. scholars cited in editors' Latin notes (Mabillon, Baluze, Tillemont, Baronius…);
   • post-1600 years;
   • modern-language paragraphs inside Latin editions (language ID);
   • modern-vocabulary prose inside Middle English and Old French editions;
   • OCR garbage.

   In original early prints (EEBO) these hits are logged but not dropped, because they are period usage there: ibid., "old English", Newton as a surname.

5. Near-duplicate removal across sources, at paragraph level (sampled 13-gram overlap on spelling-normalised text), keeping the cleanest witness.

6. Audit of the final text: tripwire re-scan (no hits left) and a list of post-1600 "canary" names with context. Every remaining hit is a legitimate pre-1600 homonym: Napoleone Orsini, Thomas Cromwell, Richard Newton the 15th-c. judge.

Known limitations

• Residual editorial prose. An editor's note that trips no filter can survive, especially in Migne's Patrologia. The worst-affected editions were moved to extended.
• Modern normalisation in editions. Punctuation, u/v, i/j and paragraphing in critical editions are modern conventions. EEBO-TCP, the largest block, is diplomatic (original spelling, abbreviations and punctuation).
• Genre skew. Roughly half the EEBO words are religious (sermons, controversy, devotion). Science, medicine and mathematics come to ~2% of EEBO; Latin and Greek add Aristotle, Euclid, Galen, Ptolemy and Aquinas.
• Size. 344M strict tokens supports a model of roughly 100M–300M parameters trained for 4–10 epochs. Much larger models will memorise it.
• Omitted for licensing. The Corpus of Middle English Prose and Verse (126 documents, ~12.6M words) was used in the build but is not included here, because its licence forbids redistribution. The pipeline in code/ can rebuild it locally.

Licences

This is a compilation of sources under different licences; each row carries its own license field. See LICENSE.md for the per-source terms and attribution. Several sources are non-commercial (CC BY-NC-SA) and several are share-alike (CC BY-SA), so the compilation as a whole should be treated as non-commercial, share-alike, research use. If you need a commercially usable subset, filter on license.

Sources

EEBO-TCP (Text Creation Partnership); Visualizing English Print; PerseusDL canonical Latin and Greek; Open Greek and Latin (Patrologia Latina, CSEL); Open Greek Corpus (First1KGreek, Perseus, PTA, Galenus Verbatim, byzantium.gr, calfa-co Patrologia Graeca, DFHG, SAWS, GLAUx); The Latin Library (via CLTK); digilibLT; Corpus Grammaticorum Latinorum; dhspriory.org Aquinas (via Geremia/AquinasOperaOmnia); scrollmapper Bible databases; DiBiLit (Deutsches Textarchiv); DraCor (Italian, German, French); Geste; OpenMedFr; heimskringla.no Eddas; IcePaHC; Anglo-Saxon Poetic Records (via sacred-texts / CLTK). Please credit these projects if you use the data.