Meinh/historical-christian-corpus

Historical Christian Corpus (Pre-1800)

A machine-readable corpus of Christian literature composed on or before 31 December 1800. Assembled for offline reasoning research, continued pretraining (CPT), fine-tuning, and historical-text analysis.

Dataset Summary

This corpus aggregates primary-source Christian texts from multiple public-domain repositories, covering the full historical span from biblical texts through the late eighteenth century. It preserves original-language material (Greek, Latin, Hebrew, Syriac, Coptic, Armenian, Ge'ez, Church Slavonic) alongside historical English translations.

Total: 3,673 documents, 4,025 MB of cleaned text.

Source breakdown

Source	Files	Notes
Internet Archive	~3,000	OCR text from scanned public-domain books
OpenChristianData	12	Curated JSONL: bible, church fathers, sermons, commentaries, hymns
CatholicCorpus	83	Catholic theological texts (text layer)
CCEL	~44	Schaff's Ante-Nicene + Nicene/Post-Nicene Fathers, classics
Wikisource	118	Transcribed public-domain texts

Period distribution

Works span eight historical periods:

1. Biblical texts and early Christian writings (pre-100)
2. Apostolic Fathers and early Church (100-325)
3. Patristic Christianity (325-750)
4. Late antiquity and early medieval (750-1050)
5. Medieval Christianity and scholastic theology (1050-1500)
6. Reformation and Counter-Reformation (1500-1648)
7. Seventeenth-century Christian thought (1648-1700)
8. Eighteenth-century Christian thought (1700-1800)

Supported Tasks

• Continued pretraining (CPT) - domain-adapt a modern base model to historical Christian literature
• Text generation - completion, style transfer, historical register modeling
• Feature extraction - embeddings for retrieval over historical theological text
• Language modeling - n-gram, neural LM, or tokenizer training

Dataset Structure

Data Fields

Field	Type	Description
id	string	Stable local record ID
source	string	Repository of origin
source_url	string	Canonical URL on source
title	string	Work title (may be file-derived)
author	string	Author name (empty if unknown)
composition_date	string	Best-known composition date or range
period	string	One of eight historical periods
language	string	Language of the downloaded text
tradition	string	Christian tradition, where known
work_type	string	commentary, sermon, confession, treatise, etc.
license	string	Licensing status of the source
text	string	Cleaned full text
char_count	int	Character count
word_count	int	Approximate word count
sha256	string	Content hash for deduplication

Data Splits

Split	Rows	Size
train	3,637	3,992 MB
validation	36	34 MB

Validation split uses seed=42, 1% of files.

Usage

Load the full corpus with the datasets library:

from datasets import load_dataset
ds = load_dataset("Meinh/historical-christian-corpus")
print(ds)
train = ds["train"]
print(train[0]["title"], train[0]["word_count"])

Stream without downloading:

ds = load_dataset("Meinh/historical-christian-corpus", streaming=True)
for row in ds["train"]:
    print(row["source"], row["title"])
    break

Filter by source or period:

ia = ds["train"].filter(lambda x: x["source"] == "internet_archive")
patristic = ds["train"].filter(lambda x: x["period"] == "patristic")

Tokenized training artifact

The tokenized/ directory contains the exact binary used for CPT:

• train.bin - 1,187,882,980 tokens (uint16, little-endian)
• val.bin - 10,286,257 tokens
• corpus_tokenizer.model - SentencePiece BPE, vocab 18,432
• corpus_tokenizer.vocab - vocab list

Load tokens with numpy:

import numpy as np
tokens = np.memmap("tokenized/train.bin", dtype=np.uint16, mode="r")
print(len(tokens), tokens[:20])

Decode with SentencePiece:

import sentencepiece as spm
sp = spm.SentencePieceProcessor()
sp.load("tokenized/corpus_tokenizer.model")
print(sp.decode([int(t) for t in tokens[:20]]))

Curation Rationale

This corpus was assembled to support offline, reasoning-first research on historical Christian literature. Design choices:

• Composition-date cutoff at 1800 - enforces historical authenticity. Works composed in 1801+ are excluded by default.
• Original-language preservation - Greek, Hebrew, Latin, Syriac, Coptic, Armenian, Ge'ez, and Church Slavonic are retained in their Unicode form. No translation or transliteration was performed.
• Source diversity - deliberately spans Catholic, Orthodox, Protestant, and other traditions. No single denomination is treated as representative.
• Doctrinal diversity preserved - the corpus contains conflicting theological positions. This is intentional and necessary for reasoning over historical debate.
• Conservative cleaning - page furniture, running headers, and OCR debris removed; short paragraphs filtered; cross-file duplicate paragraphs (500+ chars appearing in 2+ files) dropped. No semantic rewriting.

Cleaning Pipeline

Original text went through three stages:

1. Extraction - HTML/XML/EPUB/PDF text extraction; no OCR by default
2. Cleaning - scripts/clean_text.py strips IA furniture, standalone page numbers, running headers, control characters, and cross-file duplicate paragraphs
3. Tokenization - scripts/train_tokenizer.py, scripts/pack_data.py produce SentencePiece BPE (vocab 18,432, byte-fallback), packed into uint16 arrays

Total reduction: ~10.6% byte reduction (4,363 MB -> 3,903 MB).

Personal and Sensitive Information

This is a corpus of historical texts, most published before 1800. It contains:

• Historical theological polemic, including material that is antisemitic, anti-Catholic, anti-Protestant, or otherwise hostile to specific groups
• Historical discussions of violence, heresy, punishment, and moral obligation that reflect the norms of their time
• Names and biographical details of historical persons, all long deceased

No modern personal data is included. No contemporary private individuals are represented.

Considerations for Using This Data

Social Impact

The corpus documents historical Christian thought, including positions that modern readers may find objectionable. Downstream models trained on this data may reproduce historical biases. Users should:

• Not treat the corpus as a source of modern theological authority
• Not present historical opinions as universally representative of Christianity
• Not use the corpus to generate content that targets religious groups

Biases

• Geographic: Skewed toward Western European and English-language sources, especially in the post-Reformation period. Eastern Christian, African, and Asian sources are underrepresented relative to their historical importance.
• Linguistic: ~79% of the corpus is OCR'd English text from Internet Archive, which introduces OCR-specific errors and biases the token distribution.
• Tradition: Catholic and Protestant sources dominate. Orthodox, Oriental Orthodox, and non-Chalcedonian traditions are underrepresented.
• Temporal: The 17th-18th century is over-represented relative to patristic and medieval material because more of it was digitized.

Limitations

• Composition dates are uncertain for many works. The composition_date field is often empty or an estimate; the date_confidence field in the source manifest records uncertainty.
• Authorship is incomplete. Many IA items lack author metadata.
• OCR quality varies. IA text is machine-extracted from scanned images. Quality ranges from clean to noisy.
• No doctrinal filtering. The corpus contains material from mutually incompatible traditions. A model trained on it will reproduce contradictions.
• Composition-date cutoff is approximate. Books published before 1800 may contain later editorial additions; conversely, works composed before 1800 but only surviving in later editions may be missing.

Licensing Information

The corpus is a mixed-license compilation. Each source has its own terms:

Source	License
Internet Archive	Public domain (US) for pre-1929 publications; verify per item
OpenChristianData	Public domain (CC0 stated by publisher)
CatholicCorpus	Mixed; public domain + CC; some items copyrighted
CCEL	Text is public domain; CCEL XML formatting is copyrighted
Wikisource	CC BY-SA / public domain

The corpus as a whole is distributed under the most permissive interpretation of the union of these licenses: mixed-public-domain. Individual files may be subject to different terms. See LICENSE.md for details and per-source terms.

Users are responsible for verifying that their use of the corpus complies with the terms of the sources they use. For redistribution, academic use, and non-commercial research, the public-domain items are unencumbered.

Citation

If you use this corpus, cite:

@misc{historical_christian_corpus_2026,
  title  = {Historical Christian Corpus (Pre-1800)},
  author = {Meinh},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Meinh/historical-christian-corpus},
  note   = {Aggregated from Internet Archive, CCEL, CatholicCorpus, OpenChristianData, Wikisource}
}

Source Provenance

Per-item provenance is recorded in metadata/master_manifest.jsonl. Each record includes:

• local_id, source, source_url, download_url
• title, author, composition_date, date_confidence
• language, format, license
• file_path, size_bytes, sha256
• retrieved_at (UTC timestamp)

Failed downloads are logged in metadata/failed.jsonl for transparency.

Reproduction

The corpus can be regenerated from source using the scripts in scripts/:

python3 scripts/fast_downloader.py          # download from sources
python3 scripts/estimate_v2.py              # measure raw corpus
python3 scripts/clean_text.py               # clean text
python3 scripts/train_tokenizer.py          # train SentencePiece BPE
python3 scripts/pack_data.py                # pack into uint16 arrays

Requirements: Python 3.10+, requests, sentencepiece, numpy, pyarrow, pandas.

Changelog

• v1.0 (2026-10-15): Initial release. 3,673 documents, 1.27B tokens.

---

Built with the goal of preserving historical Christian literature for reasoning research. All sources are public domain or explicitly permissive. No copyrighted material is knowingly included.

Recent updates

2026-10-16

• URL stripping added to clean_text.py pipeline. Removes http(s)://..., www.*, and bare domain tokens. Cleared URL fragments from the tokenizer vocabulary and training data.
• Tokenizer retrained on URL-stripped text. Vocab size bumped from 16,384 to 18,432 (headroom for future modern-text SFT).
• Data re-tokenized end to end with the new tokenizer.
• New counts:
  • train.bin: 1,187,882,980 tokens (was 1,262,048,974)
  • val.bin: 10,286,257 tokens (was 10,648,838)
  • Clean-text reduction: 11.7% (4,363 MB -> 3,851 MB)
• Old tokenizer preserved at corpus_tokenizer_v1.model / .vocab.