FabioRS/BibliothecaPatristica

Bibliotheca Patristica

Structured OCR corpus of the Patrologia Orientalis (PO), with Patrologia Graeca (PG) and Patrologia Latina (PL) forthcoming.

The Patrologia Orientalis contains patristic texts in Oriental languages — Syriac, Armenian, Coptic, Arabic, Ethiopic — alongside Greek and Latin, making it one of the most linguistically diverse collections of early Christian literature.

Source code (GitHub): https://github.com/Fabio3rs/BibliothecaPatristica
Dataset owner on Hugging Face: FabioRS

Dataset Summary

Version	1.0 (Patrologia Orientalis)
Volumes	20 (PO002, PO003, PO006–PO019, PO021–PO023, PO025)
Pages	16,096
Scripts	Latin, Greek, Syriac, Armenian, Coptic, Arabic, French, English (less frequent), mixed
License	CC BY 4.0

OCR Pipeline

Transcription uses a multistage pipeline:

1. Tesseract — baseline OCR with Latin/Greek language packs, adaptive binarization (CLAHE + adaptive threshold), run in parallel per page
2. LLM transcription — vision model (qwen3.5:397b-cloud through ollama) produces structured XML with typed blocks, bounding boxes, and script identification
3. Deterministic similarity gate — token overlap between Tesseract and LLM outputs is computed; pages with sufficient agreement are marked deterministic_pass without further evaluation
4. LLM judge — pages that fail the gate are evaluated by a second LLM pass comparing the image against the transcription, producing fidelidade (fidelity) and usabilidade (usability) scores

Quality assurance uses a two-tier approach: a deterministic similarity gate comparing Tesseract and LLM token overlap (covering ~98.9% of pages), with LLM judge evaluation reserved for pages where the gate detects significant disagreement. This avoids the prohibitive cost of vision-based evaluation at scale while concentrating review where it matters most.

Quality Distribution

Criterion	Value	Pages
decision	deterministic_pass	15,934
decision	llm_judged	162
fidelidade	alta	104
fidelidade	media	58

Pages with decision = deterministic_pass have fidelidade = null (no LLM judge was invoked). Pages with decision = llm_judged always have a fidelidade value (alta or media); pages judged baixa or descartar are excluded from the published dataset.

Schema

Column	Type	Description
volume	string	Volume ID (e.g. PO009)
pagina	int32	Page number within volume
xml_raw	string	Structured XML transcription (see below)
original_path	string	Path of the per-page TXT inside this repo (teste/POxxx/text/*.txt)
fidelidade	string or null	alta / media / null (null = deterministic pass, no judge invoked)
usabilidade	string or null	alta / media / null (null = deterministic pass, no judge invoked)
idiomas	list[string]	Scripts detected on the page (populated only for llm_judged pages)
decision	string	deterministic_pass / llm_judged
status	string	deterministic_pass / parse_ok

XML Structure

Each page is transcribed as structured XML with typed blocks, script identification, and bounding boxes (scale 0–1000):

<pagina estado="com_texto">
  <bloco tipo="cabecalho" script="latino" bbox="60,20,940,90">
    1007 ORIGENIS 1008
  </bloco>
  <bloco tipo="texto_principal" script="latino" bbox="60,90,490,930">
    tendimus errasse Celsum, cum dixit perspicuum
    esse, qui innatae proclivitati ad peccandum ...
  </bloco>
  <bloco tipo="texto_principal" script="greco" bbox="560,90,940,930">
    Α γορῶνων τῶν Μουσών. Οἱ μὲν οὖν καθ' ἡμᾶς ...
  </bloco>
  <bloco tipo="aparato_critico" script="latino" bbox="60,840,940,920">
    (71) Vel Babylonium intellige ...
  </bloco>
  <bloco tipo="rodape" script="latino" bbox="60,930,940,990">
    Digitized by Google
  </bloco>
  <notas>Two-column layout: left in Latin, right in Greek.</notas>
</pagina>

A Note on Portuguese Metadata

This dataset was produced by a Brazilian developer using a pipeline written in Portuguese. As a result, some metadata fields and XML tags use Portuguese terms. The following glossary maps them to English:

Quality fields (fidelidade / usabilidade)

Portuguese	English	Present in dataset
alta	high	✅ yes
media	medium	✅ yes
baixa	low	❌ excluded before publish
descartar	discard	❌ excluded before publish
null	no judge invoked (deterministic pass)	✅ yes

XML block types (tipo=)

Portuguese	English
cabecalho	page header
texto_principal	main text body
aparato_critico	critical apparatus
rodape	footer
nota	footnote
nota_marginal	marginal note
outro	other

XML page states (estado=)

Portuguese	English
com_texto	contains text
vazio	blank / empty page

XML scripts (script=)

Portuguese	English
latino	Latin
grego	Greek
siriaco	Syriac
armenio	Armenian
copta	Coptic
arabe	Arabic
ethiopico	Ethiopic
cirilico	Cyrillic
georgiano	Georgian
hebraico	Hebrew
ingles	English
frances	French
misto	mixed scripts
desconhecido	unknown

Allowed block types: cabecalho, texto_principal, aparato_critico, rodape, nota, nota_marginal, outro

Allowed scripts: latino, grego, copta, siriaco, cirilico, ethiopico, armenio, arabe, hebraico, georgiano, ingles, frances, misto, desconhecido

Usage

from datasets import load_dataset

ds = load_dataset("FabioRS/BibliothecaPatristica", data_files="data/PO.parquet")

# filter by quality (llm-judged pages only)
high_quality = ds["train"].filter(
    lambda x: x["fidelidade"] == "alta"
)

# include deterministic-pass pages (fidelidade is null for those)
high_quality_with_det = ds["train"].filter(
    lambda x: x["fidelidade"] in ("alta",) or x["decision"] == "deterministic_pass"
)

# filter by script (only populated for llm_judged pages)
syriac_pages = ds["train"].filter(
    lambda x: "siriaco" in (x["idiomas"] or [])
)

Reproducing OCR on a small slice

To re-run OCR on specific pages directly from the original PDFs (e.g., pages 507–508 of volume PO011) with the full verify/judge flow:

python main2.py \
  'PO011.pdf' \
  --algorithm openai \
  --out teste \
  --lang fra+lat+grc+ell+syr+hye-calfa-n+ara+heb+eth \
  --verify-judge-llm \
  --first 507 --last 508 \
  --judge-force

"--lang" is for specifying the languages to be used in the Tesseract OCR process.

Other volumes can be processed in batch via patristicaoriental.sh, which calls main2.py over all PO*.pdf with deterministic gating + LLM judge.

Glossary (Portuguese field & tag names)

• fidelidade: fidelity of the transcription — alta (high), media (medium), null (deterministic pass, no judge invoked). Values baixa and descartar exist in the pipeline but are excluded from the published dataset.
• usabilidade: usability for downstream tasks — same labels as fidelidade.
• idiomas: list of scripts detected on the page (populated only for llm_judged pages).
• Block types: cabecalho (header), texto_principal (main text), aparato_critico (critical apparatus), rodape (footer), nota (note), nota_marginal (marginal note), outro (other).
• Script labels: latino (Latin), grego (Greek), copta (Coptic), siriaco (Syriac), cirilico (Cyrillic), ethiopico (Ethiopic), armenio (Armenian), arabe (Arabic), hebraico (Hebrew), georgiano (Georgian), ingles (English), frances (French), misto (mixed), desconhecido (unknown).
• decision: deterministic_pass (overlap gate OK) or llm_judged (sent to judge).
• status: deterministic_pass or parse_ok.

Limitations & Quality Notes

This corpus has not been reviewed by human experts.

• Pages marked deterministic_pass had strong Tesseract/LLM agreement but were not manually verified
• Pages marked fidelidade: alta or media passed automated LLM judge evaluation only
• Oriental scripts (Syriac, Armenian, Coptic, Arabic) were evaluated for structural presence and approximate extent, not character-level accuracy
• Greek diacritics and ligatures may contain errors even on otherwise well-transcribed pages
• Known unrecoverable pages are excluded from the dataset

Recommended use: LLM training, semantic search, metadata extraction, and document understanding research. Not recommended as a substitute for critical editions without manual verification.

How to Contribute

If you find transcription errors or want to contribute human-verified corrections, please open an issue or PR on the GitHub repository.

Versions

Version	Contents	Status
v1.0	Patrologia Orientalis (PO) — 20 volumes	✅ Released
v2.0	+ Patrologia Graeca (PG)	🔄 In progress
v3.0	+ Patrologia Latina (PL)	📋 Planned

Citation

If you use this dataset in your research, please cite:

@dataset{bibliotheca_patristica_2026,
  author    = {FabioRS},
  title     = {Bibliotheca Patristica},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/FabioRS/BibliothecaPatristica},
  note      = {Multistage OCR corpus of the Patrologia Orientalis.
               Not human-reviewed. CC BY 4.0.}
}

Related Work

• BibliothecaPatristica site — searchable interface over the corpus using PageFind
• latin-codec — empirical research on Latin as a compression/specification format for LLM contexts