ROADMAP.md

CatholicCorpus Roadmap

This document tracks planned releases and their deliverables. Each version has a definition of done. No dates are committed; milestones ship when they are ready.

v0.9.0 (Current Release)

The foundation release. Contains the full raw corpus, download scripts, provenance sidecars, and a text extraction pipeline (code only, not yet run).

What shipped:

• 67,772 content files across 16 collections (35.9 GB)
• Idempotent download scripts for all collections (download_script.py per task)
• Provenance tracking via _source.json sidecars
• Shared archive_helpers.py with circuit breaker, lending-restricted detection, and rate limiting
• build_manifest.py for regenerating master_manifest.json and SUMMARY.md
• Text extraction pipeline in text_extraction/ (code included for review, not yet executed)
• Project website at catholiccorpus.org
• Full documentation: README, LICENSE_NOTES, ACKNOWLEDGMENTS, BACKLOG, LENDING_RESTRICTED

Known gaps at this version:

• No token counts
• No deduplication analysis
• No formal _source.json schema documentation
• No Hugging Face dataset card
• No DOI or BibTeX citation
• Text extraction not yet run
• Patrologia Graeca vols 16 and 86 missing (98.8% coverage without them)
• 23 lending-restricted archive.org items not downloadable

v1.0.0 (Foundation Release)

The first release suitable for citation in academic work. Adds the measurements and metadata that ML/NLP researchers expect.

Definition of done:

• Token counts computed and published (overall, per-collection, per-language)
• Per-source composition breakdown (files, bytes, tokens, languages, formats)
• _source.json schema formally documented (JSON Schema + human-readable docs)
• dataset.json machine-readable manifest at corpus root
• Hugging Face dataset card following community conventions
• BibTeX citation block and DOI (via Zenodo)
• CITATION.cff at repo root
• LIMITATIONS.md published
• Corpus published on Hugging Face Datasets

v1.1.0 (Extracted Text Layer)

A pre-built plain-text layer so researchers do not need to run extraction themselves.

Definition of done:

• text_extraction/extract_text.py run across all tasks
• Per-document extraction quality scores (character count, empty-page ratio)
• Extracted text published as a separate Hugging Face configuration
• Quality report comparing extracted text to source (sample-based)

v1.2.0 (OCR Enhancement)

Re-OCR of low-quality archive.org scans to improve text extraction for older PDFs.

Definition of done:

• OCR engine selected (Tesseract 5 with theological dictionaries, or a vision-language model)
• Low-quality PDFs identified (by extraction quality scores from v1.1)
• Re-OCR applied and quality comparison report published
• Updated extracted text layer published

v1.3.0 (Critical-Edition Partnerships)

Higher-quality source texts where licensing permits.

Definition of done:

• Outreach to Leonine Commission, Brepols-adjacent OA texts, CUA Press, Editiones Cistercienses
• At least one critical-edition source integrated
• Provenance tracking updated for new sources

v2.0.0 (Aligned Bilingual Editions)

Sentence-aligned Latin-English (and Greek-English) parallel texts for translation NLP research.

Definition of done:

• Alignment methodology documented
• At least three authors aligned (Aquinas, Bonaventure, Augustine)
• Published as a separate Hugging Face configuration

How to Contribute

See CONTRIBUTING.md for how to propose new collections, flag OCR errors, or contribute to any of these milestones.