Clouda OCR v0.2.1

Clouda OCR v0.2.1 — Release Notes

Clouda OCR v0.2.1 is a maintenance and hardening release that completes the current brand migration from the historical Clouda PDF name to Clouda OCR, while preserving the existing v0.2.0 release as historical.

Highlights

• Canonical project/package identity prepared for Clouda OCR v0.2.1.
• Dataset quality gate now runs text near-duplicate leakage detection in the real pipeline.
• Cross-split near-duplicate findings now fail the quality verdict, including Arabic diacritic/tatweel/whitespace variants covered by regression tests.
• Failed quality verdicts can no longer silently produce an artifact that looks like an approved clean manifest.
• Dataset download provenance now records whether a digest is registry-pinned or self-reported.
• Unpinned dataset downloads fail closed unless the explicit opt-in environment variable is enabled.
• Reliability hardening from the engineering audit includes queue cancellation/retry fixes, duplicate-dispatch protection, SQLite claim locking, atomic publication paths, safer backup/audit naming, dead-symlink handling, Windows reserved-name protection, and dependency hardening.

Verification

Release-candidate verification reported:

• python -m pytest -q → 2169 passed, 16 skipped, 0 failed
• Ruff → pass
• Black → pass
• mypy → pass
• pip-audit → 0 known vulnerabilities
• wheel build → pass
• sdist build → pass
• clean wheel install → pass
• clean sdist install → pass
• CLI/import/runtime smoke tests → pass

Bandit still reports two previously triaged findings:

• intentional Arabic bidi characters in the RAQM corpus path
• a digest-guarded torch.load path

These were reviewed as non-blocking for this release.

Important limitations

• No final project-trained OCR model is bundled.
• OCR inference remains disabled by default unless a licensed, validated model is integrated.
• The expanded 462-page held-out benchmark is separate from this software release and remains in progress; this release does not claim a new benchmark winner or production OCR accuracy.
• Dataset integrity verification is stronger, but publisher authenticity is not yet backed by a signing infrastructure. Operators should prefer pinned trusted digests.

Historical note

v0.2.0 remains the historical Clouda PDF release.

v0.2.1 is the first release candidate using the canonical Clouda OCR project identity.

Clouda PDF v0.2.0

Clouda PDF v0.2.0

First public release of Clouda PDF, an Arabic-first, model-agnostic PDF-to-DOCX runtime with isolated data, training, and model subsystems, aimed at open, self-hostable, independently evaluated Arabic Document AI infrastructure.

Highlights

• Arabic-first PDF-to-DOCX runtime for Arabic, English, and mixed-language documents.
• Trusted born-digital PDF text extraction — embedded text must pass the canonical digital-text trust gate before direct extraction is authorized.
• RTL-aware DOCX generation — valid, editable, text-only DOCX output preserving page order, Arabic Unicode, RTL paragraph direction, footers, and page boundaries.
• Explicit page routing states — pages are classified as digital_text, blank_page, near_blank, pending_ocr_model, or review_required; scanned, image-only, or uncertain pages are kept in explicit review states instead of being presented as OCR output.
• Integrated Clouda Data Factory (clouda_data.factory) — deterministic synthetic Arabic OCR data generation (ingest, Arabic RTL rendering, atomic + composite distortions, QC, exports, JSONL/CSV manifests, seeds, resume, multiprocessing) with factory-* commands in the canonical CLI, a lossless factory-to-canonical-manifest adapter, SIL OFL 1.1 bundled fonts, and Training Experiment Framework dry-run integration. The standalone clouda-data-factory repository is superseded.
• Supporting data and training infrastructure — pre-training dataset preparation (provenance-preserving manifests, conservative Arabic normalization, validation, exact deduplication, leakage-safe splitting, JSONL export), training planning, Results Store, the loopback-only Clouda Lab dashboard, and a published metadata-only 177-page Arabic OCR benchmark (v0.1.0).

Important limitations

• No final trained OCR model is bundled. Image-only scanned pages remain in pending_ocr_model unless a licensed, validated OCR model is actually integrated.
• OCR inference remains disabled by default unless a suitable validated model is integrated.
• A larger 462-page held-out evaluation is in progress and currently paused pending additional compute capacity; production model selection remains open. This release makes no claims about that evaluation.
• Production application accuracy is not claimed.

License

Apache-2.0 applies to original Clouda OCR source code and project materials in this public repository. Third-party datasets, models, and benchmark source materials remain governed by their respective licenses — see LICENSE, NOTICE, and DATA_LICENSES.md.