enver/classical-arabic-scholastic-bilingual-corpus

Classical Arabic Scholastic Bilingual Corpus (70,192 Pairs)

Dataset Summary

The Classical Arabic Scholastic Bilingual Corpus is a curated, dense-semantic parallel dataset designed for high-fidelity translation between Classical Arabic and Academic English. It covers foundational texts across Classical Islamic Philosophy (Falsafah), Dialectical Theology (Kalām), Legal Epistemology (Uṣūl al-Fiqh), Hadith Commentary, and Spiritual Realization (Taṣawwuf).

The dataset was mined from an 80-book, 339 MB canonical library using Google's Language-Agnostic BERT Sentence Embeddings (LaBSE) executed in FP16 on Blackwell GPUs, enforcing a strict semantic cosine similarity gate ($\text{similarity} \ge 0.70$).

Dataset Structure

Each record contains:

• arabic (string): Vocalized/normalized classical Arabic sentence.
• english (string): Academic English translation.
• score (float): LaBSE cross-lingual semantic cosine similarity ($\ge 0.70$).
• source (string): The originating canonical classical work.

Example:

{
  "arabic": "وإن ادعيتم النظر فكيف يتصور أن يعلم الشئ ضرورة ثم يعلم كونه ضروريا",
  "english": "and if you claim it by reflection, then how can it be conceived that a thing is known necessarily, and then its being necessary is known by reflection.",
  "score": 0.7271,
  "source": "ghazali__al_iqtisad_fi_al_itiqad_v4_translated"
}

Source Collections & Breakdown

Collection / Master	Books	Pairs Extracted	Share	Key Works Included
Fakhr al-Dīn al-Rāzī (v4)	18	27,516	39.2%	Al-Maṭālib al-ʿĀliyah (Vols 1–9), Al-Maḥṣūl, Asās al-Taqdīs, Al-Arbaʿīn, Al-Tafsīr al-Kabīr
Imām al-Nawawī (v5)	22	20,982	29.9%	Sharḥ Ṣaḥīḥ Muslim, Al-Majmūʿ, Rawḍat al-Ṭālibīn, Al-Adhkār
Abū Ḥāmid al-Ghazālī (v4)	26	11,576	16.5%	Iḥyāʾ ʿUlūm al-Dīn, Al-Mustaṣfā, Al-Iqtiṣād, Al-Mankhūl, Al-Munqidh
Rāghib al-Iṣfahānī (Unversioned)	6	8,154	11.6%	Al-Mufradāt fī Gharīb al-Qurʾān, Jāmiʿ al-Tafsīr, Muḥāḍarāt al-Udabāʾ
Imām al-Mawwāq (v5)	3	787	1.1%	Sunan al-Muhtadīn
Classical Heritage (v5)	4	560	0.8%	Al-Futūḥāt al-Makkiyyah (Ibn ʿArabī), Al-Shifāʾ (Qāḍī ʿIyāḍ)
Mustafa Sabri (v6)	1	242	0.3%	ʿIlm Ādāb al-Baḥth wa'l-Munāẓarah
Foundational Anchors	—	375	0.5%	Core scholastic definitions weighted 25x
Total	80	70,192	100%	339 MB Source Library

Strict Versioning & OCR Noise Elimination Protocol

To eliminate translation drift and scan artifacts:

1. Versioning Priority: Enforced $\mathbf{v6} \succ \mathbf{v5} \succ \mathbf{v4} \succ \text{unversioned}$.
2. Exclusion of Corrupted OCR Scans: Mawqif al-ʿAql (Vols 1–4) and Mawqif al-Bashar by Mustafa Sabri were systematically audited and found to contain up to 80% OCR noise from 1950 letterpress scans. They were completely excluded from this dataset to prevent model poisoning. Only the 100% clean Ādāb al-Baḥth was retained.

Citation

@dataset{aynengine2026classical_bilingual,
  author = {AynEngine Research Group},
  title = {Classical Arabic Scholastic Bilingual Corpus},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/enver/classical-arabic-scholastic-bilingual-corpus}
}