AbdouMagico/cross-scriptural-verse-matcher

Cross-Scriptural Verse Matcher

A unified framework for finding theologically relevant verses across the Old Testament (OT), New Testament (NT), and the Quran based on semantic meaning and latent thematic guidance.

Architecture

• Base Encoder: intfloat/multilingual-e5-base — multilingual sentence embeddings, strong zero-shot retrieval
• Latent Guidance Head: Multi-label thematic classifier (45 theological themes) on top of embeddings
• Scripture-Type Embedding: Learned embedding for OT/NT/Quran to condition the shared space
• Loss: Combined contrastive (InfoNCE with hard negatives) + thematic BCE classification

Training Recipe

Component	Value
Base model	intfloat/multilingual-e5-base
Fine-tuning	LoRA (r=16, alpha=32) on Q,K,V,Dense
Max length	256
Batch size	32
Learning rate	3e-4
Temperature	0.05
Theme loss weight (λ)	0.3
Epochs	5
Optimizer	AdamW with cosine schedule
Precision	bf16

Dataset

Cross-scriptural verse pairs generated via LLM annotation with:

• Similarity scores (0.0–1.0)
• Relationship types: thematic, narrative, prophetic, lexical, ethical, cosmological
• Hard negatives: same-theme verses with different meaning
• 45 theological themes for latent guidance

Source datasets:

• Quran: freococo/quran_multilingual_parallel (English)
• Bible: davidguzmanr/open-bible-resources (English Standard, verse-level)

Usage

Training

python train.py

Inference

python inference.py --query "For God so loved the world" --top_k 5

Repositories

• Model: https://huggingface.co/AbdouMagico/cross-scriptural-verse-matcher
• Dataset: https://huggingface.co/datasets/AbdouMagico/cross-scriptural-pairs

Citation

Built on:

• SimCSE (Gao et al., 2021) — contrastive sentence embeddings
• E5 (Wang et al., 2022) — weakly-supervised text embedding
• multilingual-e5 (XLM-R backbone) — cross-lingual alignment