Mathani-Ayat/rattil-ayah-embed

Rattil Ayah Embed

Sentence embedder for Quran ayahs. Rattil uses it to pick confusable wrong answers for its challenges (ayahs and words that are close in meaning). It is a fine-tune of Omartificial-Intelligence-Space/Arabic-Triplet-Matryoshka-V2 (Apache-2.0, 135M parameters), which beat seven larger and Quran-tuned models on QurSim.

Results

QurSim verse pairs linked by Ibn Kathir's tafsir. 20% of verses are held out by a hash of the verse number: no training pair touches them. 459 held-out queries rank their strongly related partners among all 4,114 verses.

	Recall@10, 768-d	MRR, 768-d	Recall@10, 256-d	MRR, 256-d
Arabic-Triplet-Matryoshka-V2 (base)	0.308	0.252	0.286	0.236
rattil-ayah-embed	0.374	0.319	0.371	0.301

The Rattil index stores the first 256 dimensions as int8.

Training

1,645 QurSim strongly related pairs, MatryoshkaLoss (768/512/256/128/64) over MultipleNegativesRankingLoss, 3 epochs, batch 32, lr 2e-5, seed 13. The details are in rattil_training.json; the script is src/learning/train_ayah_embedder.py in the app repo.

Use

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Mathani-Ayat/rattil-ayah-embed", truncate_dim=256)
vectors = model.encode(["قل هو الله احد", "الله الصمد"], normalize_embeddings=True)

In the Rattil app, python src/deployment/pull_hf_assets.py --models ayah-embed puts it in runs/rattil_ayah_embed, and python src/learning/build_text_embeddings.py builds the index from it.

Limits

• QurSim uses Hafs spelling and numbering, and reflects one tafsir's links.
• The QurSim data license is not established. Keep private and research-only.