somayaeltanbouly/Badr_embedding_v0

Badr Embedding v0

Badr Embedding v0 is an Arabic sentence-transformer model specialized for answer-bearing retrieval in Islamic jurisprudence (fiqh). It is fine-tuned from mohamed2811/Muffakir_Embedding to retrieve passages that contain the ruling or information required by a fiqh question, rather than passages that are only topically similar.

The model was developed as part of the research presented in:

What Makes a Good Fiqh Retriever? Answer Retrieval for Arabic Islamic Jurisprudence
Somaya Eltanbouly, Heba Sbahi, Samer Rashwani, Abdessalam Bouchekif, Mutaz al-Khatib, Shahd Gaben, and Mohammed Ghaly.
arXiv:2608.20246

Model repository: somayaeltanbouly/Badr_embedding_v0

Model Details

Property	Value
Architecture	Sentence Transformer / bi-encoder
Base model	mohamed2811/Muffakir_Embedding
Language	Arabic
Domain	Islamic jurisprudence (fiqh)
Embedding dimension	768
Maximum sequence length	512 tokens
Similarity function	Cosine similarity
Training objective	MatryoshkaLoss + MultipleNegativesRankingLoss
Matryoshka dimensions	768, 512, 256, 128, 64

Intended Use

The model is designed for Arabic semantic retrieval, particularly:

• answer-bearing passage retrieval for fiqh questions;
• dense retrieval in retrieval-augmented generation (RAG) systems;
• semantic search over Arabic Islamic jurisprudence collections;
• retrieval experiments involving domain-specific Arabic questions and passages.

This is a retrieval model, not a standalone system for generating or validating religious rulings. Retrieved passages should be interpreted in their source context.

Usage

Install Sentence Transformers:

pip install -U sentence-transformers

Load the model and retrieve the most relevant passage:

from sentence_transformers import SentenceTransformer
import torch

model = SentenceTransformer("somayaeltanbouly/Badr_embedding_v0")

query = "ما الحكم المعتبر في جواز قصر الصلاة عند الحنابلة؟"

passages = [
    "المعتبر نية المسافة لا حقيقتها، فلو رجع قبل استكمالها فلا إعادة عليه...",
    "يجوز للمصلي قطع الصلاة لإنقاذ غريق على الصحيح من المذهب...",
    "إذا نكر السلام في التشهد لم يجزه على الصحيح من المذهب..."
]

query_embedding = model.encode(
    [query],
    convert_to_tensor=True,
    normalize_embeddings=True
)

passage_embeddings = model.encode(
    passages,
    convert_to_tensor=True,
    normalize_embeddings=True
)

scores = torch.matmul(query_embedding, passage_embeddings.T)[0]
best_idx = int(torch.argmax(scores))

print("Best passage:", passages[best_idx])
print("Score:", float(scores[best_idx]))

Matryoshka Embeddings

The model was trained with Matryoshka Representation Learning at dimensions 768, 512, 256, 128, and 64. Embeddings can therefore be truncated when lower-dimensional representations are useful for storage or retrieval efficiency.

For example:

embedding = model.encode(
    ["ما حكم المسح على الخفين؟"],
    normalize_embeddings=True
)

embedding_256 = embedding[:, :256]

When comparing truncated embeddings, normalize them again after truncation before computing cosine similarity.

Training

The model was fine-tuned using triplets consisting of:

• anchor: an Arabic fiqh question;
• positive: a passage containing the answer;
• negative: a non-answer-bearing passage.

Dataset

Split	Samples
Training	18,354
Evaluation	965

Training Configuration

Parameter	Value
Epochs	3
Batch size	128
Learning rate	1e-5
Warmup steps	42
Precision	BF16
Optimizer	AdamW
Loss	MatryoshkaLoss over MultipleNegativesRankingLoss

Evaluation

The Sentence Transformers triplet evaluation included with this checkpoint reports:

Metric	Score
Cosine Accuracy	0.9731

This score measures triplet discrimination on the model's fiqh-eval evaluation set. For retrieval evaluation and comparisons with lexical, dense, hybrid, fine-tuned, and madhhab-aware retrieval strategies, see the associated paper.

Limitations

The model is specialized for Arabic fiqh retrieval and may not generalize equally well to unrelated domains or other languages. Retrieval quality may also vary with question formulation, passage length, terminology, source style, and the availability of answer-bearing passages in the indexed corpus.

The model returns semantic similarity scores; these scores should not be treated as verification that a retrieved passage is jurisprudentially correct or sufficient on its own.

Associated Paper

What Makes a Good Fiqh Retriever? Answer Retrieval for Arabic Islamic Jurisprudence

The paper studies answer-bearing retrieval for Arabic fiqh and evaluates dense, lexical, hybrid, fine-tuned, and madhhab-aware retrieval strategies.

• Paper: https://arxiv.org/abs/2608.20246
• arXiv ID: 2608.20246

If you use this model in research, please cite the paper:

@misc{eltanbouly2026fiqhretriever,
  title        = {What Makes a Good Fiqh Retriever? Answer Retrieval for Arabic Islamic Jurisprudence},
  author       = {Somaya Eltanbouly and Heba Sbahi and Samer Rashwani and Abdessalam Bouchekif and Mutaz al-Khatib and Shahd Gaben and Mohammed Ghaly},
  year         = {2026},
  eprint       = {2608.20246},
  archivePrefix = {arXiv},
  primaryClass = {cs.IR},
  url          = {https://arxiv.org/abs/2608.20246}
}

Framework Versions

The released checkpoint was trained with:

• Python 3.10.20
• Sentence Transformers 5.3.0
• Transformers 5.3.0
• PyTorch 2.9.0+cu128
• Accelerate 1.13.0
• Datasets 5.0.0
• Tokenizers 0.22.2