SAHM

Arabic Financial Instruction-Tuning Dataset & Models

---

---

🌟 Overview

SAHM is the first large-scale Arabic financial NLP benchmark covering both modern financial analysis and Islamic/Shari’ah-compliant reasoning, introduced in our paper SAHM: Arabic Financial Instruction-Tuning Dataset And Models.

It includes 14,000+ high-quality Arabic samples across eight tasks, derived from:

• AAOIFI Shari’ah standards
• Official fatwa archives
• Corporate earnings reports
• Market news
• Business and accounting exams
• Islamic finance regulatory material

SAHM also introduces:
🟦 SAHM-7B-Instruct, an Arabic financial instruction-tuned model
🟦 A unified evaluation framework
🟦 First-of-its-kind datasets for Islamic finance + Arabic corporate analysis

---

📌 Features

Our benchmark includes eight diverse tasks:

1. Islamic Finance Shari’ah Standards QA
2. Islamic Financial Fatwa QA
3. Islamic Financial Fatwa MCQ
4. Business MCQ
5. Accounting MCQ
6. Financial Report Sentiment Analysis
7. Report Extractive Summarization
8. Event–Cause Reasoning QA

These tasks reflect real Arabic financial workflows, combining modern finance with Islamic jurisprudence (fiqh al-muʿāmalāt).

---

📁 Dataset Structure

Each dataset adheres to a unified JSON schema and standardized evaluation protocol.

Task	#Train	#Eval	Format	Capability
Shari’ah Standards QA	1621	406	QA	Islamic finance legal reasoning
Islamic Fatwa QA	11,703	250	QA	Faith-based financial rulings
Event–Cause Reasoning	160	40	QA	Financial causal inference
Extractive Summarization	160	40	Summary	Financial disclosure extraction
Fatwa MCQ	–	250	MCQ	Recognition-style reasoning
Business MCQ	381	76	MCQ	Business fundamentals
Accounting MCQ	95	24	MCQ	Numerical & IFRS reasoning
Sentiment Analysis	160	40	MCQ	Financial polarity detection

Full details appear in Table 1 of the paper.

---

🧠 SAHM-7B-Instruct Model

A 7B Arabic-centric model instruction-tuned on all SAHM datasets.
Built on top of ALLAM-7B, it achieves:

• Best MCQ performance among Arabic/open models
• +37.5 improvement in Accounting MCQ
• Strong business & sentiment accuracy
• Competent but still developing open-ended reasoning

See Table 2 in the paper for full comparison.

---

🏆 Leaderboard (MCQ Tasks)

Model	Mean Accuracy (%)
GPT-5	73.9
GPT-4o	67.0
Qwen2.5-72B	60.4
Fanar-1-9B	53.9
ALLAM-7B	56.1
SAHM-7B-Instruct (ours)	71.7

---

🏆 Leaderboard (Open-Ended QA)

Average Judge Score (0–10):

Model	Score
GPT-5	8.98
Claude 4 Sonnet	7.77
GPT-4o	7.08
Gemini 2.5 Pro	5.73
ALLAM-7B	5.05
Fanar-1-9B	4.82
SAHM-7B-Instruct	5.07

(Open-ended tasks remain significantly harder for current Arabic models.)

---

📝 Installation

git clone https://github.com/mbzuai-nlp/SAHM
cd SAHM
pip install -r requirements.txt

💾 Using the Dataset

from sahm import load_dataset

ds = load_dataset("sahm", "fatwa_qa")
print(ds["train"][0])

🤖 Using SAHM-7B-Instruct

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("mbzuai-nlp/SAHM-7B-Instruct")
model = AutoModelForCausalLM.from_pretrained("mbzuai-nlp/SAHM-7B-Instruct")

prompt = "اشرح حكم بيع المرابحة في حالة عدم تملك السلعة."
out = model.generate(**tok(prompt, return_tensors="pt"))
print(tok.decode(out[0], skip_special_tokens=True))

🗂 Repository Structure

SAHM/
│
├── data/
│   ├── shariah_standards/
│   ├── fatwa_qa/
│   ├── mcq/
│   ├── sentiment/
│   ├── summarization/
│   └── event_cause/
│
├── models/
│   └── SAHM-7B-Instruct/
│
├── docs/
│   └── assets/logo.png
│
├── evaluation/
└── README.md

📘 Citation

If you use SAHM, please cite:

@article{sahm2025,
  title={SAHM: Arabic Financial Instruction-Tuning Dataset And Models},
  author={Elbadry, Rania and Ahmad, Sarfraz and Bouch, Dani and Ahsan, Momina and Peng, Xueqing and Huang, Jimin and AlMahri, Muhra and Khalil, Marwa Elsaid and Wang, Yuxia and Lahlou, Salem and Stoyanov, Veselin and Ananiadou, Sophia and Nakov, Preslav and Xie, Zhuohan},
  year={2025},
  institution={MBZUAI}
}

🏷 License

The dataset and code are released under MIT License