Manara-3B: Sovereign SLM for the Kuwaiti Banking Sector

Manara-3B (مناره-3B) is a sovereign, bilingual (Business Arabic / English) Small Language Model (SLM) purpose-built for the Kuwaiti banking and financial sector. Built upon the Jais Arabic foundation model, Manara-3B is designed for on-premise deployment and provides deep, regulation-grounded intelligence across three critical domains.

Architecture Overview

The Manara-3B pipeline is structured as a three-phase system, following a Teacher-Student distillation paradigm.

┌─────────────────────────────────────────────────────────────────────────┐
│                         MANARA-3B PIPELINE                              │
│                                                                         │
│  ┌──────────────┐    ┌──────────────────┐    ┌──────────────────────┐  │
│  │  STEP 1      │    │  STEP 2          │    │  STEP 3              │  │
│  │  Teacher     │───▶│  Student         │───▶│  Self-Reflection     │  │
│  │  (Jais API)  │    │  (LoRA/Unsloth)  │    │  (Critic Pass)       │  │
│  │              │    │                  │    │                      │  │
│  │ 50,000 pairs │    │ 3B-param SLM     │    │ 12 Guardrails        │  │
│  │ CBK/Sharia/  │    │ 16GB VRAM        │    │ JSON Correction Log  │  │
│  │ Boursa KW    │    │ safetensors out  │    │ Recursive Retraining │  │
│  └──────────────┘    └──────────────────┘    └──────────────────────┘  │
└─────────────────────────────────────────────────────────────────────────┘

Domain Coverage

Manara-3B is trained on three core knowledge domains specific to the Kuwaiti financial ecosystem.

Domain	Coverage	Weight
CBK 2026 Regulations	Capital adequacy, AML/CFT, Open Banking, KACH, KDMS, Cybersecurity Framework	40%
Sharia-Compliant Finance	Murabaha, Wakala, Ijara, Mudarabah, Musharakah, Sukuk, AAOIFI standards	35%
Boursa Kuwait Standards	ESG reporting, financial disclosure, corporate governance, IFRS, TCFD	25%

Repository Structure

manara-3b/
├── README.md                       # This file
├── model_card.md                   # Professional HuggingFace model card
├── requirements.txt                # Python dependencies
├── config.yaml                     # Central configuration file
│
├── src/
│   ├── distillation/
│   │   ├── generate_synthetic_data.py  # Step 1: Teacher distillation pipeline
│   │   └── preprocess_dataset.py       # Step 1b: Data validation & Alpaca formatting
│   │
│   ├── training/
│   │   ├── fine_tune_lora.py           # Step 2: LoRA/Unsloth fine-tuning script
│   │   └── inference.py                # Inference engine with guardrail integration
│   │
│   └── guardrails/
│       └── self_reflective_loop.py     # Step 3: 12-guardrail critic engine
│
├── data/
│   ├── raw/                        # Raw generated data (gitignored)
│   └── processed/                  # Processed Alpaca-format datasets (gitignored)
│
├── logs/                           # Training & guardrail logs (gitignored)
│   ├── distillation.log
│   ├── training.log
│   └── self_corrections.jsonl      # Self-correction events for retraining
│
└── model_weights/                  # Saved model weights (gitignored)
    ├── manara-3b-final/            # Merged float16 model (safetensors)
    └── manara-3b-lora/             # LoRA adapters only (lightweight)

Quickstart

1. Installation

git clone https://github.com/your-org/manara-3b.git
cd manara-3b
pip install -r requirements.txt

# Recommended: Install Unsloth for 2x faster training
pip install unsloth

2. Step 1: Generate Synthetic Training Data

export JAIS_API_KEY="your-jais-api-key"

python src/distillation/generate_synthetic_data.py \
    --output_dir data/processed \
    --num_pairs 50000 \
    --model jais-13b-chat

3. Step 1b: Preprocess the Dataset

python src/distillation/preprocess_dataset.py \
    --input_file data/processed/manara_train.jsonl \
    --output_file data/processed/manara_alpaca.jsonl

4. Step 2: Fine-Tune the Model

python src/training/fine_tune_lora.py \
    --config config.yaml \
    --train_file data/processed/manara_train_alpaca.jsonl \
    --val_file data/processed/manara_val_alpaca.jsonl \
    --output_dir model_weights

5. Step 3: Run Inference with Guardrails

python src/training/inference.py \
    --model_path model_weights/manara-3b-final \
    --prompt "ما هي متطلبات نسبة كفاية رأس المال وفق تعليمات بنك الكويت المركزي؟" \
    --language ar

The 12 Self-Reflective Guardrails

Every response generated by Manara-3B is automatically verified against 12 domain-specific guardrails. Any violation is logged to logs/self_corrections.jsonl for recursive retraining.

ID	Guardrail Name	Severity
G-01	No Riba (Interest) Statements	Critical
G-02	CBK Regulatory Accuracy	Warning
G-03	Sharia Compliance Verification	Critical
G-04	Boursa Kuwait Disclosure Standards	Warning
G-05	No Fabricated Regulatory Citations	Critical
G-06	No Harmful Financial Advice	Critical
G-07	Currency & Jurisdiction Accuracy (KWD)	Warning
G-08	AAOIFI Standard Alignment	Warning
G-09	No Discriminatory Content	Critical
G-10	Data Privacy & Confidentiality (PII)	Critical
G-11	Kuwaiti Dialect & Register Appropriateness	Info
G-12	Factual Consistency & Internal Coherence	Warning

Hardware Requirements

Deployment Mode	VRAM	Notes
Training (LoRA + Unsloth)	16 GB	Optimized for single GPU on-premise
Inference (4-bit quantized)	8 GB	Suitable for production deployment
Inference (float16 merged)	16 GB	Higher accuracy, more VRAM

License

This project is licensed under the Apache 2.0 License. See LICENSE for details.

Citation

If you use Manara-3B in your research or applications, please cite:

@misc{manara3b2026,
  title        = {Manara-3B: A Sovereign Small Language Model for the Kuwaiti Banking Sector},
  author       = {Manus AI},
  year         = {2026},
  howpublished = {\url{https://github.com/your-org/manara-3b}},
  note         = {Fine-tuned on Jais, optimized for CBK regulations, Sharia finance, and Boursa Kuwait standards.}
}

Acknowledgements

This project builds upon the foundational work of the Jais team at Inception (G42) and MBZUAI, whose open-source Arabic LLM made this work possible. The fine-tuning methodology draws from the Unsloth and TRL libraries.

Downloads last month: -

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support