Omartificial-Intelligence-Space / Diraya-3B-Instruct-Ar

Diraya-3B-Instruct-Ar

Model Description

Diraya-3B-Instruct-Ar is an Arabic reasoning-specialized language model fine-tuned from Qwen2.5-3B .

This model is part of the DIRA (Diraya Arabic Reasoning AI) collection, which focuses on enhancing the logical inference and mathematical reasoning capabilities of Arabic language models.

Key Features

• Arabic-first reasoning : Optimized specifically for complex reasoning tasks in Arabic
• Structured reasoning format : Trained to output reasoning in a clear XML format
• Mathematical reasoning : Enhanced ability to solve multi-step math problems
• Instruction-tuned : Follows instructions reliably in Arabic
• Lightweight : Based on the efficient 3B parameter model architecture

Technical Details

Base Model : Qwen2.5-3B via unsloth/Qwen2.5-3B-Instruct-unsloth-bnb-4bit

Model Type : Instruction-tuned causal language model

Architecture :

• 36 transformer layers
• 16 attention heads for queries (GQA)
• 2 attention heads for keys/values
• Context length: 32,768 tokens

Training Approach :

• Fine-tuned using GPRO
• Training focused on structured reasoning output format using XML tags
• Optimized for mathematical reasoning using the Arabic GSM8K dataset
• Multiple reward functions including correctness, format adherence, and output structure

LoRA Configuration :

{
  "peft_type": "LORA",
  "r": 64,
  "lora_alpha": 64,
  "lora_dropout": 0,
  "target_modules": [
    "k_proj", "gate_proj", "o_proj", "down_proj", 
    "v_proj", "up_proj", "q_proj"
  ],
  "bias": "none",
  "inference_mode": true
}

Usage

The model is designed to output structured reasoning in the following format:

<reasoning>
[Step-by-step reasoning process in Arabic]
</reasoning>
<answer>
[Final answer in Arabic]
</answer>

Example Usage

from unsloth import FastLanguageModel

max_seq_length = 1024 # Can increase for longer reasoning traces
lora_rank = 64 # Larger rank = smarter, but slower


model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Omartificial-Intelligence-Space/Diraya-3B-Instruct-Ar",
    max_seq_length = max_seq_length,
    load_in_4bit = True, # False for LoRA 16bit
    fast_inference = True, # Enable vLLM fast inference
    max_lora_rank = lora_rank,
)


# System prompt to enforce XML structure
system_prompt = """
Respond in the following format in Arabic language only:
<reasoning>
...
</reasoning>
<answer>
...
</answer>
"""

# Prepare user question
user_question = "كل يوم، تُطعم وندي كل دجاجة من دجاجاتها ثلاث أكواب من العلف المختلط. تقدم الدجاجات وجباتهم في ثلاث وجبات منفصلة. في الصباح، تعطي قطيعها من الدجاج 15 كوبًا من العلف. في فترة ما بعد الظهر، تعطي دجاجاتها 25 كوبًا أخرى من العلف. كم عدد أكواب العلف التي تحتاجها لتقديمها لدجاجاتها في الوجبة الأخيرة من اليوم إذا كان حجم قطيع وندي 20 دجاجة؟"

# Prepare input for the model
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_question}
]
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Generate response
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.95
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Data

This model was primarily fine-tuned on:

• Arabic GSM8K Dataset :
• A comprehensive collection of grade school math problems translated to Arabic, requiring multi-step reasoning

Training and Evaluation Results

Training Progress

Figure: Composition of Rewards over Training Steps showing the evolution of model performance across different reward functions

The training utilized multiple reward functions to optimize different aspects of the model's performance:

• Correctness Reward (red): Measures the model's ability to generate the correct final answer
• Integer Reward (blue): Ensures the model outputs valid numerical responses
• Format Rewards (purple/gray): Promote proper XML structure adherence
• XML Count Reward (yellow): Fine-tunes the exact XML tag placement and completeness

As shown in the visualization, the model demonstrated consistent improvement across all reward dimensions throughout the training process. The upper values in rewards indicate higher-quality outputs that simultaneously satisfy multiple optimization criteria. This multi-objective training approach resulted in a model that not only produces correct answers but does so with clear, well-structured reasoning.

The model demonstrates strong performance on Arabic mathematical reasoning tasks, with particular strengths in:

• Producing well-structured reasoning steps
• Following the required XML output format
• Arriving at correct numerical answers for multi-step problems

Citation

If you use this model in your research, please cite:

@misc{diraya3b,
  title={Diraya-3B-Instruct-Ar: An Arabic Reasoning-Specialized Language Model},
  author={Omartificial-Intelligence-Space},
  year={2025},
  howpublished={\url{https://huggingface.co/Omartificial-Intelligence-Space/Diraya-3B-Instruct-Ar}}
}

Acknowledgements

This model builds upon the Qwen2.5-3B model by the Qwen Team and utilizes optimization techniques from Unsloth. We acknowledge their valuable contributions to the field of language modeling.

@misc{qwen2.5,
    title = {Qwen2.5: A Party of Foundation Models},
    url = {https://qwenlm.github.io/blog/qwen2.5/},
    author = {Qwen Team},
    month = {September},
    year = {2024}
}

@article{qwen2,
      title={Qwen2 Technical Report}, 
      author={An Yang and Baosong Yang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Zhou and Chengpeng Li and Chengyuan Li and Dayiheng Liu and Fei Huang and Guanting Dong and Haoran Wei and Huan Lin and Jialong Tang and Jialin Wang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Ma and Jin Xu and Jingren Zhou and Jinze Bai and Jinzheng He and Junyang Lin and Kai Dang and Keming Lu and Keqin Chen and Kexin Yang and Mei Li and Mingfeng Xue and Na Ni and Pei Zhang and Peng Wang and Ru Peng and Rui Men and Ruize Gao and Runji Lin and Shijie Wang and Shuai Bai and Sinan Tan and Tianhang Zhu and Tianhao Li and Tianyu Liu and Wenbin Ge and Xiaodong Deng and Xiaohuan Zhou and Xingzhang Ren and Xinyu Zhang and Xipin Wei and Xuancheng Ren and Yang Fan and Yang Yao and Yichang Zhang and Yu Wan and Yunfei Chu and Yuqiong Liu and Zeyu Cui and Zhenru Zhang and Zhihao Fan},
      journal={arXiv preprint arXiv:2407.10671},
      year={2024}
}