---
license: apache-2.0
tags:
- unsloth
- trl
- grpo
---
# ALLaM-Thinking: Arabic Large Language Model with Enhanced Reasoning Capabilities

[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-almaghrabima/ALLaM--Thinking-yellow)](https://huggingface.co/almaghrabima/ALLaM-Thinking)
[![Unsloth Optimized](https://img.shields.io/badge/Optimized%20with-Unsloth-green)](https://github.com/unslothai/unsloth)

## Overview

ALLaM-Thinking is an advanced Arabic Large Language Model specifically optimized for reasoning and mathematical problem-solving tasks. This model builds on state-of-the-art language model architecture and has been fine-tuned using the Unsloth library for improved performance and efficiency.

## Key Features

- **Arabic-First Design**: Built from the ground up to excel at understanding and generating high-quality Arabic text
- **Enhanced Reasoning**: Specialized in step-by-step problem solving, particularly for mathematical questions
- **Optimized Performance**: Accelerated using Unsloth for faster inference and reduced computational requirements
- **GRPO Implementation**: Utilizes Group Relative Policy Optimization for improved alignment

## Usage Example

```python
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("almaghrabima/ALLaM-Thinking")

# Initialize the model using vLLM
# Note: You should only initialize the model once, using vLLM directly
model = LLM(model="almaghrabima/ALLaM-Thinking")

# Format the prompt using chat template
text = tokenizer.apply_chat_template([
    {"role": "user", "content": "في فريق مكون من 15 لاعباً، 40% منهم يسجلون الأهداف. إذا سجل كل لاعب من اللاعبين الذين يسجلون الأهداف في المتوسط 5 أهداف خلال الموسم، فكم عدد الأهداف الكلي التي سجلها اللاعبون الذين يسجلون الأهداف؟"}
], tokenize=False, add_generation_prompt=True)

# Configure sampling parameters
sampling_params = SamplingParams(
    temperature=0.8,
    top_p=0.95,
    max_tokens=1024,
)

# Generate response
outputs = model.generate([text], sampling_params)
output = outputs[0].outputs[0].text
print(output)
```

## Answer

```
 أولاً، دعنا نجد عدد اللاعبين الذين يسجلون الأهداف.

40% من 15 لاعباً يساوي:

0.40 * 15 = 6 لاعبين

الآن، إذا كان كل لاعب من هؤلاء اللاعبين الستة يسجل في المتوسط 5 أهداف خلال الموسم، فإن إجمالي عدد الأهداف التي سجلها اللاعبون الذين يسجلون الأهداف سيكون:

6 لاعبين * 5 أهداف لكل لاعب = 30 هدفاً

لذلك، سجل اللاعبون الذين يسجلون الأهداف مجموع 30 هدفاً خلال الموسم. 
```

## Unsloth Optimization

This model has been optimized using [Unsloth](https://github.com/unslothai/unsloth), which provides significant speedups for training and inference.

## Training Details

ALLaM-Thinking was trained using a combination of techniques:

- Base architecture fine-tuned on diverse Arabic datasets
- GRPO (Group Relative Policy Optimization) for better alignment
- Specialized training on mathematical reasoning and step-by-step problem-solving

## Performance

ALLaM-Thinking demonstrates strong capabilities in:

- Mathematical problem-solving with step-by-step reasoning
- Logical analysis and deduction
- Maintaining coherence in long-form responses
- Domain-specific reasoning in technical fields

## Limitations

- Model outputs should always be verified by human experts, especially for critical applications
- May occasionally produce incorrect mathematical reasoning despite the step-by-step approach
- Limited context window compared to some larger models
- Performance may vary based on query complexity and domain specificity

## Citation

If you use ALLaM-Thinking in your research or applications, please cite:

```bibtex
@misc{almaghrabima2025allam,
  author = {Mohammed Al-Maghrabi Research},
  title = {ALLaM-Thinking: Arabic Large Language Model with Enhanced Reasoning Capabilities},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/almaghrabima/ALLaM-Thinking}}
}
```

## License

This model is released under the [Apache 2.0 License](https://opensource.org/licenses/Apache-2.0).