Fanar-Math-R1-GRPO is a reasoning-optimized language model built on QCRI/Fanar-1-9B-Instruct. This version is fine-tuned using Group Relative Policy Optimization (GRPO) from the DeepSeekMath framework on the AI-MO/NuminaMath-TIR dataset. It is designed for step-by-step mathematical problem-solving with structured reasoning in both English and Arabic.
<think> and <answer> tagstransformers, peft, and math_verify)| Component | Description |
|---|---|
| Base Model | QCRI/Fanar-1-9B-Instruct |
| Fine-Tuning | GRPO via Hugging Face TRL |
| Dataset | AI-MO/NuminaMath-TIR |
| Format | <think> ... </think> <answer> ... </answer> tagged reasoning structure |
| LoRA | Enabled (modules: q_proj, v_proj, rank=8) |
| Epochs | 1 (lightweight test configuration) |
| Tokenizer | Same as base model |
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import time
model_id = "Omartificial-Intelligence-Space/Fanar-Math-R1-GRPO"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)
def generate_with_reasoning(prompt_text):
inputs = tokenizer(prompt_text, return_tensors="pt").to(model.device)
start = time.time()
with torch.no_grad():
output = model.generate(**inputs, max_length=1024)
end = time.time()
generated = tokenizer.decode(output[0], skip_special_tokens=True)
duration = end - start
num_input_tokens = inputs["input_ids"].shape[1]
num_generated_tokens = output.shape[1] - num_input_tokens
return generated, duration, num_generated_tokens
# Example Arabic math problem
prompt_text = '''ูู ู
ุฏููุฉ ูุจูุบ ุนุฏุฏ ุณูุงููุง 1 ู
ูููู ูุณู
ุฉุ ุฅุฐุง ูุงู 60% ู
ู ุงูุณูุงู ุจุงูุบููุ ู40% ู
ู ุงูุจุงูุบูู ูุนู
ูููุ ููู
ุนุฏุฏ ุงูุนุงู
ููู ูู ุงูู
ุฏููุฉุ'''
result, time_taken, tokens = generate_with_reasoning(prompt)
print(result)
math_verify<think> and <answer> tagstransformers==4.47.1
trl==0.14.0
peft==0.14.0
datasets==2.21.0
math_verify==0.3.3
torch==2.4.1
The model is trained to follow a reasoning-first format:
<think> ุฃููุงูุ ูุญุณุจ 60% ู
ู ู
ูููู ูุณู
ุฉุ ููู 600,000. ุซู
ูุญุณุจ 40% ู
ู ูุฐุง ุงูุนุฏุฏุ ููู 240,000. </think>
<answer> 240,000 </answer>
@article{zhihong2024deepseekmath,
title={DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models},
author={Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Zhang, Mingchuan and Li, Y.K. and Wu, Y. and Guo, Daya},
journal={arXiv preprint arXiv:2402.03300},
year={2024}
}
@misc{vonwerra2022trl,
title={TRL: Transformer Reinforcement Learning},
author={von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouรฉdec, Quentin},
year={2022},
howpublished={\url{https://github.com/huggingface/trl}}
}
@misc{fanarllm2025,
title={Fanar: An Arabic-Centric Multimodal Generative AI Platform},
author={Fanar Team and Ummar Abbas and Mohammad Shahmeer Ahmad and Firoj Alam and Enes Altinisik and Ehsannedin Asgari and Yazan Boshmaf and Sabri Boughorbel and Sanjay Chawla and Shammur Chowdhury and Fahim Dalvi and Kareem Darwish and Nadir Durrani and Mohamed Elfeky and Ahmed Elmagarmid and Mohamed Eltabakh and Masoomali Fatehkia and Anastasios Fragkopoulos and Maram Hasanain and Majd Hawasly and Mus'ab Husaini and Soon-Gyo Jung and Ji Kim Lucas and Walid Magdy and Safa Messaoud and Abubakr Mohamed and Tasnim Mohiuddin and Basel Mousi and Hamdy Mubarak and Ahmad Musleh and Zan Naeem and Mourad Ouzzani and Dorde Popovic and Amin Sadeghi and Husrev Taha Sencar and Mohammed Shinoy and Omar Sinan and Yifan Zhang and Ahmed Ali and Yassine El Kheir and Xiaosong Ma and Chaoyi Ruan}},
year={2025},
url={https://arxiv.org/abs/2501.13944},
}
Happy reasoning! ๐โจ