LLaMA-3.2-3B-GRPO-GSM325

🚀 LLaMA-3.2-3B-GRPO-GSM325 is a fine-tuned version of LLaMA 3.2B, trained using GRPO (Guided Reinforcement Policy Optimization) and DeepSeek R1’s open-source recipe. This model significantly enhances the base LLaMA-3.2-3B in mathematical problem-solving, logical reasoning, and structured response generation, pushing it towards GPT-4o1-style advanced reasoning.

🔥 Trained entirely on a Free Google Colab Tesla T4 GPU: Training Notebook

🚀 With more resources and extended training, this model could be pushed even further!

Model Details

• Base Model: LLaMA 3.2B
• Fine-tuning Method: GRPO with structured reinforcement
• Dataset: 325 curated questions from GSM8K (math reasoning)
• Format Adherence: XML-based structured reasoning
• Notable Improvements:
  • Mathematical accuracy ✔
  • Logical consistency ✔
  • Structured XML-format responses ✔
  • GPT-4o1-like step-by-step reasoning ✔

---

Usage

Example Input & Output

Input (User Query)

If 2x+5=10. Solve for x.

Output (Model Response)

<reasoning>
To solve for x, we need to isolate x on one side of the equation. This can be done by subtracting 5 from both sides of the equation.
</reasoning>
<answer>
2x + 5 - 5 = 10 - 5,
2x = 5,
2x / 2 = 5 / 2,
x = 2.5
</answer>

---

Installation & Inference

Hugging Face Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Rauhan/llama-3.2-3B-GRPO-GSM325"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Using vLLM for Fast Inference

from vllm import LLM, SamplingParams

llm = LLM(model="Rauhan/llama-3.2-3B-GRPO-GSM325")
sampling_params = SamplingParams(temperature=0.7, max_tokens=256)

output = llm.generate(["<reasoning>\nA store sells apples...\n</reasoning>"], sampling_params)
print(output)

Limitations & Future Work

🚧 Limitations:

• Limited by small dataset size (325 questions)
• Training done on a single Free Google Colab Tesla T4 GPU
• Some long-form reasoning may need further fine-tuning

🚀 Future Improvements:

• Training on a larger dataset (more GSM8K questions + other logical reasoning datasets)
• Extending fine-tuning using DeepSeek R1’s full training pipeline
• Further quantization for faster and memory-efficient inference

---

License & Citation

This model is released under Apache 2.0 License. If you use this model in your research, please cite:

@misc{llama-3.2-3B-GRPO-GSM325,
  author = {Rauhan},
  title = {LLaMA-3.2-3B-GRPO-GSM325},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/Rauhan/llama-3.2-3B-GRPO-GSM325}
}

---

🚀 This model demonstrates how even small models can achieve great results with the right fine-tuning techniques! 🚀

---

About the Author

🔗 Portfolio & Contact Information:

• 🌍 Website: rauhanahmed.org
• 🏢 GitHub: github.com/rauhanAhmed
• 💼 LinkedIn: linkedin.com/in/rauhan-ahmed
• 🐦 Twitter (X): x.com/ahmed_rauh46040
• 📧 Email: [email protected]

Feel free to reach out for collaborations, AI research, or any inquiries! 🚀