🤖 Nano R1 – A Lightweight Reasoning AI Model

Nano R1 is a compact AI project inspired by the architecture of R1, designed to perform basic reasoning tasks using reinforcement learning. This project demonstrates how a minimal model can be trained from scratch to perform decision-making and reasoning using a simplified environment.

---

🧠 Core Idea

The goal of Nano R1 is to explore how reinforcement learning agents can learn logical or step-by-step reasoning behaviors, even with limited resources and a simplified architecture.

---

🚀 Features

• 📚 Trains on GSM8K (Grade School Math 8K) dataset
• 🧠 Trained using Reinforcement Learning (RL)
• 🧮 Performs step-by-step reasoning to solve math problems
• 🔬 Compact model size suitable for low-resource training
• 📈 Basic reward feedback loop for learning
• 📦 Modular code structure for ease of experimentation

---

📚 Tech Stack

• 🐍 Python
• 🧠 TensorFlow / PyTorch (whichever you used)
• 📊 NumPy
• 🤖 Hugging Face Datasets

---

base_model: unsloth/Qwen2.5-3B-Instruct-unsloth-bnb-4bit library_name: peft license: apache-2.0 pipeline_tag: text-generation language:

• en tags:
• unsloth
• grpo
• trl
• transformers
• qwen2.5
• text-generation-inference
• PyTorch
• gsm8k

---

Model Description

• Developed by: Jeesan Abbas
• License: Apache license 2.0
• Finetuned from model: unsloth/Qwen2.5-3B-Instruct-unsloth-bnb-4bit

🙋‍♂️ Author

Made with ❤️ by Jeesan Abbas