🌟 Welcome to the HRM-Mini Repository

This repository provides a lightweight environment for exploring and testing the reasoning capabilities of HRM with minimal computational resources.

It implements a 1-step gradient approximation and a hierarchical reasoning architecture, making it suitable for general-purpose applications.

Depending on hardware performance and setup, training typically takes several minutes to a few hours.

🚀 Getting Started

Follow the steps below to quickly set up and launch HRM Mini on an Ubuntu 24 server. HRM Mini is also compatible with Windows and macOS.

🧩 Step 1: Create a Virtual Environment

Using Miniconda:

conda create -n mini python=3.12
conda activate mini

📦 Step 2: Install Dependencies

Make sure the requirements.txt file is in your project directory. Then, first install PyTorch by following the appropriate installation steps from PyTorch's installation guide, and install the remaining dependencies by running:

pip install -r requirements.txt

🧠 Step 3: Try Out HRM-Mini

cd HRM-Mini
jupyter notebook --ip=127.0.0.1 --port=8888

After launching, click the link that looks like this:

http://localhost:8888/?token=xxxxxx

Log in through your browser and open the following notebooks to begin your reasoning experiment:

1_prepare_data.ipynb - Shows the data preparation process, including how to augment data.
2_hrm_mini_train.ipynb - Demonstrates the training process with augmented data.
3_hrm_mini_test.ipynb - Allows you to test HRM-Mini's results with your own Sudoku puzzle.

💡 Tips

• Feel free to modify the notebooks to adjust parameters or experiment with your own datasets.
• The notebook runs smoothly on most modern laptops and desktops.

🧾 License & Acknowledgment

HRM-Mini is part of the Hierarchical Reasoning Model (HRM) project. If you find it useful, please ⭐ the repository and cite the related work.