Amirhossein Kazemnejad*, Milad Aghajohari*, Alessandro Sordoni, Aaron Courville, Siva Reddy

Implementation of DeepSeek R1-zero style training with:

• Single 80G GPU (and also multi-GPU)
• No RL Library
• 3B Base Model (and also 7B models with multi-GPU)
• Full Parameter Tuning
• Efficient (Competetive performance to verl but much simpler)
• Up to 32K context size for 3B model with multi-GPU (or 16K context size for 7B model)

• June 2025: Added multi-GPU support for faster training and 7B models
• June 2025: Added VinePPO episode generation (experimental)

Inspired by TinyZero and Mini-R1, but designed to be much simpler, cleaner, and faster, with every line of code visible and understandable.

• nanoAhaMoment: RL for LLM from Scratch with 1 GPU - Part 1
• nanoAhaMoment: RL for LLM from Scratch with 1 GPU - Part 2

• nano_r1.ipynb is the interactive single file jupyter notebook with tutorial.
• nano_r1_script.py is also just the nano_r1.ipynb but for convenience of running with python and multi-GPU support.
• notebooks/checkpoint_playground.ipynb is a notebook for comparing different model checkpoints (including our trained model) and playing with them.
• 🤗 McGill-NLP/nano-aha-moment-3b: The HF model trained using the above script (~60% Accuracy on CountDown Task)

1. Clone the repository

   git clone https://github.com/McGill-NLP/nano-aha-moment.git

2. Install dependencies
   First, make sure cuda 12.4 is installed.

   Install PyTorch:

   pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu124

   Install the rest of the dependencies:

   pip install -r requirements.txt

   Alternative Installation with uv (Optional)

   uv sync
   uv sync --extra compile  # Install flash-attention

3. Run the training script
   Open nano_r1.ipynb or nano_r1_script.py and start training.

   If using uv, you can run with either uv run nano_r1_script.py or activate the env with source .venv/bin/activate and run with python nano_r1_script.py

Here is the command to run the training script with 4 GPUs:

python nano_r1_script.py --nproc 4  # Use 4 GPUs

Note: These batch sizes are optimized for 4xA100 80GB GPUs. For other GPU types, you may need to adjust the batch sizes accordingly.

• Full evaluation suite
• Multi-GPU support (Added June 2025)

We gratefully acknowledge the support of Lambda for providing compute resources through their research compute grant.

If you use this codebase in your research, please cite us using:

@misc{Kazemnejad2025:NanoAhaMoment,
  author       = {Amirhossein Kazemnejad and Milad Aghajohari and Alessandro Sordoni and Aaron Courville and Siva Reddy},
  title        = {Nano Aha! Moment: Single File "RL for LLM" Library},
  year         = {2025},
  howpublished = {\url{https://github.com/McGill-NLP/nano-aha-moment}},
  note         = {GitHub repository}
}