Grokking

An implementation of the OpenAI 'Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets' paper in PyTorch.

Installation

• Clone the repo and cd into it:

  git clone https://github.com/danielmamay/grokking.git
  cd grokking

• Use Python 3.9 or later:

  conda create -n grokking python=3.9
  conda activate grokking
  pip install -r requirements.txt

Usage

The project uses Weights & Biases to keep track of experiments. Run wandb login to use the online dashboard, or wandb offline to store the data on your local machine.

• To run a single experiment using the CLI:

  wandb login
  python grokking/cli.py

• To run a grid search using W&B Sweeps:

  wandb sweep sweep.yaml
  wandb agent {entity}/grokking/{sweep_id}

References

Code:

• openai/grok

Paper:

• Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets