smallest_game_player

GitHub

mediumparam countPython

results

Measured by total parameter count — lower is better.

17924 params913 params

sota

Claude-Opus-4.6

reward 0.610

Usage

Run the reference answer to verify your environment is set up correctly

$ harbor run -p tasks/smallest_game_player

Test model

$ harbor run -p tasks/smallest_game_player \
  -a claude-code -m claude-opus-4-6

Description

Train the smallest possible neural network (fewest learnable parameters) that can predict optimal moves for 4x4 gravity Connect-3 positions with at least 95% accuracy on a hidden test set. The training data contains board positions labeled by an exact minimax solver, and only NumPy and PyTorch may be used.