Overview

This repo contains the training, test, and data generation infrastructure of the 1st place solution to the Game-Playing Strength of MCTS Variants Kaggle competition.

Reproducing the winning ensemble

The models used in the winning solution can be replicated by running

./TrainWinningEnsemble.sh

in the root of the repository. This will install dependencies, download training data, train the models, and save the models to a models subdirectory of the current working directory. You will need an NVIDIA GPU with at least 8 GB of VRAM and at least 16 GB of system RAM to run it.

Data generation/annotation

• Text generation models for this repo's GAVEL implementation are available at https://www.kaggle.com/datasets/jsday96/gavel-models
• The main entrypoint for running GAVEL is GAVEL/GenerateGames.py
• A utility for generating rulesets with ordinary instruction-tuned LLMs and few-shot prompting is located at DataGeneration/NewGameGeneration/GenerateNewGames.py
• Data annotation utilities are scattered throughout the DataGeneration, GAVEL, StartingPositionEvaluation, and Reannotation directories.

Starting position analysis

Source code for the utility which runs tree searches on game starting positions to compute additional game balance & search speed features is located at StartingPositionEvaluation/EclipseWorkspace/AnalyzeGameStartingPosition/src/analyzeGameStartingPosition/AnalyzeGameStartingPosition.java. It depends on the Ludii player.