This project provides a pipeline for finetuning large language models (LLMs) specifically for GDScript generation, a domain-specific language for the Godot game engine. It addresses the performance degradation of general-purpose LLMs on less common languages by creating specialized models trained on human-written GDScript code scraped from GitHub. This benefits game developers using Godot by offering more accurate and consistent code generation.

How It Works

The core methodology involves creating a dataset of comment:code pairs derived exclusively from MIT-licensed GDScript projects on GitHub. Unlike other approaches that use LLMs for dataset generation, this project uses LLMs (like GPT-3.5-turbo) solely for labeling human-written code snippets. This ensures the training data consists of authentic, human-created code, leading to more robust and syntactically correct GDScript generation.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Requires GitHub and OpenAI API keys for dataset generation.
• Finetuning requires significant GPU resources (e.g., multiple A100 80GB GPUs recommended for LLaMA models).
• Inference demo available via notebook: Google Colab Link

Highlighted Details

• Finetuned models demonstrate significantly higher consistency in GDScript syntax compared to GPT-4/3.5-turbo.
• Variants trained on code-specific base models can outperform general models on complex instructions.
• Dataset generation involves scraping GitHub, identifying GDScript files, splitting them into functions, and generating comments using an LLM.
• Pre-assembled dataset godot_dodo_4x_60k (approx. 60k rows) is available.

Maintenance & Community

• Project is authored by Markus Sobkowski.
• Acknowledgements mention MIT-licensed Godot projects and fluidstack.io for GPU instances.
• Citation details are provided for the project, LLaMA, and Stanford Alpaca.

Licensing & Compatibility

• Dataset is exclusively sourced from MIT-licensed GitHub repositories.
• No explicit license is mentioned for the repository's code itself, but the data sourcing implies a permissive approach.

Limitations & Caveats

The primary weakness identified is a loss in verbosity and context awareness, where models may assume external scope due to the nature of scraped code samples. This could be improved with a more sophisticated dataset construction.