AI Dungeon is an open-ended text adventure game that allows users to create and experience infinite stories. It is targeted at players who enjoy interactive fiction and creative writing, offering a unique AI-powered narrative experience.

How It Works

The game leverages a fine-tuned GPT-2 language model to generate story content based on user inputs. The underlying data format involves transforming text adventures into a JSON tree structure, which is then converted into a large text file for model training. This approach allows the AI to learn narrative flow and generate coherent, contextually relevant story continuations.

Quick Start & Requirements

• Install locally: git clone --branch master https://github.com/AIDungeon/AIDungeon/, cd AIDungeon, ./install.sh, ./download_model.sh, source ./venv/bin/activate, ./play.py
• Prerequisites: NVIDIA GPU with 12GB+ VRAM and CUDA installed is recommended for faster turn times. Without it, turns can take minutes.
• Resources: Fine-tuning the largest GPT-2 model requires significant hardware, such as 8x 32GB V100 GPUs (e.g., AWS p3dn.24xlarge) or 8x 16GB V100 GPUs (e.g., SageMaker p3.24xlarge) with reduced batch sizes.
• Links: Website, Mobile App, Play Online, Colab Demo

Highlighted Details

• Offers online play, mobile apps, and a Colab notebook for accessibility.
• Supports custom model fine-tuning with detailed data formatting and hardware recommendations.
• Follows a Git Flow branching model with develop for contributions and master for stable releases.

Maintenance & Community

• Open-source project welcoming contributions via pull requests.
• Community resources include a website, email, Twitter, Reddit, and Discord.
• Development primarily occurs on the develop branch.

Licensing & Compatibility

• The README does not explicitly state a license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

Running the game locally without a high-end NVIDIA GPU significantly impacts performance, leading to multi-minute response times per turn. Fine-tuning the model requires substantial and specific cloud or on-premises GPU resources.