Typing Tiny Stories

Description:

Typing Tiny Stores is a proof-of-concept web application that demonstrates the practical use of large language models (LLMs) running on local hardware in real-time.

How to Play:

1. Click the screen to begin the game.
2. Type the text as it appears. Incorrect text will be highlighted in red.
3. Use the backspace to correct any errors.
4. Lines end with a space; typing the space progresses you to the next line.
5. If you type enough words correctly, you will have the option to write your own single-line story. The LLM will then continue the story from where you left off.

AI Model:

This game serves as a proof-of-concept, showcasing the potential of focused LLMs for specific tasks, such as dynamic language systems for NPCs or validating dynamic user inputs. The model is optimized for speed, capable of generating a few hundred tokens per second on a mobile CPU. While it may not be the most robust model with extensive use, it is highly suitable for a typing game.

The LLM powering Typing Tiny Stores is trained on the TinyStories dataset (Eldan, R. and Li, Y., 2023), which consists of short stories with a limited vocabulary intended for three-year-olds. A smaller model, based on the llama2 architecture, was trained on this dataset, resulting in a lightweight model with approximately 300K parameters and a size of less than 1 MB. This makes it exceptionally fast and efficient for use in this game.