TokenPrint: An Open-Source 3D Visual Debugger for LLMs and Transformers

TokenPrint is a new open-source project designed to demystify the internal operations of Large Language Models (LLMs) and transformers. By providing an interactive 3D visualization environment, it allows developers and researchers to trace the journey of a token as it moves through various layers of a model. Users can inspect embeddings, attention mechanisms, projections, and residual streams in real-time. The tool emphasizes transparency by categorizing data as real, derived, conceptual, or simulated, ensuring users understand the source of the information being visualized. Beyond simple observation, TokenPrint aims to serve as an educational and experimental platform, supporting tasks like activation analysis and model debugging. By moving away from static diagrams and toward spatial, explorable computation, the project offers a more intuitive way to understand how LLMs process information. The project is available on GitHub and can be tested via its web interface.