transformer-explainer

Transformer Explainer is an interactive website that teaches people how GPT-style AI language models work by letting them watch a real model run in their browser. Rather than reading a static explanation, visitors type their own text and see, step by step, how each internal part of the model processes it and arrives at a prediction for the next word. The model running inside the page is GPT-2, an earlier and smaller version of the architecture behind modern large language models, which is compact enough to run entirely in a web browser without sending data to a server.

The tool was built by a research team at Georgia Tech and was presented as a poster at IEEE VIS 2024, a major conference on data visualization. It also has an accompanying academic paper on arXiv. The goal is educational: making the internals of transformer models visible and understandable to people who are curious but not necessarily working researchers.

For anyone who wants to run it locally rather than using the hosted version, the setup requires Node.js and NPM, the standard JavaScript development tools. After cloning the repository and running a few install commands, the visualization is available at a local address in your browser.

The same team has built similar interactive explainers for other types of AI models, including one for image-generating diffusion models, one for convolutional neural networks used in image recognition, and one for generative adversarial networks. Those are separate projects linked from the README. This repository covers only the transformer explainer. The software is released under the MIT license, meaning it can be freely used and adapted.