natfii/residual-walker

Residual Walker is a visualization tool that lets you watch, in real time, how a language model thinks its way through a sentence. Every layer of a transformer model updates its internal state by adding small corrections to it, once for attention and once for a small internal network called the MLP. This project captures that internal state at every one of those addition steps, squashes the high dimensional numbers down into three dimensions, and draws the result as a path moving through 3D space in your browser, one step at a time, until the model settles on the next word. You run a small language model locally through this tool rather than calling an outside service. A launcher script sets up a private Python environment, installs the right version of PyTorch for your machine, whether you have an NVIDIA graphics card or not, and lets you pick a model from a list. A Windows only prebuilt version needs no Python installed at all. As the path is drawn, different colors mark what kind of update just happened: white marks where a word's path begins, blue segments show information pulled in from other words in the sentence, and orange segments show facts or associations being recalled. At any point along the path, the tool can show what word the model would guess next if it stopped right there, so you can watch a model change its mind partway through, for example locking onto "Paris" while processing a sentence about the capital of France. The tool also lets you click on points along the path to pause and inspect them closely, showing which earlier words in the sentence were read from and how strongly. For a subset of supported models it can additionally load a more advanced lens, based on outside published research, that tries to guess what the model is silently planning to say several steps ahead of the word it actually outputs next. There is also a feature for nudging the internal state mid-generation to see how that changes what the model eventually says.