v0.2

Second release. Same deal as 0.1: open index.html, train a tiny n-gram model in the browser, no server.

You can now train on Romeo and Juliet or the first chapters of Dracula without hunting down a .txt file. VT323 is self-hosted, so the typeface still loads offline. A few real bugs got fixed along the way: next-word sampling was accidentally deterministic, the explanation view highlighted the prompt as if the model had written it, and training could insert undefined tokens.

The UI is new. There is a short intro, then you pick a text, write a prompt, and read the output on the same page. Dark mode follows the OS.

v0.1

Initial version, working with a simple n-grams with laplace smoothing. No external libraries used for the model generation, it's all in ~600 JS lines. No web server required, just open index.html in a modern Web Browser.