LLM Lab — Train a Tiny LLM in Your Browser
Pick a corpus, watch a byte-level transformer train live, then sample from it — or load the sample James trained on this site's copy. Everything happens on your device.
This is a hands-on look at what language model training actually is: a small transformer — embedding, a few attention layers, a tied output head — trained with backpropagation in a web worker by a trainer written in plain TypeScript, with no ML framework involved. Sixty seconds gets it from noise to word fragments; the sample James trains overnight on this site’s approved copy shows where the same workflow ends up. Its output recites and remixes what it has seen — it is not the grounded JTB chatbot, so treat anything it says about James as wordplay, not fact.
Privacy by architecture: nothing you paste, fetch, train, or generate leaves your device. The only network requests are fetching this site’s sample model file or a text URL you explicitly ask for.
The lab
Five steps: corpus, configure, train, generate, keep
The interactive lab needs JavaScript — above, the page explains what it does and why nothing leaves your device. With JS on, five cards appear here: pick a corpus, size the model, train it live in a web worker, sample text from it, and download the model file.