Browser-only language model
TriTalk miniGPT
This is a toy GPT text model trained on 2,463 public posts from the TriTalk forum. The cleaned corpus contains 130,608 word tokens, and the page downloads the model weights once, then generates text entirely on your device.
Model snapshot
Preparing browser demo…
Completion
Version 0.0.1
The training happened off-line in Python using a small decoder-only transformer. The result was then exported into static weight files that JavaScript can load.
When you press generate, the browser tokenizes your prompt, runs the model layer by layer, samples the next word, appends it, and repeats.
No API calls are made during inference. After the model loads, the generation loop is local.
It's not very good is it?
This model is tiny by modern standards. It has a short context window, a word-level vocabulary, and very limited reasoning ability. It is just a proof of concept.
It also inherits the quirks of the training data. You should expect triathlon jargon, forum-like phrasing, repetition, and some odd completions involving Fossies, crocs, and the word "your" in weird places.