Gloopy
A tiny alien character language model trained from scratch, with chat inference running in the browser.
Technology stack
Python / PyTorch / ONNX Runtime Web / JavaScript
An 8.7-million-parameter character model trained on generated single-turn alien conversations, packaged with its tokenizer and browser chat interface.
The website runs inference on the visitor’s device without a chat backend. Training tools, dataset preparation, and command-line chat support the experiment.
Behind this project
Gloopy: Training a Tiny Language Model from Scratch
Trace next-token training, causal attention, and ONNX browser inference while examining how a small character model sets expectations and invites experimentation.