Generated by a 2.7M-parameter transformer, trained from scratch on public-domain poetry — Tagore, Naidu, Dickinson, Poe, and more. It runs on a small server now, so there's nothing to download; just start typing.
The model runs server-side — a few million parameters, milliseconds per request. No download, no setup, always ready.
Embeddings, attention, softmax, sampling — identical math to the notebook that trained it, just running remotely instead of in your browser. The model architecture follows Andrej Karpathy's "Let's Build GPT" — a well established way to build a transformer.
Each request is generated fresh. Nothing you type is saved or logged beyond what's needed to return a response.
The poetry it learned from — Tagore, Naidu, Kabir, Dickinson, Poe, and more — comes from Project Gutenberg. Every book is in the public domain, royalty free, and free of copyright.