Skip to content
dormouse

Start here

What this is

Section titled “What this is”

dormouse is a byte-level language-model trainer that fits on one workstation: a 16 GB consumer GPU plus host RAM. Every mechanism it has is judged by held-out bits-per-byte at a fixed step budget, and this site is the record of that judgement — including the parts where the judgement went against the code.

If you are here for one number, it is on the status page. If you are here because you are about to build on this repo, read the agent rulebook §1 first; it exists because breaking it cost a run, a week, or a claim. Where every document lives is the map of this site back to the files it is generated from.

The culture on display here is unusual and deliberate. This project publishes retractions next to the claims they retract, marks a number unverified when the gate that would verify it is not runnable, and refuses to call anything “bit-for-bit” without naming the external file the reference came from (ADR-0020). A knowledge base that hid that would be a marketing site, and the only thing a reader of a training log really needs is to know which lines to trust.