Zrodlo documentation
Zrodlo trains every weight of a decoder language model: pretraining from a random initialization and continued pretraining of an existing model. It writes the Transformers model directory that Ster then posttrains. It is one Rust binary, `zrodlo`, built on Candle, and runs without any other Wisent product.
Start here#
- Install Zrodlo: with Stado or from source, and what it needs.
- Quick start: a tokenizer, a corpus, a tiny GPT-2, a resumed run and a held-out score.
Reference#
- CLI reference: every command on its own page, with every flag and what it refuses.
- What Zrodlo reads and writes: configurations, corpora, model directories and checkpoints.
- Refusals: what each command refuses, and that a refusal leaves nothing behind.
Zrodlo Desktop#
Zrodlo Desktop is a native macOS window over the installed zrodlo. It reads the commands the binary lists (zrodlo --json commands list) and shows one form per command, the command line it will run, the progress a run writes and its JSON result or refusal. A flag that takes a path has Choose…, and a flag with a fixed set of values (--device, --precision, --decay-scope, --decay-shape) is a menu of exactly those values. Stop ends a running command (a stopped training run keeps the checkpoints it wrote, which checkpoint resume continues), and quitting the app stops every run. The first run that names a file in Desktop, Documents or Downloads makes macOS ask whether Zrodlo may read that folder; the run waits until it is answered. Install it with stado product install zrodlo --surface desktop.
After stado product update zrodlo --surface cli, Reload Commands (⌘R, or ↻ in the toolbar) reads the binary's commands again without quitting; each form keeps what was typed in a flag the binary still takes and drops a value it no longer accepts.