Start with seeds
Author tasks and choose the harnesses and models that generate and judge them.
TRAINING DATA / GITHUB
Better training data starts with the whole trajectory.
Author the tasks. Capture the work. Judge the result. Turn accepted model sessions into training data.
THE PROJECT, IN PRACTICE
What you can do with it.
Keep the development loop
Capture model trajectories, including tool interactions, corrections, and the steps that led to the result.
Judge before publishing
Independent review accepts or rejects traces. Rejected work can be retried before it enters the prepared dataset.
Keep the queues moving
Seed authoring, tracing, judging, formatting, privacy checks, local storage, and Hugging Face publication run as independent queues.
FROM IDEA TO USE
How it works.
- 01
Configure the roles
Choose author and judge models, providers, project storage, and publication preferences.
- 02
Generate and review
Run enabled queues and inspect accepted traces, rejections, retries, and throughput.
- 03
Prepare and publish
Format accepted trajectories for training and publish to your configured Hugging Face dataset.
python3 -m pip install moonshiner
moonshiner
# Inspect the active project
moonshiner status
moonshiner doctorThe first run opens setup. Model providers may require credentials and incur their own usage costs.
TAKE IT FOR A SPIN
Start here.
Install Moonshiner, run the setup assistant, and select your author and judge roles. Start with a small project and inspect the accepted traces before scaling up.
STATUS & REQUIREMENTS
Pi and Codex can author traces and seeds. Claude Code is supported as a judge. Provider access, credentials, and upstream terms depend on your selected configuration.