Docs

Start with getting started. It takes you from an empty folder to a server answering with an adapter.

1. Getting startedFrom an empty folder to a Jeff server answering with an adapter, in five steps.
2. When to use JeffFor decisions you make often and fast, a small model with an adapter beats a big one. For the hard few, hand off.
3. Request formatThe state, the three question types, and the rules for options that every adapter shares.
4. Python and TypeScript clientsCall a Jeff server from Python or TypeScript, choose an adapter by name, and handle errors.
5. Serving several adaptersOne Jeff server, one base model, any number of adapters, chosen per request by name.
6. Preparing requests in advancePut the unchanging parts first so the server can prepare them before the changing part arrives.
7. Data guidelinesA ten-point checklist for an adapter's data set, so it trains well on v1.2 today and carries over to v1.3 unchanged.
8. Training an adapterBuild clean data in Jeff's row format, train a LoRA adapter on one GPU, and measure it honestly. With the lessons from the nine official adapters.
9. Adapter kitHelper scripts that check your data for format errors, leaks and shortcuts, split it, mix in replay data and measure the adapter three ways.
10. Submitting an adapterAdd your adapter to JeffHub with a pull request: the manifest, the test set, the shortcut report and a data card.
11. Suggest an adapterDescribe an adapter you would like Jeff to have. Others upvote ideas, and the most wanted ones may be trained by the maintainers.
12. Roadmap: v1.3 long-term supportWhat the stable v1.3 base will change, what stays fixed for its whole life, and what it means for adapters you build now.