Chinese Jev Cookbook
Datawhale's 11-chapter Chinese Jupyter course for Jev, with 18 recipes, a 3D smart home, and local Laya fine-tuning.
Overview
Jev Cookbook (github.com/datawhalechina/jev-cookbook) is Datawhale's 11-chapter Jupyter tutorial for Jev, written in Chinese. The README calls the course Jev 入门教程. You learn Choice, Score, and Noul, and the rule that one question asks one thing so code can combine the answers. Chapter 3 covers speculative fan-out, confidence gating, composite scoring, and intent routing. Chapter 4 replays 18 recipes from the official cookbook, including reranking, row-by-row semantic search, structure recovery, function calls, citation checks, and guardrails. Chapter 5 is a voice-driven 3D smart home. Chapter 6 sets up a Laya versus Jev run on 231 public questions and scores accuracy, a majority-class floor, Brier, and ECE, with no retries and no fallback. Chapter 7 has 12 runnable games and apps. Chapter 9 puts a judgment in front of the Pi agent and in front of DeepSeek Harness. Chapter 10 fine-tunes the open model Laya on your own machine with RLCD. Offline examples run without an API key. The site is an unofficial community translation of the TypeSafe docs. The README says the English original at docs.typesafe.ai is the authority. Original tutorial text is CC BY-NC-SA 4.0. The README names project lead 王熠明 (Bald0Wang). On 6 October 2026 the repo had 184 stars and 23 forks. It was created on 20 September 2026.
Problem: Chinese-language readers had almost no runnable path from a first Jev question to recipes, an agent gate, and a local fine-tune.
Built for: Developers and students who can use Python and Jupyter, want a Chinese Jev course, and do not need a machine-learning background.
First indexed on Jev Directory: 2026-10-06
Creator and team
- Name
- Datawhale
- Organization
- Datawhale
- Handle
- @datawhalechina
“适合中国宝宝的 Jev 入门教程”
How Jev is used
- Role in the product flow
- A teaching path: typed questions in notebooks, then the same shapes inside recipes, a smart home, agent gates, and a local Laya server
- Primitives
- ChoiceScoreNoul
- State in
- A state plus one typed question at a time. Choice, Score, and Noul are the 3 primitives. The course says to ask one thing per question and let code combine the answers.
- Decision out
- A structured answer with a probability and a confidence, not generated text you have to parse. Later chapters route, gate, or score from those answers.
- Read the site, or clone the repo and run main/setup_env.sh. The chapter guide asks for Python 3.10+ and Jupyter.
- Chapters 1 to 3 are the core path: what Jev is, 5 booklets on the mechanism, then the 4 architecture patterns.
- Set JEV_RUN_MODE=offline to walk the code on saved example responses. A TypeSafe key is optional and is required only for live calls.
- Chapter 4 runs the 18 recipe notebooks. Chapter 5 speaks to a 3D room. Chapter 7 runs 12 projects.
- Chapter 9 gates Pi tool calls and reviews a DeepSeek Harness session. Chapter 10 builds a Chinese dataset, fine-tunes Laya with RLCD, and serves it locally.
This repo teaches the call. It does not replace the English docs. The official cookbook listing on this directory points at docs.typesafe.ai, which the Datawhale README says is the authority. Ten levels of Jev (disler-ten-levels-of-jev) is a different course: 30 TypeScript examples and a Vue lab, in English. receptron-laya is the Node library that runs Laya weights, not a tutorial. y0usaf-pi-jev is a Pi extension you install. Chapter 9 here is a notebook about a Pi gate, plus a DeepSeek Harness plugin whose own README says the offline demo does not call TypeSafe. Chapter 6's notebook is built to score Laya and Jev on the same 231 public questions. The compare.md stored in the repo does not finish that job: Laya-local has no scorable rows. A separate note in the chapter 6 README, dated 26 September 2026, covers 5 other providers on the same question file. That note says none of them returned 231 valid answers, and it says not to treat the composite scores as a final ranking. This page does not repeat that table. Cite the Datawhale README, the chapter guides, and docs.typesafe.ai.
Sourced performance claims
- Root README: 11 runnable chapters, 18 official recipes replayed one by one, 12 runnable game and app projects, and a knowledge base it describes as 500+ files. Chapter 11's own guide says the snapshot has 21 sections. License file: original tutorial content is CC BY-NC-SA 4.0. Third-party projects keep their own licenses.Source: github.com/datawhalechina/jev-cookbook README and LICENSE
- Chapter 6: 231 public questions, 48 easy, 72 original, and 111 hard. The comparison is local Laya versus hosted Jev on accuracy, majority-class floor, Brier, and ECE, with no retries, no fallback, and a budget ledger.Source: github.com/datawhalechina/jev-cookbook chapter 6 notebook and README
- Checked-in compare.md for that notebook demo: Jev accuracy 0.9587 on 121 scorable rows, Brier 0.0582, ECE 0.0339, p50 0.79s, p95 1.11s, cost 0.0020 USD. Laya-local has 0 scorable rows and a blank accuracy. This file is not a finished 231-question head-to-head.Source: jev-cookbook main/06 chapter, benchmark/runs/notebook-demo/multi/compare.md
- Public GitHub repo datawhalechina/jev-cookbook had 184 stars and 23 forks on 2026-10-06. Created 2026-09-20. Language Python. Homepage https://datawhalechina.github.io/jev-cookbook/.Source: GitHub API 2026-10-06
Features and stack
Features
- 11 chapters in Jupyter, from a first multi-question call to a local fine-tune
- 5 core booklets: System One, state, primitives, confidence, and building with TypeSafe
- 18 recipe notebooks replayed from the official cookbook, with a note at the end of each
- Voice-driven 3D smart home lab with cost stats
- 12 projects: snake, minesweeper, werewolf, a web page that opens those games, a maze, a moving target, a browser agent, Dou Dizhu, blackjack, sudoku, a Mario repro, and the smart home
- Pi tool-call gate and a DeepSeek Harness (DSH) decision plugin
- Local Laya fine-tuning with RLCD, a Chinese dataset, and a local Jev-compatible server
- Offline examples that run without an API key
- Chapter 11 knowledge snapshot: unofficial Chinese docs plus community articles. English docs.typesafe.ai stays authoritative
Stack
- Python 3.10+
- Jupyter
- TypeSafe API for live calls
- Laya for the local model chapter
- Pi and DeepSeek Harness in chapter 9
- CC BY-NC-SA 4.0 for original tutorial text
Pricing: Original tutorial text is CC BY-NC-SA 4.0, so the license bars commercial use of that text. Offline notebooks do not need a key. Live calls spend your TypeSafe key. The DeepSeek Harness chapter can also use a DeepSeek key when you leave demo mode.
Links
FAQ
- Is this the official TypeSafe cookbook?
- No. Datawhale says the site is an unofficial community translation. The README says the English original at docs.typesafe.ai is the authority. The official cookbook listing on this directory is the English recipes. This page is the Chinese tutorial repo.
- Can I run the Jev Cookbook without an API key?
- Yes for the offline path. The chapter guide says JEV_RUN_MODE=offline uses saved example responses so you can learn the code path. Those examples are not evidence of accuracy or calibration. A TypeSafe key is optional and is only needed when you want the same code to call the live model.
- What did Laya score against Jev?
- The chapter 6 notebook is set up to compare local Laya with hosted Jev on 231 public questions and to report accuracy, a majority-class floor, Brier, and ECE. The compare.md checked into the repo shows Jev with 121 scorable rows (accuracy 0.9587, Brier 0.0582, ECE 0.0339) and Laya-local with 0 scorable rows. This listing does not treat that file as a finished head-to-head.
- What is DSH in chapter 9?
- The plugin README in apps/dsh-jev-decision says DSH is DeepSeek Harness. The plugin lets that harness call Jev for a structured judgment before it acts. Its offline demo runs without an API key. The chapter 9 README says a live run sends at most 4 requests.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev & TypeSafe
What Jev is, how System One fits your stack, and where this directory ends and TypeSafe docs begin.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.