JEV / LEARNING ROADMAP
Jev learning paths: beginner, Python experiments, and agents
Choose an entry by what you can already do. Finish one small artifact before adding another capability. You do not need to complete every community project.
01 / 06
01 · Beginner: follow one ticket
- Before starting
- No code or account needed.
- Practice plan
- Read the ticket case. Separate customer claims from verified facts, state one decision goal, and write three distinct questions.
- Completion artifact
- A state → questions → actions sketch, including what still needs human verification.
Learn here first
Continue at the source
Next step → 02 · No code: choices and fallback
02 / 06
02 · No code: choices and fallback
- Before starting
- Distinguish state from questions. No API key needed.
- Practice plan
- Complete the four-step practice. Change one fact and reconsider the primitive. For missing evidence, specify what to collect first.
- Completion artifact
- A request draft with an unknown or human-review branch. Practice here uses synthetic examples; follow Playground requirements for live runs.
Learn here first
Continue at the source
Next step → 03 · Python: run an offline project
03 / 06
03 · Python: run an offline project
- Before starting
- Know virtual environments, dependencies, and JSON. Datawhale requires Python 3.10+; check each project README.
- Practice plan
- Start with a support-router fixture and inspect request and response fields. Add empty and conflicting inputs, verify fallback, then consider live mode.
- Completion artifact
- A rerunnable local script and three test records labeled fixture or live.
Learn here first
Continue at the source
Next step → 04 · Product work: evaluate failure costs
04 / 06
04 · Product work: evaluate failure costs
- Before starting
- Have a fixed task and runnable request; be able to label expected outcomes.
- Practice plan
- Learn threshold and coverage tradeoffs with synthetic examples, then evaluate held-out data. Split results by language, label, and error cost; avoid tuning on the test set.
- Completion artifact
- An evaluation sheet covering errors, human fallback, costs, and latency definitions; retain failures.
Learn here first
Continue at the source
Next step → 05 · Agents and apps: inspect the feedback loop
05 / 06
05 · Agents and apps: inspect the feedback loop
- Before starting
- Know basic evaluation and distinguish permissions, legal actions, and model judgments.
- Practice plan
- Study tool gates, then choose one community project. Trace visible state, candidates, code checks, outcomes, and the next state; keep hidden referee information out of model inputs.
- Completion artifact
- A loop diagram and one failure trace identifying who detects failure and where execution stops or falls back.
Learn here first
Continue at the source
Next step → 06 · Research and local models: a separate experiment
06 / 06
06 · Research and local models: a separate experiment
- Before starting
- Be able to fix baselines, split data, and track versions. Check model cards for training hardware and licensing.
- Practice plan
- Read the research question and controls before reproducing. Laya is a separate open-weight route; evaluate on the same tasks and splits, separating pseudo-label agreement from human-labeled performance.
- Completion artifact
- An experiment comparison sheet: sources, splits, model hashes, metrics, and limits. Unreproduced results remain source reports.
Learn here first
Continue at the source
Next step → Jev learning resources: four sites mapped
Connecting the eleven Datawhale chapters
A task map, not a reproduction. Original explanations and exercises live here; notebooks, engineering code, and reports remain in the source repository. The Chinese online documentation is a separate reference.
| Source chapter | Start here | Original source |
|---|---|---|
| 01 · Introduction | 01 · Beginner: follow one ticket | Datawhale ↗ |
| 02 · Core concepts | 02 · No code: choices and fallback | Datawhale ↗ |
| 03 · Architecture patterns | 03 · Python: run an offline project | Datawhale ↗ |
| 04 · 18 recipes | From concepts to working designs | Datawhale ↗ |
| 05 · Smart-home experiment | 05 · Agents and apps: inspect the feedback loop | Datawhale ↗ |
| 06 · Model evaluation | 04 · Product work: evaluate failure costs | Datawhale ↗ |
| 07 · 12 app projects | Which of the twelve projects should you read? | Datawhale ↗ |
| 08 · Research | 06 · Research and local models: a separate experiment | Datawhale ↗ |
| 09 · Agent integration | 05 · Agents and apps: inspect the feedback loop | Datawhale ↗ |
| 10 · Local Laya models | 06 · Research and local models: a separate experiment | Datawhale ↗ |
| 11 · Source and version index | Three checks before using a resource | Datawhale ↗ |
Which of the twelve projects should you read?
Pick one to study state, decisions, execution, and feedback. Links open source code directories; check each README for modes and dependencies. Jev Games Web is the shared presentation entry.
| Project | What to study | Source code |
|---|---|---|
| 01 · Gridloop / Snake | Start: local observation and collision checks | Gridloop / Snake ↗ |
| 02 · Minesweeper | Start: let code resolve rule-determined moves | Minesweeper ↗ |
| 03 · Werewolf | Later: hidden roles and evidence sources | Werewolf ↗ |
| 04 · Jev Games Web | Presentation: navigation and traces, not another strategy | Jev Games Web ↗ |
| 05 · Maze | Start: local decisions and planning code | Maze ↗ |
| 06 · Moving target | Later: distinguish action probability from hits | Moving target ↗ |
| 07 · Browser Use | Apps: verify goals independently after actions | Browser Use ↗ |
| 08 · Doudizhu | Later: hidden cards and legal candidates | Doudizhu ↗ |
| 09 · Blackjack | Later: compare with fixed rules and random conditions | Blackjack ↗ |
| 10 · Sudoku | Start: filter by constraints before choosing | Sudoku ↗ |
| 11 · Mario emulator | Later: frame timing and action duration affect feedback | Mario emulator ↗ |
| 12 · Smart home | Apps: intent, dispatch, and simulation modes | Smart home ↗ |
Three checks before using a resource
- Identify offline fixtures, archived replays, and real API modes before running.
- Record state, expected labels, actions, and failure causes, rather than only successful screenshots.
- Check source dates, model versions, and project licenses before adapting a product.
Learning-path questions
Can I learn without Python?
Yes. Start with understanding and no-key practice. Learn to design state, questions, and fallback before deciding whether to learn Python.
Does finishing a game prepare a production agent?
Games teach feedback loops. Production also needs permissions, argument validation, independent success checks, and evaluation on real tasks.
Can local Laya replace Jev directly?
Evaluate it separately. Similar input and output formats do not imply identical training, quality, calibration, or service behavior.