agentic-rl
33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
This repository is mostly documentation or write-up rather than runnable code.
Overview
Agentic RL is a beginner-friendly Chinese tutorial in 33 Markdown chapters on training LLMs into agents that plan, use tools and interact over several turns, from basic concepts through SFT, RLHF and GRPO to hands-on training with TRL. It includes a minimal runnable training example.
Its link to Jev is a dedicated chapter on when not to use RL now that System One decision models exist, plus chapters 26 to 33 on building vertical decision models. A companion minimal-decision-bench runs a reproducible three-way comparison on the same schema and data: a keyword rules baseline, a 102M Chinese MacBERT encoder (the recommended route) and a Qwen3.5-0.8B LoRA as an LLM control. The README says encoder and LLM quality were statistically indistinguishable in its findings.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | For a bounded judgment task, should you use an external decision model, an encoder classifier, an LLM or RL training?(inferred) | — |
How it works
- Learn agentic RL concepts and GRPO
- Read the chapter on externalizing judgment to models like Jev
- Run the three-way decision bench (rules, encoder, LLM LoRA)
- Choose a route based on measured quality and cost
Strengths
- Reproducible comparison with a one-page findings summary
- Runnable minimal examples
- CI on the repository
Caveats
- Chinese-language content
- Jev itself is discussed rather than benchmarked in the bench
Getting started
Start from the chapter index; for conclusions read minimal-decision-bench/FINDINGS.md.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- notebooks / scripts (TRL)
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Agentic RL 中文零基础教程（33 章）：从概念到 GRPO 实战，含 TRL 最小可跑示例；26–33…
wh000wh000/awesome-jev-live
Frequently asked questions
How does it relate to Jev?
A chapter discusses when an external decision model like Jev makes RL unnecessary, and later chapters build vertical decision models.
What does the bench compare?
Keyword rules, a 102M MacBERT encoder and a Qwen3.5-0.8B LoRA on the same task.
Is it in English?
No, it is written in Chinese.
Related projects
Jev-compatible open System One engine with a Claude Code-style terminal REPL showing live probabilities, plus an experiment harness.
ProjectprototypePythonBilingual research report on what Jev is, how its claims compare with independent evidence, and the open-source alternatives.
ProjectresearchDatawhale's Chinese Jev course: notebooks on the three primitives, translated recipes, evaluation and local Laya fine-tuning.
ArticleproductionPython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Pre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPythonInstruction dataset of 119 rows compiled from TypeSafe's public Jev documentation, split by Choice, Noul and Score.
ProjectresearchPython