Top Repos· Jev-Based· Eval· Last updated 2026-09-21

cookiespiggy / agentic-rl

“Agentic RL 中文零基础教程（25 章）：从概念到 GRPO 实战，含 TRL 最小可跑示例。第 25 章讲清 Jev / TypeSafe System One 判别模型与 RL 的能力边界 | Chinese Agentic RL tutorial, 25 chapters + Jev-vs-RL boundary analysis”

Jev-Based rank #30In Eval #4

• 103Stars
• —Stars Δ / 1d
• —Stars Δ / 7d
• —Stars Δ / 30d
• 13Forks
• 0Watchers

GitHub ↗Official docs ↗Back to ranking

TL;DR

cookiespiggy/agentic-rl showed up as Jev / System One related work: Open replica or independent decision engine — not Jev, not endorsed.

The author describes it as: “Agentic RL 中文零基础教程（25 章）：从概念到 GRPO 实战，含 TRL 最小可跑示例。第 25 章讲清 Jev / TypeSafe System One 判别模型与 RL 的能力边界 | Chinese Agentic RL tutorial, 25 chapters + Jev-vs-RL boundary analysis” That is their claim, not a bench this site ran.

Who it’s for

• Agent / harness authors who need a typed gate, route, or prune step
• Anyone who needs the unofficial line: jev.pro does not issue keys and is not TypeSafe

What’s inside

Top-level entries on GitHub at generation time: .gitignore, 01-什么是Agentic-RL-概念与愿景.md, 02-基础知识-强化学习入门.md, 03-LLM与Post-Training基础.md, 04-从LLM-RL到Agentic-RL.md, 05-Agentic-RL的应用场景.md, 06-PPO算法详解.md, 07-DPO直接偏好优化.md, 08-GRPO深度解析.md, 09-RLVR可验证奖励的强化学习.md, 10-Reward-Shaping奖励设计进阶.md, 11-训练框架选型与搭建.md, 12-数据准备与SFT阶段.md, 13-训练环境搭建.md. Read those before you trust a one-line install.

How to use

1. Clone and follow the author’s Python toolchain. Do not invent a package name.

git clone https://github.com/cookiespiggy/agentic-rl.git
cd agentic-rl

2. If the project calls hosted Jev, you still need a TypeSafe console key or a documented gateway.
3. Send a complete state and typed questions. Code owns control flow after the probabilities.

Hosted Jev still looks like a function call — wrapper or not:

{
  model: "jev-1.13.0",
  state: { source: "cookiespiggy/agentic-rl", brief: "..." },
  questions: {
    gate: { type: "noul", instructions: "Should we act on this state?" },
    route: { type: "choice", instructions: "Pick one next system.", criteria: ["code", "review", "human"] },
  },
}

Try this

Put Jev on one real brief

Wire cookiespiggy/agentic-rl so one real state (ticket, diff, or utterance) becomes a Choice plus a Noul. Do not ask Jev to write the user-facing reply.

Schema-safe ≠ correct

On a case cookiespiggy/agentic-rl gets wrong, write down the schema-legal answer it returned and the human label. Do not “fix” it by asking Jev to generate an explanation.

FAQ

Does this replace the TypeSafe SDK?

Only if the author implemented a full client. Most harnesses still call the vendor SDK or raw HTTP.

Are the star chips live?

Star, fork, and watcher chips refresh from the public GitHub API via a scheduled job. The prose on this page is editorial.