Aim for hard-to-verify tasks.

Hey, I'm zhuzilin — an engineer driven by curiosity.

I'm particularly interested in the training dynamics of LLM post-training: how models evolve during this stage, and how we can build better systems and tools to observe, understand, and improve the process.

• Currently, I work on RL at @Z.ai and am one of the core maintainers of slime.
• Previously, I was fortunate to work hands-on across the LLM training stack at @WeChat — from pretraining and SFT to RLHF — while also building training infrastructure along the way.
• Before the LLM era, I worked on deep learning training frameworks and contributed to projects including TensorFlow and PyTorch.
• I'm also fascinated by JavaScript engines and runtimes. I once read through the ES5 specification and wrote es from scratch, and contributed bug fixes to bun in its early days.

My avatar is Frieren, from Frieren: Beyond Journey's End.

我是 zhuzilin，一个兴趣使然的工程师~

我尤其关注 大语言模型后训练动力学（training dynamics）：模型在这一阶段究竟是如何变化的，以及我们能否构建更好的系统和工具，去观测、理解并进一步改进这个过程。

• 目前，我在智谱的 RL 团队打工，也是 slime 的核心维护者之一。
• 在此之前，我很幸运地在 @WeChat 参与了从预训练到后训练的全流程，以及参与开发了支撑这些训练任务的基础设施和训练框架。
• 前大模型时代，我主要从事深度学习训练框架相关的工作，也曾为 TensorFlow、PyTorch 等项目贡献代码。
• 我对 JavaScript 引擎和运行时 也很感兴趣。读过 spec，写过解释器 es，还给早期的 Bun 修过一些 bug。

头像是 芙莉莲（Frieren），来自 《葬送的芙莉莲》。

btw, this is my cat~