笔记NOTES
DRIFTING · 走神
做的时候 顺手记的 Write-ups from the build
项目里的设计笔记和踩坑记录，放不进简历一行字的那些。
Design notes and postmortems from things I've shipped — the parts that don't fit in a resume bullet.
DPO vs. GRPO：教小模型写「高效」SQL
DPO vs. GRPO: Teaching a Small Model to Write Efficient SQL
哪里翻车、哪里奏效，以及为什么要同时用偏好学习和执行感知的强化学习，Qwen3-4B 才不再写「对但慢」的查询。
What broke, what worked, and why we needed both preference learning and execution-aware RL to get Qwen3-4B to stop writing correct-but-slow queries.
阅读 →READ →为 LLM 数据清洗设计多模型路由 Pipeline
Designing a Multi-Model Routing Pipeline for LLM Data Curation
在阶跃星辰搭 GPT / Claude / Qwen 路由的笔记：异步执行、本地并行，还有一个改变了我们调用每个模型方式的 prompt 位置发现。
Notes from building a GPT/Claude/Qwen routing system at Stepfun — async execution, local parallelisation, and a prompt-placement finding that changed how we called every model.
阅读 →READ →