~
Tip: ⌘P to print or save as PDF中文版本 →

Yu Ze

LLM Post-Training Engineer · Alibaba Quark Search

📧 TODO@example.com🌐 yuze.dev💻 github.com/houpanpan🔗 linkedin.com/in/TODO

Summary

AI/ML engineer with 7+ years of experience, focused on large language model post-training (SFT, RLHF) and multimodal alignment. Currently owning the post-training stack at Alibaba Quark Search end to end — from data production through training strategy, eval, and serving. Previously at ByteDance on multimodal LLM pretraining, with first-hand experience in scaling, training stability, and data quality.

Experience

Alibaba · Quark Search
2024.XX — present
LLM Post-Training Engineer · Senior
  • Own the SFT × RLHF stack for text and multimodal models in the Quark Search product: data → training → eval → deploy.
  • TODO: scale line — "training scale X-YB params / Zk SFT samples per month / DAU level".
  • TODO: business impact line — "rerank / answer-generation lifted internal metric X by Y pp (anonymized)".
  • TODO: engineering impact line — "rebuilt the post-training pipeline, cut iteration time from N to M days".
  • TODO: cross-team line — collaboration with data / inference / algo teams.
ByteDance · Multimodal LLMs
2022.XX — 2024.XX
Multimodal LLM Pretraining Engineer
  • Pretraining for multimodal LLMs: image-text data cleaning, training strategy, scaling-behaviour analysis.
  • TODO: subsystem you owned.
  • TODO: scale fact — tokens / GPU·hours / parameter count.
  • TODO: outcome line — improvement on a benchmark or downstream task.

Selected Projects

  • yuze.dev— This site. Next.js 16 + MDX + Tailwind v4, statically exported to GitHub Pages.

Skills

Languages / Frameworks
Python · PyTorch · Transformers · TRL · TypeScript (occasionally)
Training
Megatron-LM · DeepSpeed · Internal frameworks · FSDP
Inference / Serving
vLLM · SGLang · TensorRT-LLM
Post-training methods
SFT · PPO · DPO · GRPO · KTO · Reward Modeling
Data / Eval
LLM-as-judge · Custom eval pipelines · pandas / numpy · Data cleaning / dedup

Education

Beihang University (BUAA) · TODO college / major / degree
2015 — 2019