Yujie Qian
San Francisco, California, United States
Summary
🤩
Rockstar🎓
Top SchoolYujie Qian is an NLP and foundation-model researcher with 11 years of experience building pretraining and retrieval systems, currently a Member of Technical Staff at Thinking Machines Lab in San Francisco. He holds an MIT PhD and has driven research-to-product work across academia and industry, including founding research scientist work at Voyage AI where he developed embeddings and rerankers for retrieval. His contributions span large-scale pretraining (GLM) and applied low-resource NLP, with notable open-source work fine-tuning SuperGLUE tasks and implementing a fast cloze model in the widely used THUDM/GLM repository. A repeat Google intern and Tsinghua collaborator, he combines deep research (CSAIL, ML for chemistry and structured documents) with hands-on engineering that optimizes model training and inference. Colleagues describe him as someone who routinely bridges cutting-edge papers and practical system tweaks—tuning hyperparameters and processors that materially improve downstream task performance.
11 years of coding experience
9 years of employment as a software developer
Doctor of Philosophy - PhD, Computer Science, Doctor of Philosophy - PhD, Computer Science at Massachusetts Institute of TechnologyBachelor of Engineering - BEng, Computer Science, Bachelor of Engineering - BEng, Computer Science at Tsinghua UniversityChangzhou Senior High School
Chinese, English
Github Skills (5)
pytorch10
machine-learning10
nlp10
python10
configuration-management9
Programming languages (4)
Github contributions (5)
THUDM/GLM
Jan 2021 - Jul 2021
GLM (General Language Model)
Role in this project:
ML Engineer
ML EngineerContributions:32 commits, 2 pushes, 2 comments in 5 months
Contributions summary:Yujie primarily contributed to the fine-tuning of configuration files for various SuperGlue tasks, including ReCoRD, MultiRC, BoolQ, COPA, RTE, WiC, and WSC. These changes involved modifying parameters like batch sizes, learning rates, and maximum sequence lengths. Furthermore, the user implemented a "fast cloze model" within the PVP framework, suggesting efforts to optimize the model's performance for masked language modeling tasks. The user also made changes to the WSC processor, which included negative sampling.
glmlanguage-modelmultimodaldialogue-systems
thomas0809/MolScribe
Mar 2021 - Jan 2023
Robust Molecular Structure Recognition with Image-to-Graph Generation
Contributions:1 review, 88 commits, 5 PRs in 1 year 9 months
chemistrydeep-learningmolecule
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.