OSGym: Super-Scalable Distributed Data Engine for Generalizable Computer Agents
Data infrastructure and algorithms for training generalizable computer-using agents at scale.
- Result
- Research intern · Paper co-author · Only high school student in the lab
- Areas
- Computer-Using Agents, Data Engineering, Reinforcement Learning, Back-End Development
Contribution
Training data for agents that operate computers
At MIT CSAIL's OSGym Lab, I engineered data-generation, pruning, and cleaning algorithms for computer-using-agent datasets.
The work contributed to OSGym, a super-scalable distributed data engine designed to support more generalizable computer agents, and resulted in a co-authored paper.