We study code intelligence.
What Makes Interaction Trajectories Effective for Training Terminal Agents?
SWE-Lego: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving