SWE-Prime: Fewer Trajectories, Better Performance

TL;DR - SWE-Prime is a two-stage data-selection method for supervised fine-tuning of software-engineering agents that filters successful trajectories by quality and representativeness, then selects useful segments for loss computation. Using only 10% of trajectories outperformed training on the full resolved dataset, showing that cleaner supervision can beat greater data volume.

• Screens trajectories using process quality, result quality, and dataset representativeness.
• Evaluates semantic step segments for solution contribution, learnability, and potential risks.
• Retains all segments as context during training but computes loss only on selected segments.
• Achieved relative gains of up to 12.2% on SWE-Bench Pro and 24.2% on SWE-Bench Verified.