SWE-Prime: Fewer Trajectories, Better Performance
SWE-Prime is a multi-granularity, two-stage SFT data selection method that filters training data at trajectory and segment levels. The first stage performs trajectory-level screening based on process quality, result quality, and data representativeness. The second stage performs segment-level selection by grouping consecutive steps into semantic segments and assessing each segment based on contribution to the final solution, learnability, and potential risks. During SFT, all segments remain in the sequence to preserve context.
SWE-Prime proposes a two-stage supervised fine-tuning data selection method to improve large language models' ability to resolve real-world software issues. It filters successful agent trajectories at trajectory and segment levels to reduce noisy supervision and encourage better problem-solving behaviors.
The method introduces multi-granularity filtering: trajectory-level screening for process quality, result quality, and representativeness, followed by segment-level selection based on contribution, learnability, and risk. Segments are kept in sequence during SFT to preserve context.
This approach addresses a key challenge in agent training: not all successful trajectories are equally instructive. By selecting higher-quality supervision data, it may reduce training cost and improve model reliability for software engineering tasks.
Improved data efficiency in fine-tuning could lower training costs and enhance performance of coding agents, benefiting companies building AI-powered software development tools.
Observable next signals include empirical results comparing SWE-Prime against baselines on software engineering benchmarks, adoption of similar data selection methods in agent training pipelines, and potential open-source releases of filtered datasets.