(논문 요약) SWE-LEGO; PUSHING THE LIMITS OF SUPERVISED FINE-TUNING FOR SOFTWARE ISSUE RESOLVING

(논문 요약) SWE-LEGO: PUSHING THE LIMITS OF SUPERVISED FINE-TUNING FOR SOFTWARE ISSUE RESOLVING (Paper)

핵심 내용

SWE-Lego dataset:32k high quality task instances and 18k validated trajectories (real+synthetic data)
refined SFT
- error masking: tool call 에서 에러가 난 경우 (Incorrect Implementation, Localization Error) call 부분 mask
- difficulty-based curriculum