Researchers have introduced SWE-Prime, a novel two-stage method for selecting data to fine-tune large language models for software engineering tasks. This approach filters training data at both the trajectory and segment levels to improve the quality of supervision and mitigate undesirable behaviors. Experiments on SWE-Bench Pro and SWE-Bench Verified demonstrate that SWE-Prime's curated 10% subset of trajectories can lead to significant performance gains compared to using the full resolved dataset. AI
IMPACT This method could lead to more efficient and effective training of LLMs for complex software engineering tasks, potentially improving AI-driven code generation and debugging.
RANK_REASON The cluster contains a research paper detailing a new method for improving LLM performance. [lever_c_demoted from research: ic=1 ai=1.0]
- agent trajectory datasets
- arXiv
- Hugging Face
- large-language models
- supervised fine-tuning
- SWE Bench Pro
- SWE-bench Verified
- SWE-Prime
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
