SWE-bench Verified
SWE-bench Verified 是 SWE-bench 的 500 题人工校验子集,目标是剔除那些离开 GitHub issue 附加上下文就无法完成的题目,从而更稳定地衡量 agentic coding 系统解决真实软件工程问题的能力。
- 该 benchmark 使用真实开源 Python 仓库中的 GitHub issue,而不是面试题或竞赛题。
- 评分依据是关闭原始 issue 的 PR 所对应的真实单元测试。
- 文章强调它衡量的是完整 agent,包括模型、提示、工具和交互循环,而非单独模型。
- 在《[llm-wiki/raw/anthropic/Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet](/raw/anthropic/Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet.md)》中,Anthropic 用它来展示 Claude 3.5 Sonnet + Agent scaffold 的系统效果。
- 文章还指出这个 benchmark 的运行成本高,且容易受到环境配置、隐藏测试和评分噪声影响。
ℹ️ Conflict:
- 不同论文或团队汇报的分数若使用不同 scaffold、运行设置或修复后的 harness,横向比较时需要谨慎。
- [llm-wiki/raw/anthropic/Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet](/raw/anthropic/Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet.md)
- Claude 3.5 Sonnet
- Anthropic
- Agent scaffold
- Agentic coding evals
- Anthropic Engineering, “Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet”, 2025-01-06: https://www.anthropic.com/engineering/swe-bench-sonnet
- 原文摘录:[llm-wiki/raw/anthropic/Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet](/raw/anthropic/Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet.md)