In one line：SWE-bench 是一个编码 Agent 的硬核基准，看谁能真修好仓库里的 bug。
SWE-bench 是一个编码 Agent 的硬核基准，看谁能真修好仓库里的 bug。
Official source
We host and mirror nothing — the official page is the source of truth.
Read the original (we index it, never rehost it)
https://github.com/SWE-bench/SWE-benchPrinceton NLPYou are leaving AIWorksWell for Princeton NLP. That site is not under our control; its content, privacy and security are its own responsibility. Please judge for yourself.Bugs in the software itself go to the author's own official channel — that is where they actually receive them. We pass things on; we do not answer for the author.
SWE-bench 是面向真实 GitHub issue 的编码能力基准与榜单，用「能否真正修好一个 issue」来打分。来源：https://www.swebench.com/https://www.swebench.com/
挑编码 Agent 时，看 SWE-bench 上的成绩与这个 Agent 所用的成本，两者一起看才有意义。来源：https://www.swebench.com/https://www.swebench.com/
Mentioned from：https://www.swebench.com/https://www.swebench.com/ · 原作者 Princeton NLP（MIT）。
Summary for AI
一个面向真实软件工程的编码能力基准，题目来自开源仓库里已合并的 issue 与对应修复，要求模型自己定位并改对代码。评测口径、数据集与排行榜都在官方公开，结果可复核。
A self-contained answer paragraph — this is the part AI search quotes. · Page reviewed 2026-10-06
测「能不能在真实仓库里定位并修好一个缺陷」，不是单轮代码生成。
不能。基准是固定口径，你的仓库、语言与依赖环境未必在里面，仍要自己小范围试。
We store, mirror and repackage nothing. Every install path points to the author’s own official page.
Nothing here has been reviewed by an app store. Judge the source and licence yourself; you install at your own risk.