Salesforce/Libra
收藏资源简介:
Libra数据集是一个用于代码定位训练和评估的数据集,核心思想是“优化环境而非代理”。数据集包含13个配置,分为两类:Prompter Configs(12个仓库)和SWE-bench Evaluation Config。Prompter Configs是合成的代码定位问答对,由Prompter代理生成,每个实例包含一个问题陈述、黄金文件路径、黄金函数、推理跟踪、源代码块等。SWE-bench Evaluation Config是来自SWE-bench Lite的真实错误报告,用于评估Locator代理。数据集的生成过程详细描述了Prompter代理如何生成每个问答对。
The Libra Dataset is a dataset for training and evaluating code localization. Its core insight is to "optimize the environment, not the agent." The dataset contains 13 configs organized into two categories: Prompter Configs (12 repos) and SWE-bench Evaluation Config. Prompter Configs are synthetic code-localization Q/A pairs generated by the Prompter agent, each including a problem statement, gold files, gold functions, reasoning trace, chunk content, etc. The SWE-bench Evaluation Config consists of real bug reports from SWE-bench Lite, augmented with gold file and function labels for evaluating the Locator agent. The README also details how the Prompter agent generates each Q/A pair.




