CodeSense
收藏资源简介:
CodeSense 是一个面向真实世界代码语义推理的基准和数据分析集,旨在评估代码大语言模型(LLMs)在实际软件工程(SE)任务中的能力。该数据集包含来自真实世界存储库的 Python、C 和 Java 软件项目,并通过执行测试、收集执行跟踪和构建细粒度语义推理任务的基准数据集来创建。CodeSense 数据集包括 2125 个 Python、876 个 C 和 875 个 Java 唯一函数,基于这些函数,我们整理了 4495 个样本及其基准数据。该数据集的创建过程包括收集真实世界的软件项目、开发自动提取基准任务的框架、构建和执行项目以及记录运行时执行值和跟踪。CodeSense 数据集的应用领域包括测试输入生成、漏洞检测、故障定位、代码修复、重构和功能验证等软件工程任务。
CodeSense is a benchmark and data analytics dataset for real-world code semantic reasoning, aiming to evaluate the capabilities of code large language models (LLMs) in practical software engineering (SE) tasks. This dataset contains Python, C, and Java software projects sourced from real-world repositories, and was constructed by executing tests, collecting execution traces, and developing benchmark datasets for fine-grained semantic reasoning tasks. The CodeSense dataset includes 2125 unique Python functions, 876 unique C functions, and 875 unique Java functions. Based on these functions, we curated 4495 samples along with their benchmark data. The creation process of the CodeSense dataset involves collecting real-world software projects, developing a framework for automatically extracting benchmark tasks, building and executing projects, as well as recording runtime execution values and traces. Application scenarios of the CodeSense dataset cover software engineering tasks such as test input generation, vulnerability detection, fault localization, code repair, refactoring, and functionality verification.
CodeSense 数据集概述
基本信息
- 数据集名称: CodeSense Evaluation Platform
- 目标领域: 代码语义理解与软件工程
- 支持语言: Python, C, Java
核心特点
- 评估维度: 细粒度代码推理能力
- 数据来源: 真实世界软件项目
- 评估方式: 提供执行轨迹和工具支持
应用场景
- 大型语言模型(LLMs)在软件工程实践中的能力评估
- 超越粗粒度输入/输出任务的深度语义理解




