LLMCSA Benchmark Dataset
收藏资源简介:
该数据集由四川大学团队构建,是首个面向大语言模型代码溯源任务的大规模基准数据集,包含来自DeepSeek、Claude、Qwen和ChatGPT四种主流模型的91,804个代码样本,覆盖Python、Java、C和Go四种编程语言。数据通过受控生成流程采集,包含带注释与无注释两种编码设置,重点捕捉模型在代码风格、结构组织和词汇偏好等方面的生成指纹。该数据集旨在支持软件溯源、安全审计等场景下的模型级代码归属分析研究。
This dataset, constructed by a team from Sichuan University, represents the first large-scale benchmark dataset targeted at code provenance tasks for large language models (LLMs). It includes 91,804 code samples from four mainstream models, namely DeepSeek, Claude, Qwen, and ChatGPT, covering four programming languages: Python, Java, C, and Go. The data is collected through a controlled generation pipeline, with two coding configurations: annotated and unannotated, and it focuses on capturing the generative fingerprints of models in terms of code style, structural organization, and vocabulary preference. This dataset is intended to support research on model-level code attribution analysis in scenarios such as software provenance and security auditing.
DCAN数据集概述
数据集状态
- 数据集即将发布。
数据集内容
- 代码与数据集即将推出。
备注
- 数据集详情页面(https://github.com/mtt500/DCAN)显示,相关资源尚未公开。




