THROWBENCH
收藏资源简介:
THROWBENCH是由Free University of Bozen-Bolzano和Univ. Bordeaux等机构的研究人员创建的一个数据集,包含超过2400个短小的人类编写的程序,这些程序使用四种不同的编程语言编写。大多数程序在运行时抛出异常(由于程序中的错误)。该数据集旨在评估大型语言模型是否能正确预测运行时程序行为,特别是预测程序是否抛出异常以及抛出哪种类型的异常。数据集中的程序是从RunBugRun数据集中筛选出来的,RunBugRun数据集是基于IBM的CodeNet数据集,包含了多种编程语言编写的编程竞赛提交作品。THROWBENCH数据集可用于评估和改进大型语言模型在代码理解方面的能力。
THROWBENCH is a dataset created by researchers from institutions including the Free University of Bozen-Bolzano and Univ. Bordeaux. It contains over 2,400 short, human-written programs developed in four distinct programming languages. Most of these programs throw runtime exceptions due to internal errors in the source code. This dataset is designed to evaluate whether large language models (LLMs) can correctly predict runtime program behaviors, specifically whether a program will throw an exception and what type of exception it will throw. The programs in THROWBENCH are filtered from the RunBugRun dataset, which is based on IBM's CodeNet dataset and consists of programming contest submissions written in a variety of programming languages. The THROWBENCH dataset can be used to evaluate and improve the code comprehension capabilities of large language models (LLMs).




