CodeGolf Bench
收藏资源简介:
CodeGolf Bench是由独立研究者Vedant Padwal创建的一个多语言代码生成基准数据集,旨在评估大型语言模型在60种编程语言中生成简洁高效代码的能力。该数据集基于代码高尔夫(code golf)竞赛平台,包含115个涵盖艺术、计算、游戏、数学、序列和转换等六类问题,每个问题均配有标准测试用例和实时人类解决方案基线。数据集的构建过程通过自动化API从code.golf平台获取问题陈述、语言支持和验证框架,确保了数据的动态更新与低污染风险。该数据集主要应用于评估LLM在资源受限环境(如物联网设备)下的代码优化能力,解决现有基准在语言覆盖度、静态问题集和人类基线动态性方面的局限性。
CodeGolf Bench is a multilingual code generation benchmark dataset created by independent researcher Vedant Padwal, aiming to evaluate the capability of large language models (LLMs) to generate concise and efficient code across 60 programming languages. Built upon the code golf competition platform, this dataset contains 115 problems spanning six categories: art, computing, games, mathematics, sequences, and transformations. Each problem is equipped with standard test cases and real-time human solution baselines. The dataset's construction process retrieves problem statements, language support information and validation frameworks from the code.golf platform via automated APIs, ensuring dynamic data updates and low contamination risk. This benchmark is mainly used to assess the code optimization ability of LLMs in resource-constrained environments such as IoT devices, addressing the limitations of existing benchmarks in terms of language coverage, static problem sets and the dynamism of human baselines.




