FuzzyBench
收藏资源简介:
FuzzyBench是由滑铁卢大学、康奈尔大学和哈佛大学的研究团队构建的一个大规模模糊函数数据集,旨在为训练程序即权重(PAW)范式提供支持。该数据集包含1000万条样本,每条样本由自然语言规范、输入文本和目标输出构成三元组,数据通过GPT-5.2模型分阶段生成,覆盖文本处理、搜索匹配、自定义分类、代码生成、安全验证等超过800个任务类别。数据集的创建过程采用两阶段流水线:首先生成模糊函数的自然语言规范,随后为每个规范生成输入输出对,并经过独立模型验证以确保质量。该数据集的核心应用是训练神经编译器,以解决传统符号规则难以处理的模糊任务(如日志分类、格式修复、语义搜索等),推动轻量级本地化人工智能系统的发展。
FuzzyBench is a large-scale fuzzy function dataset constructed by research teams from the University of Waterloo, Cornell University, and Harvard University, aimed at supporting the Program-as-Weight (PAW) paradigm. The dataset contains 10 million samples, each consisting of a triplet of natural language specification, input text, and target output. Generated in stages using the GPT-5.2 model, it covers over 800 task categories including text processing, search matching, custom classification, code generation, security verification, and more. The dataset was developed through a two-stage pipeline: first, natural language specifications for fuzzy functions are generated, then input-output pairs are created for each specification, and quality is validated via independent models to ensure reliability. The core application of this dataset is to train neural compilers to address ambiguous tasks that are difficult for traditional symbolic rules to handle, such as log classification, format repair, and semantic search, thereby advancing the development of lightweight localized artificial intelligence systems.





