hoisoserious/formal-grammar-llm-benchmark

该数据集是一个基于概率上下文无关文法（PCFG）的形式语言基准，用于评估语言模型的记忆和泛化能力。它包含多种文法配置的语法序列（训练和测试）以及通过编辑距离和语法编辑构造的非语法序列。This dataset is a benchmark for formal languages based on probabilistic context-free grammars (PCFGs), designed to evaluate language models memorization and generalization capabilities. It includes grammatical sequences (train and test) from various grammar configurations, as well as non-grammatical sequences constructed via edit distance and grammar edits.