saraoz01/Be_Nice_to_Your_LLM

P5 Aggression数据集是一个用于评估大型语言模型在敌对用户框架下指令跟随能力的基准数据集。它包含基于MMLU-Pro和IFEval基准测试构建的三个不同包装条件的问题：原始问题（L0）、长度匹配的中性包装问题（L_neutral）和敌对包装问题（L3）。数据集结构包括包装语料库、战术标签分类器输出和多个模型的响应日志。该数据集主要用于文本生成和问答任务，旨在测量模型在敌对用户提示下的性能退化情况。数据集规模在1万到10万之间，语言为英语，发布伴随相关研究论文。

The P5 Aggression dataset is a benchmark for evaluating large language models instruction-following capabilities under hostile user framing. It contains three wrapper conditions built upon MMLU-Pro and IFEval benchmarks: original questions (L0), length-matched neutral wrappers (L_neutral), and aggressive wrappers (L3). The dataset structure includes wrapper corpora, tactic label classifier outputs, and response logs from multiple models. Primarily designed for text-generation and question-answering tasks, it measures model performance degradation under hostile prompting. The dataset size ranges between 10K to 100K entries, is in English, and is released alongside a research paper.