unlearning-cleanslate/eval-olmo-3-7b-simnpo-baseline

该数据集用于评估语言模型记忆化（memorization）行为。包含文本长度、窗口数量、记忆化窗口数量及比例、覆盖率、概率统计量（p_z）等特征。每个样本包括多个窗口的详细信息，如窗口索引、起始字符、目标文本、对数概率、是否被记忆化等。还包含评估模型名称、窗口大小、步长、阈值以及内容元数据（ID、标题、创作者、年份）。适用于分析模型在训练数据中记忆特定片段的现象。This dataset is used to evaluate language model memorization behavior. It includes features such as text length, number of windows, memorized windows count and fraction, coverage, and p_z statistics (mean, median, min, max, std). Each sample contains detailed information for each window, including index, start character, target text, log probability, memorization flag, etc. It also includes evaluation model name, window size, stride, threshold, and content metadata (ID, title, creators, year). This dataset is suitable for analyzing the phenomenon of models memorizing specific segments from training data.