unlearning-cleanslate/eval-23-debug-qwen3-8b-simnpo-gentle-bm25-6t-target-100-localtrain-checkpoint-1
收藏资源简介:
该数据集用于评估语言模型对文本的记忆化程度,包含每个文本样本的统计特征(如文本长度、窗口数、记忆化窗口数、记忆化分数、覆盖度、概率分布统计等)以及每个窗口的详细记忆化信息(如起始字符、结束字符、记忆化标志、对数概率、目标令牌的排名等)。此外,还包括评估模型、窗口大小、步长、阈值等参数,以及内容元数据(ID、标题、创建者、年份)。数据集仅包含训练集,共4663个样本。
This dataset is designed for evaluating the memorization degree of language models on text. It contains statistical features per text sample (e.g., text length, number of windows, memorized windows, memorized fraction, coverage, probability distribution statistics) and detailed window-level information (e.g., start/end characters, memorization flag, log probabilities, target token ranks). Additionally, it includes evaluation parameters (model, window size, stride, threshold) and metadata (content ID, title, creators, year). The dataset consists of a single training split with 4663 examples.




