unlearning-cleanslate/not-exp-fsid-curated-gemma-12b-it-target-100

该数据集用于研究语言模型的记忆和遗忘行为，包含多个配置：forget和retain配置可能涉及模型对文本内容的记忆评估，其中forget配置包括请求ID、内容标题、前缀、后缀、记忆分数等特征，用于分析模型在特定规则下的遗忘情况；forget_pool配置包含歌词、创作者、年份等详细信息，用于池化分析记忆内容；retain配置包括文本和规则名称，用于评估模型保留的内容；retain_pool配置提供更详细的记忆和再生指标，如ROUGE-L分数、困惑度等，用于深入分析模型对文本窗口的记忆和生成能力。数据集分片包括baseline、bm25_10B、bm25_6T、igm_10B等，表示不同实验条件或模型版本下的数据。

This dataset is developed to study the memory and forgetting behaviors of language models, and it encompasses multiple configurations. The forget and retain configurations are designed for evaluating the model's memory of textual content: the forget configuration includes features such as request ID, content title, prefix, suffix, and memory score, which are used to analyze the model's forgetting performance under specific rules; the forget_pool configuration contains detailed information including lyrics, creator, and year, for pooling analysis of memorized content; the retain configuration includes text and rule name, aiming to evaluate the content retained by the model; the retain_pool configuration provides more detailed memory and regeneration metrics such as ROUGE-L score and perplexity, for in-depth analysis of the model's memory and generation capabilities towards text windows. The dataset splits include baseline, bm25_10B, bm25_6T, igm_10B, and other variants, representing data under different experimental conditions or model versions.