CodeDoes/OpenMementos

OpenMementos-228K是一个包含228,557个推理轨迹的数据集，这些轨迹带有块分割和压缩摘要（mementos）的注释，源自OpenThoughts-v3。该数据集旨在教导语言模型在长形式推理过程中管理自己的上下文，通过将推理分割成块并将每个块压缩成摘要（memento）。数据集包含两个子集：default用于训练就绪的数据，full包含中间管道输出。数据集还提供了关于块和摘要大小、压缩比以及数据来源的详细统计信息。数据集采用MIT许可证，适用于微调模型以进行memento风格的生成。

OpenMementos-228K is a dataset of 228,557 reasoning traces annotated with block segmentation and compressed summaries (mementos), derived from OpenThoughts-v3. The dataset is designed to teach language models to manage their own context during long-form reasoning by segmenting reasoning into blocks and compressing each block into a summary (memento). It includes two subsets: default for training-ready data and full with intermediate pipeline outputs. The dataset provides detailed statistics on block and summary sizes, compression ratios, and data sources. It is licensed under MIT and is intended for fine-tuning models for memento-style generation.