A dataset of 228,557 reasoning traces annotated with block segmentation and compressed summaries (mementos), derived from OpenThoughts-v3.
Memento is a framework for teaching language models to manage their own context during long-form reasoning. Instead of generating one long, unstructured chain-of-thought, memento-trained models segment their reasoning into blocks, compress each block into a dense summary (a memento), and continue reasoning from mementos alone.… See the full description on the dataset page:
https://huggingface.co/datasets/CodeDoes/OpenMementos.
