Dataset Catalog

This dataset is used for the paper “Hyper‑multi‑step: The Truth Behind Difficult Long‑context Tasks”. It contains a series of long‑context retrieval tasks, divided into simple and difficult categories. Simple tasks include direct key‑to‑value and multi‑step key‑value retrieval; difficult tasks involve logic‑based retrieval, multi‑match retrieval, etc. Each task has a filename indicating the task type and context size. Columns include the full prompt, gold keys (correct answer keys), and gold values (correct answer values).