This dataset contains the NIAH delayed-query erasing task used by KVEraser.
It is intended for the standalone KVEraser release and mirrors the paper
protocol for Qwen3-8B contexts from 1K to 32K tokens.
Samples are generated from the RULER-style needle-in-a-haystack task using
Paul Graham essay text as haystack material. Each prompt contains two matching
number-bearing needle sentences for one… See the full description on the dataset page:
https://huggingface.co/datasets/Graph-COM/KVEraser.
