This repository provides the datasets essential for both training and evaluating MemAgent, our framework designed for long-context LLMs. The data is organized to facilitate various types of experiments, including main task evaluations, model training, and out-of-distribution (OOD) tasks.
The datasets are primarily derived from the HotpotQA dataset, enriched with synthetic long-context multi-hop question-answering data to push the… See the full description on the dataset page:
https://huggingface.co/datasets/BytedTsinghua-SIA/hotpotqa.
