MemGym/memgym_publicreadme

MemGym-IR是一个多跳问答基准数据集，专为研究记忆系统在多跳问答中的应用而设计。数据集包含经过验证的实例和基准测试结果，涵盖了3跳、4跳和5/6跳的问题实例。数据生成过程包括三个阶段：增长（Grow）、制作（Craft）和校准（Calibrate），所有LLM调用均通过AWS Bedrock完成。数据集采用MIT许可证发布，使用、复制、修改和再分发需保留MIT版权和许可通知。

MemGym-IR is a multi-hop question-answering benchmark dataset designed for academic research on memory systems for multi-hop QA. It includes verified instances and benchmark results covering 3-hop, 4-hop, and 5/6-hop question instances. The data generation process involves three stages: Grow, Craft, and Calibrate, with all LLM calls made through AWS Bedrock. The dataset is released under the MIT License, allowing use, copying, modification, and redistribution for any purpose, provided the MIT copyright and permission notice are preserved in any redistribution.