David-beakr/long-term-memory-seed
收藏资源简介:
这是一个用于评估AI助手(特别是Beakr系统)长期记忆能力的内部数据集,版本为v0.2。数据集包含7个时间线,每个时间线由一系列事件和一个探针问题组成,旨在测试长期记忆的五个核心能力:信息提取(IE)、多会话推理(MR)、知识更新(KU)、时间推理(TR)和弃权(ABS)。数据集基于LongMemEval(ICLR 2025)基准进行设计,但规模较小,适用于内部迭代和评估。它采用混合评分方法,结合确定性检查(如子字符串匹配)和LLM法官评分,以确保准确性和效率。数据集主要用于测试记忆层在实时学习事件序列时的表现,包括处理更正、冲突、时间引用、多会话合成和重复工作流等场景。数据集不包含固定文档检索测试,而是自包含的事件序列。
This is an internal dataset for evaluating the long-term memory capabilities of AI assistants (specifically the Beakr system), version v0.2. The dataset contains 7 timelines, each consisting of a sequence of events and a probe question, designed to test five core memory abilities: Information Extraction (IE), Multi-Session Reasoning (MR), Knowledge Updates (KU), Temporal Reasoning (TR), and Abstention (ABS). It is grounded in the LongMemEval (ICLR 2025) benchmark but is smaller in scale (7 timelines), suitable for internal iteration and evaluation. The dataset uses a hybrid scoring approach, combining deterministic checks (e.g., substring matching) with LLM judge scoring to ensure accuracy and efficiency. It focuses on testing how the memory layer learns from event sequences in real-time, including scenarios like corrections, conflicts, temporal references, multi-session synthesis, and repeated workflows. The dataset does not test retrieval over fixed documents; each timeline is self-contained.




