agentscope-ai/ReMe_longmemeval_clean_s_v2
收藏资源简介:
LongMemEval ReMe Cleaned-S 是 LongMemEval Cleaned-S 数据集的修正版本。它保留了原始问题和haystack会话,但将答案和支持会话的ground truth替换为来自final_groundtruth_cleaned_s.json的审核值。修正解决了不准确的答案和证据会话问题,包括证据发生在问题时间之后的情况,从而避免了未来信息泄露到评估中。数据集包含500条记录,涉及知识更新、多会话、单会话助手、单会话偏好、单会话用户和时间推理等多种问题类型。修正统计显示,15.4%的记录在修正字段上有所变化,包括答案和证据会话ID的调整。
LongMemEval ReMe Cleaned-S is a corrected version of the LongMemEval Cleaned-S dataset. It retains the original questions and haystack conversations, but replaces the answers and ground truth of the supporting conversations with reviewed values sourced from final_groundtruth_cleaned_s.json. The corrections address inaccurate answers and evidence conversations, including instances where the evidence occurred after the question's timestamp, thus preventing future information leakage during model evaluations. The dataset contains 500 records covering multiple problem types such as knowledge update, multi-turn conversations, single-turn assistant interactions, single-turn preference, single-turn user scenarios, and temporal reasoning. Correction statistics show that 15.4% of the records have changes in their corrected fields, including adjustments to answer and evidence conversation IDs.




