mteb/ConvoMem

ConvoMem是一个基于对话记忆检索的数据集，属于大规模文本嵌入基准（MTEB）。该数据集专注于从对话中检索证据，涵盖事实、偏好、弃权和用户状态变化等多个方面。任务类别为文本到文本检索，涉及社交和口语领域。数据集包含多个子配置，如弃权证据、助手事实证据、变化证据等，每个配置包括语料库、查询、相关度判断和排名靠前的文档。数据来源于LMEB，语言为英语，适用于评估嵌入模型在长时程记忆检索任务中的性能。ConvoMem is a dialogue-memory retrieval dataset part of the Massive Text Embedding Benchmark (MTEB). It focuses on evidence retrieval for facts, preferences, abstention, and changing user state from conversations. The task category is text-to-text retrieval, covering social and spoken domains. The dataset includes multiple configurations such as abstention_evidence, assistant_facts_evidence, changing_evidence, etc., each comprising corpus, queries, relevance judgments (qrels), and top-ranked documents. It is derived from LMEB, in English, and designed for evaluating embedding models on long-horizon memory retrieval tasks.