rlhn/remove-680K

Remove 680K数据集是级联LLM框架RLHN的一部分，用于准确重标记现有IR/RAG训练数据集中的难负样本。该数据集包含原始查询、正例和经过清理的硬负例，每个训练对中删除了一个假负例。数据集可用于微调嵌入、ColBERT或多元向量以及重排模型。The Remove 680K dataset is part of the cascading LLM framework RLHN, designed to accurately relabel hard negatives in existing IR/RAG training datasets. It contains the original queries, positives, and cleaned hard negatives, with one false negative removed from each training pair. The dataset can be used to fine-tune embeddings, ColBERT, or multi-vector reranker models.