R1 reasoning traces
收藏资源简介:
R1推理链数据集是由约翰斯霍普金斯大学的研究团队创建的,包含了从MS MARCO数据集中提取的超过63.5万个查询和段落的R1推理链例子。该数据集用于训练能够利用测试时间计算的reranking模型RANK1,该模型在信息检索的reranking设置中展示了卓越的性能。数据集的内容来源于MS MARCO的积极示例、Tevatron的抽样否定示例、mT5-13B的硬否定示例等,经过仔细的质量过滤和混合,最终形成了用于训练的高质量数据集。该数据集的应用领域主要在于信息检索,旨在解决如何通过reranking提高检索相关性的问题。
The R1 Reasoning Chain Dataset was developed by a research team at Johns Hopkins University, containing over 635,000 R1 reasoning chain examples of queries and passages extracted from the MS MARCO dataset. This dataset is designed for training the RANK1 reranking model that leverages test-time computation, which has demonstrated exceptional performance in the information retrieval reranking setup. The dataset's content is sourced from positive examples from MS MARCO, sampled negative examples from Tevatron, hard negative examples from mT5-13B, among other sources. After rigorous quality filtering and mixing, it is finalized as a high-quality dataset for training purposes. The primary application domain of this dataset is information retrieval, aiming to solve the problem of improving retrieval relevance through reranking.

- 1Rank1: Test-Time Compute for Reranking in Information Retrieval约翰斯霍普金斯大学 · 2025年



