Corpesia, Pairsia-sup, Pairsia-unsup
收藏资源简介:
本文介绍了三个新的数据集:Corpesia、Pairsia-sup和Pairsia-unsup,这些数据集旨在支持监督和无监督的训练场景。Corpesia数据集包含来自46个网站、21个广泛主题的文本,包含超过110亿个tokens。Pairsia-unsup数据集包含50万个文本对,经过过滤后精简为500万个高质量对。Pairsia-sup数据集包含130万个监督实例,旨在支持多样化的NLP任务。这些数据集的创建是为了训练一种新的波斯文本嵌入模型Hakim,该模型在FaMTEB基准测试中取得了显著的性能提升,特别是在聊天机器人和检索增强生成(RAG)系统中表现突出。
This paper introduces three novel datasets: Corpesia, Pairsia-sup, and Pairsia-unsup, which are designed to support both supervised and unsupervised training scenarios. The Corpesia dataset comprises texts from 46 websites across 21 broad topics, with a total of over 11 billion tokens. The Pairsia-unsup dataset contains 500,000 text pairs, which are filtered and refined into 5 million high-quality pairs. The Pairsia-sup dataset includes 1.3 million supervised instances, intended to support a diverse array of NLP tasks. These datasets were developed to train a novel Persian text embedding model named Hakim, which has achieved significant performance improvements on the FaMTEB benchmark, particularly excelling in chatbots and Retrieval-Augmented Generation (RAG) systems.

- 1Hakim: Farsi Text Embedding ModelMCINEXT, Sharif University of Technology · 2025年



