Yuan embedding 2.0 en Yuan embedding 2.0 en is an embedding model specifically designed for English text retrieval tasks. Built on the foundation of Qwen/Qwen3 Embedding 0.6B, we have further optimized it for both Retrieval and Reranking tasks. The key work is as follows: Data Augmentation Hard negative sampling: Dual evaluation is conducted using a Rerank model and LLM to filter out high quality positive and negative samples LLM synthesized data: The Yuan2 M32 is employed to perform LLM based rewriting on the query data within the training dataset Loss function design Multi Task loss Matryoshka Representation Learning The Retrieval tasks adopt InfoNCE with in batch negative The Reranking tasks adopt InfoNCE with in batch negative Yuan embedding 2.0 en 是专门为英文文本检索任务设计的嵌入模型。我们在Qwen/Qwen3 Embedding 0.6B的基础上，针对Retrieval任务与Reranking任务进行了进一步优化。主要工作如下： 数据增强 Hard negative sampling：利用Rerank模型与LLM进行双重评估，筛选出高质量正负样本 LLM合成数据：利用Yuan2 M32针对训练数据query进行LLM重写 损失函数设计 Multi Task loss Matryoshka Representation Learning Retrieval任务使用InfoNCE with in batch negative 针对Reranking任务使用InfoNCE with in batch negative Usage 使用示例：
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy