AIArchiveInfo/Yuan-embedding-2.0-en

Byte-identical preservation mirror of `IEITYuan/Yuan-embedding-2.0-en` at revision `78cf57002fc6`, archived 2026-09-21 by AIArchive. All credit belongs to the original authors. No weights were trained, fine-tuned, or altered in any way. The original license (apache-2.0) is included verbatim and continues to govern this copy.

<h2 align="left">Yuan-embedding-2.0-en</h2>

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

<h2 align="left">Usage</h2>

pip install -U sentence-transformers==3.4.1

使用示例：

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("IEITYuan/Yuan-embedding-2.0-en")
sentences_1 = ["data-1", "data-2"]
sentences_2 = ["data-3", "data-4"]
embeddings_1 = model.encode(sentences_1, normalize_embeddings=True)
embeddings_2 = model.encode(sentences_2, normalize_embeddings=True)
similarity = embeddings_1 @ embeddings_2.T
print(similarity)