Hugging Face Overview
Overview
Yuan-embedding-2.0-en is an 0.8 billion parameter embedding model from IEITYuan, engineered for English text retrieval and reranking. It builds upon the foundation of Qwen/Qwen3-Embedding-0.6B and incorporates several key optimizations to enhance its performance in semantic search applications.
Key Capabilities
- Optimized for English Text Retrieval: Specifically designed to generate high-quality embeddings for English text, facilitating accurate semantic search.
- Enhanced for Reranking Tasks: Beyond initial retrieval, the model is also fine-tuned to improve the ranking of search results.
- Advanced Data Augmentation:
- Hard Negative Sampling: Employs a dual evaluation process using a Rerank model and an LLM to filter high-quality positive and negative samples, improving model robustness.
- LLM-Synthesized Data: Leverages the Yuan2-M32 model to rewrite query data within the training dataset, expanding and diversifying the training examples.
- Sophisticated Loss Function Design: Incorporates a multi-task loss function and Matryoshka Representation Learning. It uses InfoNCE with in-batch-negative for both retrieval and reranking tasks, which is crucial for learning effective representations.
Good For
- Generating embeddings for English text.
- Improving the accuracy of semantic search systems.
- Enhancing the relevance and order of retrieved documents through reranking.
- Applications requiring robust text similarity and contextual understanding in English.