BGE Reranker Base
| Property | Value |
|---|---|
| Parameter Count | 278M |
| License | MIT |
| Languages | Chinese, English |
| Framework | PyTorch, ONNX |
What is bge-reranker-base?
BGE Reranker Base is a cross-encoder model designed for efficient reranking of search results and document pairs. Unlike traditional embedding models, it directly computes relevance scores between query-document pairs, offering higher accuracy for ranking tasks. The model contains 278M parameters and supports both Chinese and English languages.
Implementation Details
The model utilizes a cross-encoder architecture based on XLM-RoBERTa, performing full-attention over input pairs to generate accurate relevance scores. It's optimized using cross-entropy loss and can be deployed using PyTorch or ONNX for efficient inference.
- Achieves MAP scores of 81.27% on CMedQAv1 and 84.10% on CMedQAv2
- Supports both PyTorch and ONNX runtime for flexible deployment
- Includes Safetensors implementation for enhanced security
Core Capabilities
- Bilingual reranking support for Chinese and English content
- Efficient reranking of top-k search results
- Strong performance on medical domain tasks
- Compatible with text-embeddings-inference systems
Frequently Asked Questions
Q: What makes this model unique?
The model uniquely combines cross-encoder architecture with bilingual capabilities, offering superior reranking performance compared to traditional bi-encoder models while maintaining practical inference speeds for production use.
Q: What are the recommended use cases?
The model is ideal for reranking top-k search results from a first-stage retriever, particularly in bilingual contexts or medical information retrieval scenarios. It's best used as a second-stage ranker after initial candidate retrieval.
