jina-reranker-v1-turbo-en
| Property | Value |
|---|---|
| Parameter Count | 37.8M parameters |
| Architecture | JinaBERT with ALiBi |
| License | Apache 2.0 |
| Paper | JinaBERT Paper |
| Maximum Sequence Length | 8,192 tokens |
What is jina-reranker-v1-turbo-en?
jina-reranker-v1-turbo-en is a high-performance text reranking model designed for blazing-fast operation while maintaining competitive accuracy. Built on the innovative JinaBERT architecture, it represents a balanced compromise between speed and effectiveness with its 6-layer structure and 37.8M parameters.
Implementation Details
The model employs knowledge distillation techniques, learning from a larger teacher model (jina-reranker-v1-base-en) to maintain high accuracy while significantly improving inference speed. It utilizes a symmetric bidirectional variant of ALiBi, enabling it to process sequences up to 8,192 tokens in length.
- 6-layer architecture with 384 hidden size
- Knowledge distillation for optimal performance
- BF16 tensor type for efficient processing
- Supports multiple integration methods including API, sentence-transformers, and transformers.js
Core Capabilities
- Extended sequence length processing (up to 8,192 tokens)
- Competitive NDCG@10 score of 49.60 on BEIR datasets
- 85.13% Hit Rate on LlamaIndex RAG tasks
- Multilingual support focusing on English content
- Compatible with various deployment environments including browser-based applications
Frequently Asked Questions
Q: What makes this model unique?
The model's unique value proposition lies in its optimal balance between speed and accuracy, achieved through knowledge distillation and the innovative JinaBERT architecture. Its ability to handle long sequences up to 8,192 tokens sets it apart from traditional rerankers limited to 512 tokens.
Q: What are the recommended use cases?
The model is ideal for search and retrieval systems requiring fast reranking of results, RAG (Retrieval-Augmented Generation) applications, and any scenario where quick but accurate document relevance scoring is needed. It's particularly suitable for production environments where processing speed is crucial but accuracy cannot be significantly compromised.
