ko-reranker
| Property | Value |
|---|---|
| Parameter Count | 560M |
| License | MIT |
| Languages | Korean, English |
| Framework | PyTorch |
| Base Model | BAAI/bge-reranker-large |
| Research Paper | Lost in Middle (Liu et al., 2023) |
What is ko-reranker?
ko-reranker is a specialized Korean language reranking model fine-tuned from the BGE-reranker-large architecture. Unlike traditional embedding models, it directly outputs similarity scores between questions and documents, making it particularly effective for improving RAG (Retrieval-Augmented Generation) systems.
Implementation Details
The model utilizes CrossEntropy loss optimization and accepts question-document pairs as input, producing unrestricted similarity scores. With 560M parameters and F32 tensor type support, it demonstrates improved performance over the base model, achieving an MRR of 0.87 compared to 0.84 for the original BGE-reranker.
- Fine-tuned on translated MS MARCO triplets dataset with 499,184 samples
- Supports both Korean and English text inputs
- Implements efficient batch processing with maximum sequence length of 512
Core Capabilities
- Direct similarity scoring between questions and documents
- Enhanced Korean language understanding for reranking tasks
- Seamless integration with SageMaker deployment
- Superior performance in document relevance ranking
Frequently Asked Questions
Q: What makes this model unique?
The model specifically addresses Korean language reranking needs, offering improved performance over generic rerankers with a 96% accuracy in right-context identification, compared to 93% without reranking.
Q: What are the recommended use cases?
The model excels in RAG applications, document retrieval systems, and any scenario requiring precise ranking of Korean language content based on relevance to queries. It's particularly effective when document ordering significantly impacts downstream task performance.
