ms-marco-MiniLM-L-6-v2
| Property | Value |
|---|---|
| License | MIT |
| Framework | PyTorch, JAX |
| Performance | NDCG@10: 74.30, MRR@10: 39.01 |
| Speed | 1800 documents/second |
What is ms-marco-MiniLM-L-6-v2?
ms-marco-MiniLM-L-6-v2 is a specialized cross-encoder model designed for information retrieval tasks, particularly excelling in passage ranking. It represents version 2 of the architecture, bringing significant improvements over its predecessors in both performance and efficiency.
Implementation Details
The model is built on the MiniLM architecture with 6 layers, optimized specifically for the MS Marco Passage Ranking task. It can be easily implemented using the SentenceTransformers library and is designed to efficiently process query-passage pairs for ranking purposes.
- Optimized for passage ranking and reranking tasks
- Built on efficient MiniLM architecture with 6 layers
- Supports maximum sequence length of 512 tokens
- Achieves strong performance while maintaining reasonable inference speed
Core Capabilities
- High-quality passage ranking with NDCG@10 score of 74.30 on TREC DL 19
- Efficient processing at 1800 documents per second on V100 GPU
- Query-passage pair scoring for information retrieval
- Suitable for production environments requiring balance of speed and accuracy
Frequently Asked Questions
Q: What makes this model unique?
This model strikes an optimal balance between performance and speed, achieving nearly identical results to the larger L-12 variant while being significantly faster. It represents a sweet spot in the efficiency-performance tradeoff curve of the MS Marco cross-encoder family.
Q: What are the recommended use cases?
The model is ideal for information retrieval systems requiring reranking of passages, particularly in production environments where both quality and speed are crucial. It's especially effective when used in combination with first-stage retrievers like ElasticSearch for a retrieve & rerank architecture.
