ms-marco-MiniLM-L-4-v2
| Property | Value |
|---|---|
| License | Apache 2.0 |
| Downloads | 1,632,962 |
| Processing Speed | 2,500 docs/sec on V100 GPU |
| Performance (NDCG@10) | 73.04 on TREC DL 19 |
What is ms-marco-MiniLM-L-4-v2?
The ms-marco-MiniLM-L-4-v2 is a specialized cross-encoder model designed for passage ranking tasks, particularly optimized for the MS Marco dataset. This version 2 model represents a significant improvement over its predecessors, offering an excellent balance between performance and processing speed.
Implementation Details
The model can be implemented using either the Transformers library or SentenceTransformers framework. It's specifically designed for query-passage pair ranking and can process up to 2,500 documents per second on a V100 GPU.
- Achieves 73.04 NDCG@10 on TREC DL 19 dataset
- MRR@10 score of 37.70 on MS Marco Dev set
- Supports both PyTorch and JAX frameworks
- Maximum sequence length of 512 tokens
Core Capabilities
- Efficient passage ranking and reranking
- Query-passage pair scoring
- Integration with both Transformers and SentenceTransformers frameworks
- Batch processing of multiple query-passage pairs
Frequently Asked Questions
Q: What makes this model unique?
This model offers an optimal balance between performance and speed, achieving strong NDCG@10 scores while maintaining efficient processing speeds. It's particularly well-suited for production environments where both accuracy and speed are crucial.
Q: What are the recommended use cases?
The model excels in information retrieval tasks, particularly when used in conjunction with initial retrieval systems like ElasticSearch. It's ideal for reranking passages based on relevance to queries in search applications and document retrieval systems.
