ms-marco-TinyBERT-L2-v2
| Property | Value |
|---|---|
| Model Type | Cross-Encoder |
| Performance (NDCG@10) | 69.84 |
| Speed | 9000 docs/sec |
| Author | cross-encoder |
| Hub URL | https://huggingface.co/cross-encoder/ms-marco-TinyBERT-L2-v2 |
What is ms-marco-TinyBERT-L2-v2?
ms-marco-TinyBERT-L2-v2 is a lightweight and efficient cross-encoder model specifically designed for the MS Marco passage ranking task. As part of the Version 2 series, it represents a significant improvement over its predecessor, offering an optimal balance between performance and speed. The model excels at information retrieval tasks, particularly in query-passage matching scenarios.
Implementation Details
The model can be implemented using either the Transformers library or SentenceTransformers framework. It processes query-passage pairs to produce relevance scores, making it ideal for re-ranking applications. The model supports a maximum sequence length of 512 tokens and can be easily integrated into existing search pipelines.
- Achieves 32.56 MRR@10 on MS Marco Dev set
- Processes 9000 documents per second on a V100 GPU
- Supports both PyTorch and Transformers implementations
- Optimized for production environments
Core Capabilities
- Fast and efficient passage ranking
- Query-passage relevance scoring
- Integration with ElasticSearch for retrieve & re-rank workflows
- Batch processing of multiple query-passage pairs
Frequently Asked Questions
Q: What makes this model unique?
The model stands out for its exceptional speed (9000 docs/sec) while maintaining competitive performance metrics. It's particularly suitable for applications requiring real-time response with moderate accuracy requirements.
Q: What are the recommended use cases?
The model is ideal for information retrieval systems that need to re-rank passages efficiently, particularly in search engines, question-answering systems, and document retrieval applications where speed is crucial.
