efficient-splade-VI-BT-large-doc
| Property | Value |
|---|---|
| License | cc-by-nc-sa-4.0 |
| Paper | View Paper |
| MRR@10 (MS MARCO dev) | 38.0 |
| Inference Latency | 0.7ms |
What is efficient-splade-VI-BT-large-doc?
This is a specialized document encoder that forms part of a two-model architecture for efficient passage retrieval. It represents the document-side component of the SPLADE (Sparse Lexical AndMask Distillation) architecture, optimized for both performance and efficiency. The model achieves an impressive balance between retrieval quality and computational efficiency, with a MRR@10 of 38.0 on MS MARCO dev set while maintaining extremely low inference latency of 0.7ms.
Implementation Details
The model utilizes a modified DistilBERT architecture with specific optimizations for document encoding. It implements several efficiency-focused techniques including L1 regularization, FLOPS-regularized middle-training, and separate document/query encoders to achieve state-of-the-art performance while maintaining competitive latency.
- Achieves 97.8% R@1000 on MS MARCO dev set
- Optimized for sparse representation learning
- Implements bag-of-words approach with neural enhancement
- Utilizes knowledge distillation for improved efficiency
Core Capabilities
- Fast document encoding with 0.7ms inference latency
- Efficient passage retrieval using sparse representations
- Competitive performance comparable to traditional BM25
- Scalable document indexing for large-scale applications
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its exceptional efficiency-performance trade-off, achieving similar latency to traditional BM25 systems while maintaining competitive retrieval performance. The separation of document and query encoders allows for optimized inference speeds.
Q: What are the recommended use cases?
The model is specifically designed for large-scale passage retrieval tasks where both efficiency and effectiveness are crucial. It's particularly well-suited for applications requiring fast document indexing and retrieval with near state-of-the-art performance.
