| Property | Value |
|---|---|
| Parameters | 22.3M |
| License | MIT |
| Base Model | cross-encoder/ms-marco-MiniLM-L-6-v2 |
| Paper | ColBERT Paper |
What is col-minilm?
col-minilm is an efficient passage ranking model based on the ColBERT architecture, specifically optimized for use with Vespa.ai. It's built upon the MS-Marco-MiniLM-L-6-v2 base model and trained using the ColBERT training routine, offering a perfect balance between performance and efficiency.
Implementation Details
The model employs a contextualized late interaction approach over BERT, making it approximately 2x faster than its medium counterpart while maintaining competitive performance. It achieves an impressive MRR@10 of 0.364 on the MS Marco dev set, with strong recall metrics across different k values.
- Recall@50: 0.816
- Recall@200: 0.905
- Recall@1000: 0.939
- ONNX export support for Vespa.ai integration
Core Capabilities
- Efficient passage search and ranking
- Optimized for production deployment with Vespa.ai
- Contextual embedding generation
- Support for both query and document encoding
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its efficient architecture that maintains high-quality results while being significantly faster than larger alternatives. It's specifically optimized for production deployment with Vespa.ai and includes ONNX export capabilities.
Q: What are the recommended use cases?
This model is ideal for production-scale passage ranking tasks, particularly when integrated with Vespa.ai. It's especially suitable for applications requiring efficient search and ranking over large document collections while maintaining high accuracy.