stella_en_400M_v5
| Property | Value |
|---|---|
| Parameter Count | 435M |
| License | MIT |
| Paper | MRL Paper |
| Base Model | Alibaba-NLP/gte-large-en-v1.5 |
What is stella_en_400M_v5?
stella_en_400M_v5 is an advanced sentence embedding model trained using Multiple Representation Learning (MRL), offering flexible dimension options ranging from 512 to 8192. Built upon Alibaba-NLP's GTE architecture, it simplifies prompt engineering by providing two main prompts for sentence-to-passage (s2p) and sentence-to-sentence (s2s) tasks.
Implementation Details
The model implements a unique architecture that supports multiple embedding dimensions through separate linear projection layers. It achieves state-of-the-art performance on the MTEB benchmark, with the 1024-dimension version performing nearly as well as the 8192-dimension version.
- Supports multiple embedding dimensions: 512, 768, 1024, 2048, 4096, 6144, 8192
- Maximum sequence length of 512 tokens
- Implements memory-efficient attention mechanisms
- Compatible with both SentenceTransformers and Transformers libraries
Core Capabilities
- Semantic text similarity assessment
- Passage retrieval and ranking
- Document clustering
- Pair classification tasks
- Cross-encoder capabilities for various NLP tasks
Frequently Asked Questions
Q: What makes this model unique?
The model's key innovation lies in its implementation of Multiple Representation Learning, allowing for flexible dimension choices while maintaining high performance. The simplified prompting system with just two main prompts makes it particularly user-friendly.
Q: What are the recommended use cases?
The model excels in retrieval tasks, semantic similarity matching, and document classification. The 1024-dimension version is recommended for most applications, offering an optimal balance between performance and computational efficiency.
