stella-mrl-large-zh-v3.5-1792d
| Property | Value |
|---|---|
| Parameter Count | 326M |
| Training Method | MRL (Margin Ranking Loss) |
| License | MIT |
| Paper | MRL Paper |
What is stella-mrl-large-zh-v3.5-1792d?
stella-mrl-large-zh-v3.5-1792d is a sophisticated Chinese language model specifically designed for sentence embeddings with variable dimensions. Built upon the stella-large-zh-v3-1792d architecture, it implements the MRL (Margin Ranking Loss) training methodology to create versatile embeddings that can be truncated to different dimensions (128-1792) while maintaining performance.
Implementation Details
The model generates embeddings of up to 1792 dimensions that can be truncated to smaller sizes in multiples of 128. Performance scales with dimension size, with optimal results typically achieved at higher dimensions. The model requires no prefix text and supports direct text encoding.
- Flexible dimension selection (128-1792)
- Normalized embeddings post-truncation
- Optimized for retrieval and semantic matching tasks
- Strong performance on CMTEB benchmark
Core Capabilities
- Sentence similarity scoring with high accuracy
- Text retrieval with MAP@100 scores over 82% for some tasks
- Classification tasks with accuracy up to 88%
- Reranking capabilities with MAP scores around 68%
- Clustering with v-measure scores above 54%
Frequently Asked Questions
Q: What makes this model unique?
The model's ability to generate variable-dimension embeddings while maintaining strong performance makes it highly flexible for different application requirements and resource constraints. Users can choose the embedding dimension that best balances performance and computational resources.
Q: What are the recommended use cases?
The model excels in Chinese language tasks including semantic similarity matching, information retrieval, text classification, and document clustering. It's particularly well-suited for applications where embedding dimension flexibility is important.
