stella-base-en-v2
| Property | Value |
|---|---|
| Dimension | 768 |
| Sequence Length | 512 |
| Language | English |
| License | MIT |
| Framework | PyTorch |
What is stella-base-en-v2?
stella-base-en-v2 is a specialized text embedding model designed for retrieval and semantic similarity tasks. It's notable for requiring no prefix text or special instructions, making it simple to use while achieving strong performance across various benchmarks. The model generates 768-dimensional embeddings and can handle sequences up to 512 tokens in length.
Implementation Details
The model employs multiple advanced training techniques including contrastive learning with hard negatives, Elastic Weights Consolidation (EWC), and cosent loss. It achieves an impressive average score of 62.61 on the MTEB benchmark, with particularly strong performance in Classification (75.28) and STS tasks (83.02).
- Mean pooling for text vector generation
- Optimized for both short and long text understanding
- Trained with multiple specialized loss functions
- No need for special prefixes or instructions
Core Capabilities
- Strong performance in classification tasks (75.28 on MTEB)
- Excellent semantic textual similarity (83.02 on STS)
- Efficient retrieval and reranking capabilities
- Balanced performance across multiple task types
Frequently Asked Questions
Q: What makes this model unique?
The model stands out for its simplicity of use (no prefixes needed) while maintaining competitive performance across diverse tasks. It's particularly strong in classification and semantic similarity tasks.
Q: What are the recommended use cases?
The model excels in text similarity comparison, document retrieval, semantic search, and classification tasks. It's particularly well-suited for applications requiring semantic understanding of English text without complex preprocessing.
