bge-micro-v2
| Property | Value |
|---|---|
| Parameter Count | 17.4M |
| Model Type | Sentence Transformer |
| License | MIT |
| Tensor Type | FP16 |
What is bge-micro-v2?
bge-micro-v2 is a highly efficient sentence embedding model that represents the second iteration of a two-step distillation process from BAAI/bge-small-en-v1.5. It maps sentences and paragraphs to 384-dimensional dense vector spaces, making it particularly suitable for tasks like semantic search, clustering, and similarity comparison.
Implementation Details
The model utilizes a BERT-based architecture with mean pooling and has been optimized for both performance and efficiency. It supports a maximum sequence length of 512 tokens and maintains the case sensitivity of input text. The implementation can be easily accessed through both the sentence-transformers library and HuggingFace Transformers.
- 384-dimensional embeddings
- Mean pooling architecture
- FP16 precision for efficient inference
- Supports both standard and cross-encoder applications
Core Capabilities
- Semantic text similarity (achieving over 80% correlation on standard benchmarks)
- Document clustering (demonstrated by strong v-measure scores)
- Information retrieval (showing competitive MAP scores)
- Classification tasks (with accuracy typically above 70%)
Frequently Asked Questions
Q: What makes this model unique?
The model stands out for its exceptional efficiency-to-performance ratio, offering strong embedding capabilities in a very compact 17.4M parameter package. It's particularly notable for achieving competitive results across various benchmark tasks while maintaining a small footprint.
Q: What are the recommended use cases?
The model excels in semantic search applications, document similarity comparison, clustering tasks, and as a feature extractor for downstream tasks. It's particularly well-suited for applications where computational resources are limited but high-quality embeddings are required.
