Snowflake Arctic Embed M
| Property | Value |
|---|---|
| Parameter Count | 109M |
| Embedding Dimension | 768 |
| License | Apache-2.0 |
| Paper | Technical Report |
What is snowflake-arctic-embed-m?
Snowflake Arctic Embed M is a medium-sized text embedding model that represents the optimal balance between performance and efficiency in the Arctic Embed family. Based on the intfloat/e5-base-unsupervised architecture, it achieves state-of-the-art retrieval performance with an MTEB Retrieval Score (NDCG@10) of 54.90, surpassing comparable models in its class.
Implementation Details
The model is trained through a sophisticated multi-stage pipeline that begins with pretraining on approximately 400M samples of mixed public datasets and proprietary web search data. This is followed by focused training on 1M carefully curated triplets of query-positive-negative documents, utilizing hard negative mining techniques. The model produces 768-dimensional embeddings and supports a context window of 512 tokens.
- Multi-stage training pipeline with hard negative mining
- Optimized for retrieval tasks with 768-dimensional embeddings
- Supports both sentence-transformers and Hugging Face transformers implementations
- Includes specialized query prefixing for enhanced retrieval performance
Core Capabilities
- State-of-the-art retrieval performance in its size class
- Efficient processing with 109M parameters
- Versatile implementation options including JavaScript support via Transformers.js
- Optimized for both symmetric and asymmetric similarity search
Frequently Asked Questions
Q: What makes this model unique?
The model achieves superior retrieval performance through its specialized training pipeline and optimization techniques, offering a perfect balance between model size and performance. It outperforms other base-sized models while maintaining practical deployment requirements.
Q: What are the recommended use cases?
The model excels in text retrieval tasks, semantic search applications, and document similarity comparisons. It's particularly well-suited for enterprise-scale deployments where both accuracy and efficiency are crucial.
