NovaSearch/stella_en_1.5B_v5
NovaSearch/stella_en_1.5B_v5 is a 1.5 billion parameter embedding model developed by NovaSearch, built upon Alibaba-NLP's gte-large-en-v1.5 and gte-Qwen2-1.5B-instruct. Trained with MRL, it offers multiple output dimensions (512d to 8192d) for embeddings, with 1024d providing near-optimal performance. This model is optimized for retrieval tasks, supporting both sentence-to-passage (s2p) and sentence-to-sentence (s2s) similarity with specific prompt templates.
Loading preview...
NovaSearch/stella_en_1.5B_v5: A Multi-Dimensional Embedding Model
NovaSearch/stella_en_1.5B_v5 is a 1.5 billion parameter embedding model developed by NovaSearch, leveraging the architectures of Alibaba-NLP/gte-large-en-v1.5 and Alibaba-NLP/gte-Qwen2-1.5B-instruct. It is specifically trained using the MRL method, which enables it to generate embeddings in multiple dimensions: 512, 768, 1024, 2048, 4096, and 8192.
Key Capabilities & Features
- Flexible Embedding Dimensions: Users can choose from various output dimensions, with 1024d offering a strong balance of performance and efficiency, showing only a 0.001 MTEB score difference compared to 8192d.
- Simplified Prompting: The model provides two standardized prompts for common tasks:
s2p(sentence-to-passage): "Instruct: Given a web search query, retrieve relevant passages that answer the query.\nQuery: {query}"s2s(sentence-to-sentence): "Instruct: Retrieve semantically similar text.\nQuery: {query}"
- Standard Integration: Compatible with
SentenceTransformersandtransformerslibraries for easy text encoding. - Optimized for Retrieval: Primarily designed for tasks requiring semantic similarity and passage retrieval.
Usage Considerations
- The recommended sequence length for input is 512 tokens, as the model was trained on datasets of this length.
- While higher dimensions generally yield better performance, 1024d is often sufficient for most applications.
- A technology report and code for Jasper and Stella models are available, detailing their distillation and SOTA embedding techniques (arXiv:2412.19048, GitHub).