The SFR-v2-small embedding model offers an enhanced context window of 8,000 tokens and significantly improves user experience and model capabilities.
Required Editions
| Available in: All Editions supported by Data Cloud. See Data Cloud edition availability. |
The following provides technical information about the SFR-v2-small embedding model.
Note If Salesforce removes this model due to regulatory changes affecting LLMs developed by
Chinese entities, you'll receive the lesser of 90 days notice or the maximum time allowed by
applicable law to rebuild your search index with an alternative model.
| MoDel Info | Details |
|---|---|
| Model type | Text embedding model that accepts text as input and generates 1024-dimensional continuous embedding vectors. It is trained with contrastive loss on query and document pairs. |
| Model architecture | 434M parameter language model |
| Context window | 8192 |
| Model lineage | No previous production models |
| Languages | English |
| Finetuned from model | gte-large-en-v1.5 |
| Quantized from model | N/A |
| Disclaimers and copyright attributions | Apache license 2.0 |
Technical Specifications
- Number of layers: 24
- Embedding dimension: 1024
- Number of parameters: 434 million
- Model context size: 8,000
- Supported language: English
Performance and Evaluation
The following captures public benchmarking scores and comparrisons to other embedding models.
- Improvement of +7.5% (15% relative) in retrieval performance over e5-large-v2
- Increased context length from 512 to 8,000 tokens, allowing more flexible knowledge processing
| MoDel Info | MTEB (56 tasks) | Retrieval (15 tasks) | Model Size | Embedding Dimension | Context Length |
|---|---|---|---|---|---|
| e5-large-v2 | 62.2 | 50.6 | 335 million | 1024 | 512 tokens |
| SFR-v2-small | 65.9 | 58.1 | 434 million | 1024 | 8192 tokens |