\
Webinar · Jul 20\
\
Fine-tuning open-source models: is it time to move off Frontier Lab models?

# SFR-Embedding-Code-2B_R

Salesforce

SFR-Embedding-Code-2B_R is a 2B parameter embedding model by Salesforce for multilingual code/text retrieval, achieving 67.4% NDCG@10 on CoIR benchmark

| Property          | Value                                                          |
|------------------|----------------------------------------------------------------|
| Model Size       | 2B parameters                                                 |
| Developer        | Salesforce Research                                           |
| Paper            | [CodeXEmbed: A Generalist Embedding Model Family](https://arxiv.org/abs/2411.12644) |
| License          | Research-only (Subject to Gemma Terms of Use)               |
| Performance      | 67.4% NDCG@10 on CoIR Benchmark                              |

## What is SFR-Embedding-Code-2B_R?

SFR-Embedding-Code-2B_R is a state-of-the-art embedding model designed for multilingual and multi-task code and text retrieval. Developed by Salesforce Research, it represents the largest model in the SFR-Embedding family, demonstrating superior performance compared to existing open-source code embedding models.

## Implementation Details

The model can be easily implemented using either the Transformers or Sentence Transformers libraries. It supports a maximum sequence length of 32,768 tokens and requires specific instruction formatting for queries. The model architecture is based on Gemma and has been fine-tuned for code retrieval tasks.

- Built on Gemma architecture with 2B parameters
- Supports both code and text embedding generation
- Implements instruction-based query formatting
- Offers flexible integration through popular ML frameworks

## Core Capabilities

- Multilingual code retrieval and understanding
- High-performance text-to-code matching
- Superior embedding quality (67.4% NDCG@10)
- Long context support up to 32K tokens
- Flexible API support for both Transformers and Sentence Transformers

## Frequently Asked Questions

### Q: What makes this model unique?

The model stands out for its superior performance in code retrieval tasks, achieving the highest NDCG@10 score (67.4%) among comparable models on the CoIR benchmark. It's also notable for its large parameter count and versatility in handling both code and text embedding tasks.

### Q: What are the recommended use cases?

The model is specifically designed for research purposes in code retrieval, documentation matching, and code-text alignment tasks. It's particularly useful for applications requiring high-quality code search, code documentation linking, and similar code detection.