SFR-Embedding-2 R

Salesforce/SFR-Embedding-2_R is a text embedding model that produces dense vector representations optimized for information retrieval and semantic search tasks. Developed by Salesforce Research, it uses a multi-stage training approach to achieve strong performance across diverse embedding benchmarks.

Key Capabilities

The model accepts a query paired with a one-sentence task instruction and supports input sequences up to 4,096 tokens. For retrieval, only queries require an instruction prefix; documents can be passed as-is. This instruction-following architecture enables fine-grained control over the embedding behavior for different downstream tasks.

Benchmark Performance

On the MTEB classification suite, SFR-Embedding-2_R achieves 92.72% accuracy on AmazonCounterfactualClassification, 97.31% on AmazonPolarityClassification, and 61.04% on AmazonReviewsClassification. For retrieval on the ArguAna dataset, it obtains an NDCG@10 of 62.34, MAP@10 of 53.91, Recall@10 of 89.05%, and Precision@1 of 37.77%. In clustering tasks, it reports a v-measure of 54.02 on ArxivClusteringP2P and 48.82 on ArxivClusteringS2S.

Additional Details

The model is tagged as a feature-extraction pipeline and is compatible with both transformers and sentence-transformers libraries. It was created on 2024-06-14 and has accumulated over 3.9 million all-time downloads.

Ethical considerations: This release is intended for research purposes only. Users should evaluate accuracy, safety, and fairness before deploying in high-risk applications. Refer to Salesforce’s acceptable use policies for further guidance.