NV-Embed v2

NV-Embed-v2 is an embedding model that produces dense vector representations of text, optimized for retrieval, semantic search, and other natural language tasks. It is built on a Mistral-7B-v0.1 decoder-only LLM and uses a latent attention pooling layer to generate pooled embeddings, which consistently outperforms mean pooling or last-token approaches. The model removes the causal attention mask during contrastive training to enhance representation learning, and applies a two-stage instruction-tuning method: first, contrastive training on retrieval datasets with in-batch negatives and curated hard negatives; second, blending non-retrieval tasks into instruction tuning to improve both retrieval and non-retrieval accuracy.

Benchmark Performance

As of August 30, 2024, NV-Embed-v2 holds the No. 1 position on the Massive Text Embedding Benchmark (MTEB) leaderboard with an overall score of 72.31 across 56 tasks. It also ranks No. 1 in the retrieval sub-category with a score of 62.65 across 15 tasks, making it a strong foundation for retrieval-augmented generation (RAG) pipelines. Additionally, it achieved the highest scores in the Long Doc section and the second-highest scores in the QA section of the AIR Benchmark, which covers out-of-domain information retrieval topics beyond MTEB.

Technical Highlights

• Pooling: Latent-Attention (embedding dimension: 4096).
• Training: Hard-negative mining methods that use positive relevance scores to remove false negatives.
• Acceptance: The underlying research paper (NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models) was accepted at ICLR 2025 as a Spotlight paper.
• Compression: The paper provides analysis of model compression techniques (pruning, knowledge distillation, quantization) for generalist embedding models.

For detailed technical descriptions, refer to the paper: NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models.