nvidia/nv-embedqa-mistral-7b-v2¶
Model Information¶
The nvidia/nv-embedqa-mistral-7b-v2 model is optimized for text question-answering retrieval. It transforms textual information into dense vector representations, facilitating efficient semantic search and retrieval tasks.
- Model Developer: NVIDIA
- Model Release Date: April 12, 2025
- Supported Languages:
- Primary: English (US)
- Potential Support: Other languages under research
Model Architecture¶
- Base Model: Fine-tuned Mistral-7B
- Architecture Type: Transformer encoder
- Layers: 32
- Embedding Dimension: 4096
- Attention Mechanism: Bi-directional attention (converted from original causal attention)
- Pooling Method: Latent-attention pooling
- Training Approach: Two-stage contrastive instruction tuning with hard-negative mining
Benchmark Scores¶
| Benchmark | Metric | Score |
|---|---|---|
| MTEB (Massive Text Embedding Benchmark) | Overall Score | 69.32 |
| MTEB Retrieval Subset (15 tasks) | Score | 59.35 |
| BeIR Benchmark (NQ, HotpotQA, FiQA, TechQA) | Recall@5 | 72.97% |
Note: The model outperforms previous leading embedding models such as E5-mistral-7b-instruct and SFR-Embedding on these benchmarks.