---
title: NV-Retriever: Improving text embedding models with effective hard-negative mining
url: https://www.ml-quant.com/papers/arxiv/2407.15831/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-09-26
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2407.15831
source_url: https://arxiv.org/abs/2407.15831
featured: 2024-07-24
citations: 108
topic: Other
---


# NV-Retriever: Improving text embedding models with effective hard-negative mining

Hard-Negative Mining: The article suggests positive-aware mining methods for fine-tuning text embedding models, with the NV-Retriever-v1 model outperforming previous methods in the MTEB Retrieval benchmark.

- Source: https://arxiv.org/abs/2407.15831
- Identifier: arXiv:2407.15831
- Released: 2024-07-22
- First featured: Quant Letter No. 58 (2024-07-24): https://www.ml-quant.com/issues/2024-07-24/
- Citations (Semantic Scholar): 108
- Published in: not yet
- Topic: Other

## Related

- [EAGLES: Efficient Accelerated 3D Gaussians with Lightweight EncodingS](https://www.ml-quant.com/papers/arxiv/2312.04564/): The paper introduces a technique that uses quantized embeddings and a coarse-to-fine training strategy to optimize 3D Gaussian splatting, reducing memory storage and speeding up novel-view scene synthesis.
- [Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions](https://www.ml-quant.com/papers/arxiv/2411.14405/): The research investigates the OpenAI o1 model's ability, enhanced by Chain-of-Thought fine-tuning and innovative reasoning strategies, to adapt to wider domains lacking clear standards.
- [Dataset Size Recovery from LoRA Weights](https://www.ml-quant.com/papers/arxiv/2406.19395/): A new task called dataset size recovery aims to determine the number of samples used to train a model, with a method called DSiRe proposed for this purpose.
- [Cell Morphology-Guided Small Molecule Generation with GFlowNets](https://www.ml-quant.com/papers/arxiv/2408.05196/): The research introduces a new HCI-guided molecular design method that uses an unsupervised multimodal joint embedding to generate new molecules with similar phenotypic effects to a given image target.
- [OWL2Vec4OA: Tailoring Knowledge Graph Embeddings for Ontology Alignment](https://www.ml-quant.com/papers/arxiv/2408.06310/): The paper presents OWL2Vec4OA, an improved ontology embedding system that uses edge confidence values from seed mappings to enhance the effectiveness of ontology alignment tasks.
- [Unlocking NACE Classification Embeddings with OpenAI for Enhanced Analysis and Processing](https://www.ml-quant.com/papers/arxiv/2409.11524/): The article introduces a new technique to convert the NACE classification of economic activities in the European Union into simpler forms, while maintaining its hierarchical structure. This enhances the efficiency of tasks such as clustering and classification.