Provence-reranker-debertav3-v1
| Property | Value |
|---|---|
| Parameter Count | 430 million |
| Context Length | 512 tokens |
| License | CC BY-NC 4.0 |
| Paper | arXiv:2501.16214 |
| Architecture | DeBERTa-v3-large based |
What is provence-reranker-debertav3-v1?
Provence is a sophisticated context pruning model designed to enhance retrieval-augmented generation (RAG) systems. Developed by Naver Labs Europe, it intelligently removes irrelevant sentences from retrieved passages based on user questions, improving both processing speed and reducing noise in the context provided to language models.
Implementation Details
Built on DeBERTa-v3-large architecture, Provence processes input by encoding all sentences in a passage simultaneously, enabling it to capture complex relationships between sentences. The model was trained on a combination of MS Marco and Natural Questions datasets, making it robust across various domains.
- Automated threshold-based sentence selection
- Joint sentence encoding for better coreference handling
- Plug-and-play compatibility with any LLM
- Support for batch processing multiple contexts
Core Capabilities
- Context pruning with adjustable thresholds (0.1-0.5)
- Passage reranking based on relevance scores
- Title preservation option for better context maintenance
- Support for English language processing
- Efficient batch processing with customizable batch sizes
Frequently Asked Questions
Q: What makes this model unique?
Provence stands out for its ability to maintain context quality while reducing irrelevant information, achieved through its joint sentence encoding approach and automatic threshold-based selection system. It's particularly notable for working across various domains without requiring domain-specific training.
Q: What are the recommended use cases?
The model is ideal for RAG applications where context efficiency is crucial, such as question-answering systems, document summarization, and information retrieval tasks. It's particularly valuable when working with large documents where context reduction can significantly improve processing speed and response quality.
