PISCO-solar
| Property | Value |
|---|---|
| Parameter Count | 10.9 billion |
| Base Model | upstage/SOLAR-10.7B-Instruct-v1.0 |
| License | CC BY-NC 4.0 |
| Developer | Naver Labs Europe |
| Compression Rate | 16x |
What is pisco-solar?
PISCO-solar is an innovative context compression model specifically designed for Retrieval Augmented Generation (RAG) applications. Built on the SOLAR-10.7B backbone, it implements a unique two-adapter architecture that enables efficient document compression and question answering capabilities.
Implementation Details
The model architecture consists of two key components: an encoder adapter that compresses input contexts into 8 embedding vectors, and a decoder adapter that processes these embeddings alongside queries to generate answers. This design enables a 16x compression rate for documents up to 128 tokens while maintaining remarkable accuracy.
- Efficient document compression with 8 embedding vectors per document
- 5x faster inference speed with pre-compressed document collections
- Minimal accuracy loss (0-3%) on QA benchmarks
- Support for documents up to 128 tokens in length
Core Capabilities
- High-accuracy response generation from compressed documents
- Domain-agnostic performance across various data types
- Efficient RAM utilization through document compression
- Seamless integration with existing RAG pipelines
Frequently Asked Questions
Q: What makes this model unique?
PISCO-solar's distinctive feature is its ability to compress documents while maintaining high accuracy in question answering tasks. The 16x compression rate combined with 5x faster inference makes it particularly valuable for large-scale RAG applications.
Q: What are the recommended use cases?
The model is optimized for RAG-based question answering systems where document retrieval and processing speed are crucial. It's particularly effective when working with large document collections that need to be pre-compressed for efficient retrieval and inference.
