RyanStudio/qwen3.5-2b-rag-rewriter
RyanStudio/qwen3.5-2b-rag-rewriter is a 2.3 billion parameter Qwen3.5-based language model developed by RyanStudio, fine-tuned for Retrieval Augmented Generation (RAG) rewriting tasks. It significantly improves retrieval performance, achieving up to 96.89% Top-10 accuracy in RAG benchmarks. This model is optimized for rephrasing queries to enhance the relevance and effectiveness of information retrieval systems.
Loading preview...
RyanStudio/qwen3.5-2b-rag-rewriter: Enhanced RAG Performance
This model, developed by RyanStudio, is a 2.3 billion parameter variant of the Qwen3.5 architecture, specifically fine-tuned for Retrieval Augmented Generation (RAG) rewriting. Its primary function is to rephrase or rewrite user queries to improve the accuracy and relevance of information retrieved from external knowledge bases.
Key Capabilities and Features
- Optimized for RAG Rewriting: Designed to enhance retrieval performance by intelligently modifying queries.
- Significant Performance Gains: Demonstrates substantial improvements over raw retrieval, with Top-10 accuracy reaching 96.89% and Top-1 accuracy at 85.21% in internal benchmarks.
- Efficient Training: Utilizes Unsloth and Huggingface's TRL library for faster training.
- Base Model: Fine-tuned from Qwen/Qwen3.5-2B.
- Dataset and Benchmarks: Training and evaluation datasets were generated using Claude 4.6 Sonnet, and embeddings were created with all-MiniLM-L12-v2.
Use Cases
This model is ideal for applications requiring improved information retrieval, such as:
- Question Answering Systems: Enhancing the quality of retrieved documents for more accurate answers.
- Search Engines: Optimizing query understanding to return more relevant search results.
- Chatbots and Virtual Assistants: Improving the ability of conversational agents to fetch precise information based on user input.