Model Overview
The caraman/Qwen2.5-7B-mtrag-query-rewriter-final is a specialized 7.6 billion parameter model based on the Qwen 2.5 Instruct architecture. Developed by caraman, it is specifically fine-tuned for query rewriting in multi-turn conversational retrieval systems. The model's primary function is to take a conversation history and a current question, then rewrite the question into a standalone, search-friendly query by resolving pronouns and incorporating necessary context.
Key Capabilities & Features
- Specialized Query Rewriting: Transforms conversational queries into explicit, self-contained search queries.
- Performance: Achieved an nDCG@5 of 0.531 in the SemEval-2026 Task 8 (MTRAGEval), ranking 8th out of 38 systems. It provides a 13.7% relative gain over a no-rewriting baseline.
- Fused LoRA Weights: The Low-Rank Adaptation (LoRA) weights are fused directly into the base model, simplifying direct inference without separate adapter loading.
- Optimized for Apple Silicon: Trained using the MLX framework, indicating potential optimization for Apple Silicon environments.
- Domain-Specific Temperature: Recommends specific inference temperatures (e.g., 0.0 for technical docs, 0.2 for Wikipedia) for optimal performance across different domains.
Ideal Use Cases
This model is particularly well-suited for applications requiring:
- Multi-turn Conversational Search: Enhancing retrieval accuracy in chatbots or virtual assistants that handle follow-up questions.
- Information Retrieval Systems: Improving the effectiveness of search engines by providing clearer, context-rich queries.
- Semantic Search: Generating more precise queries for systems that rely on understanding the full intent behind a user's request.