hotchpotch/query-crafter-japanese-Qwen3-1.7B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:2BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 2, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

The hotchpotch/query-crafter-japanese-Qwen3-1.7B is a 1.7 billion parameter model based on the Qwen3 architecture, developed by Yuichi Tateno. It is specifically designed for efficient generation of diverse search queries from Japanese text documents. This model excels at creating various query types, including keywords, questions, and summaries, making it ideal for training vector search systems and rerankers in Japanese information retrieval.

Loading preview...

Model Overview

The query-crafter-japanese-Qwen3-1.7B is a lightweight, 1.7 billion parameter model developed by Yuichi Tateno, optimized for generating high-quality search queries from Japanese text. It is part of a collection of models designed to efficiently create diverse question-query pairs for information retrieval tasks, particularly for training vector search systems and rerankers, without the overhead of larger commercial LLMs.

Key Capabilities

  • Diverse Query Generation: Produces seven distinct categories of queries from Japanese documents, including keywords, synonym_keywords, query, alt_query, title, faq, and summary.
  • High Performance: Achieves an average score of 0.8701 in relevance evaluation against DeepSeek-R1 using the BAAI/bge-reranker-v2-m3, demonstrating competitive quality.
  • Exceptional Speed & Cost-Efficiency: Processes approximately 48,000 tokens/second for input and generates 2,200 tokens/second for output, enabling query generation for 10,000 documents in under 100 seconds. Offers significant cost savings compared to API services for large-scale query generation.
  • Apache 2.0 License: Available for unrestricted use in projects.

Ideal Use Cases

  • Training Vector Search Systems: Generating synthetic query data to improve the performance of vector search and retrieval-augmented generation (RAG) systems.
  • Reranker Training: Creating diverse query-document pairs for fine-tuning reranking models.
  • Large-Scale Japanese Information Retrieval: Efficiently generating queries for vast collections of Japanese documents where speed and cost are critical factors.