mahiyama/query-understanding-ja-4b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The mahiyama/query-understanding-ja-4b model is a 4.5 billion parameter Qwen3.5-4B variant, fine-tuned with LoRA to convert Japanese e-commerce search queries into structured JSON. It excels at parsing complex queries into product type, search terms, and inclusion/exclusion filters, achieving performance comparable to a 27B parameter model on specific tasks. This model is optimized for query understanding in Japanese product search systems, providing a structured output for improved search relevance.

Loading preview...

Japanese Query Understanding Model

This model, mahiyama/query-understanding-ja-4b, is a specialized 4.5 billion parameter model based on Qwen/Qwen3.5-4B. It is fine-tuned using LoRA to transform natural language Japanese e-commerce search queries into a structured JSON format. This structured output includes product_type, search_terms, include filters, exclude filters, and unresolved conditions, making it ideal for enhancing search system capabilities.

Key Capabilities

  • Structured Query Output: Converts complex Japanese search queries (e.g., "ロジクール以外の静音ワイヤレスマウス 黒") into a consistent JSON schema.
  • Detailed Filtering: Identifies and categorizes explicit inclusion and exclusion conditions (e.g., brand, color, size, feature) from the query.
  • High Accuracy: Achieves an Exact Match score of 69.5% and Attribute F1 of 81.7% on a 600-query Japanese evaluation set, performing similarly to the much larger Qwen3.5-27B model with few-shot prompting.
  • Efficient Inference: Designed for efficient deployment, offering fast inference times suitable for online search applications, especially when batching queries.

When to Use This Model

  • Japanese E-commerce Search: Ideal for developers building or improving search engines for Japanese online stores.
  • Query Understanding: When you need to programmatically interpret user intent from free-form search queries to apply structured filters and boost/demote results.
  • Proof of Concept: Useful for exploring the application of small language models to specialized query understanding tasks, particularly with public datasets like the Amazon Shopping Queries Dataset (ESCI).

Note: This model is a proof of concept (PoC) and not production-ready out-of-the-box. It requires further fine-tuning and evaluation with domain-specific data for real-world applications.