AGmind/qmd-query-expansion-ru

TEXT GENERATIONConcurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AGmind/qmd-query-expansion-ru is a 2 billion parameter Russian query expansion model developed by AGmind, based on Qwen/Qwen3-1.7B. It is specifically fine-tuned to generate relevant Russian query expansions for the qmd local search engine, addressing the limitations of the English-only stock model. This model ensures valid output contracts and handles Russian morphology, providing Hyde, lexical, and vector expansions.

Loading preview...

Overview

AGmind/qmd-query-expansion-ru is a specialized 2 billion parameter model designed for Russian query expansion within the qmd local search engine. It serves as a direct replacement for the stock tobil/qmd-query-expansion-1.7B model, which is English-only and fails to process Russian queries effectively, often hallucinating English boilerplate or breaking the output format.

Key Capabilities

  • Russian Language Support: Specifically trained on Russian data to generate accurate and relevant expansions for Russian search queries.
  • Valid Output Contract: Achieves a 99.8% valid output contract rate (1 hyde:, 3 lex:, 2 vec:) for Russian queries, compared to 0.0% for the stock model.
  • Morphological Awareness: Addresses the limitations of the Porter stemmer in qmd by covering Russian word forms and synonyms in its lexical expansions.
  • Reduced Hallucinations: Eliminates English boilerplate hallucinations, maintaining Russian output for 99.3% of responses.
  • Drop-in Replacement: Easily integrated into qmd by updating a single line in the index.yml configuration.

Training Details

The model is based on Qwen/Qwen3-1.7B and fine-tuned using LoRA (r16/α32 all-proj) over 5 epochs, following the upstream qmd finetuning recipe. It was trained on 5,075 Russian queries from MIRACL-ru, Mr.TyDi-ru, and Yandex.Q, distilled using DeepSeek with a rule-based reward filter.

Use Cases

This model is ideal for qmd users who need to perform effective local search in Russian, ensuring that query expansions are linguistically appropriate and maintain the expected output format for improved search relevance.

Finetunes
2 models