Arthur-75/storm-qwen3-4B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 4, 2026Architecture:Transformer0.0K Featherless Exclusive Cold

Arthur-75/storm-qwen3-4B is a Qwen3-based language model, part of the STORM (Stepwise Token Optimization with Reward-guided beaM search) framework, designed for self-supervised lexical query expansion. This model excels at generating retrieval-effective keywords for search queries, outperforming larger proprietary models in lexical retrieval tasks. It is optimized for efficient, transparent keyword generation for BM25 indexes and supports zero-shot transfer across 18 languages.

Loading preview...

Overview

Arthur-75/storm-qwen3-4B is a Qwen3-based language model developed within the STORM (Stepwise Token Optimization with Reward-guided beaM search) framework. STORM is a self-supervised approach for lexical query expansion, specifically designed to train rewriters through generation guided by retrieval metrics. Unlike traditional LLM query rewriting that can produce retrieval-ineffective terms, STORM uses a beam search approach where candidate expansions are scored against a BM25 index, pruning low-reward continuations. This method converts retrieval rewards into a token-level signal, focusing exploration on vocabulary that is effective for retrieval.

Key Capabilities

  • Lexical Query Expansion: Generates semantically related keywords for search queries, optimized for lexical retrievers like BM25.
  • Retrieval-Guided Generation: Utilizes a novel reward-guided beam search to ensure generated keywords are highly effective for retrieval.
  • Efficiency: Enables 0.6B-8B backbones to match or surpass competitive LLM rewriters and larger proprietary models while maintaining the speed of plain BM25 retrieval.
  • Multilingual Support: Demonstrates strong zero-shot transfer capabilities across 18 languages (MIRACL benchmark), outperforming dedicated multilingual dense retrievers.

Good For

  • Improving BM25 Retrieval: Enhancing the performance of lexical search systems by generating effective query expansions.
  • Infrastructure-Light Solutions: Providing a competitive alternative to dense neural retrieval without requiring complex index rebuilding when models change.
  • Multilingual Search Applications: Generating keywords for queries in various languages without specific fine-tuning for each.
  • Keyword Generation: Creating comma-separated lists of keywords from user queries, as demonstrated in the provided example.

Finetunes
1 models