Arthur-75/storm-qwen3-0.6B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 4, 2026Architecture:Transformer0.0K Featherless Exclusive Cold

Arthur-75/storm-qwen3-0.6B is a 0.6 billion parameter Qwen3-based model developed by Arthur-75, designed for self-supervised lexical query expansion. It utilizes the STORM (Stepwise Token Optimization with Reward-guided beaM search) framework to train a rewriter through generation guided by retrieval metrics. This model excels at generating retrieval-effective keywords for search queries, offering an infrastructure-light alternative to dense neural retrieval.

Loading preview...

Model Overview

Arthur-75/storm-qwen3-0.6B is a 0.6 billion parameter model based on the Qwen3 architecture, developed by Arthur-75. It implements the STORM (Stepwise Token Optimization with Reward-guided beaM search) framework for self-supervised lexical query expansion. This approach trains a rewriter by guiding generation with retrieval metrics, scoring candidate expansions against a BM25 index, and pruning low-reward continuations. This process transforms the retrieval reward into a token-level signal, focusing exploration on retrieval-effective vocabulary.

Key Capabilities

  • Lexical Query Expansion: Generates semantic-related keywords from a given query to improve retrieval effectiveness.
  • Retrieval-Guided Generation: Utilizes a novel method where a beam of candidate expansions is scored against a BM25 index, providing token-level supervision.
  • Efficiency: Offers a competitive, infrastructure-light alternative to dense neural retrieval, retrieving as fast as plain BM25.
  • Multilingual Transfer: Demonstrates zero-shot transferability to 18 languages (MIRACL dataset), outperforming dedicated multilingual dense retrievers.
  • Scalability: The STORM framework is shown to be effective across various backbone sizes, from 0.6B to 8B parameters.

Good For

  • Improving Search Relevance: Enhancing the performance of lexical retrievers like BM25 by addressing vocabulary mismatch.
  • Query Rewriting: Generating effective query terms for information retrieval systems.
  • Applications Requiring Fast Retrieval: Ideal for scenarios where maintaining the speed of standard inverted indexes is crucial.
  • Multilingual Search: Applicable for zero-shot keyword generation in diverse linguistic contexts.

Finetunes
1 models