Arthur-75/storm-qwen3-8B

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 4, 2026Architecture:Transformer0.0K Featherless Exclusive Cold

Arthur-75/storm-qwen3-8B is an 8 billion parameter Qwen3-based language model developed by Arthur-75, specifically designed for lexical query expansion. It utilizes a self-supervised framework called STORM (Stepwise Token Optimization with Reward-guided beaM search) to generate retrieval-effective keywords. This model excels at enhancing traditional lexical retrievers like BM25 by overcoming vocabulary mismatch, offering an infrastructure-light alternative to dense neural retrieval, and supports zero-shot transfer to 18 languages.

Loading preview...

STORM Qwen3-8B Keyword Generator

Arthur-75/storm-qwen3-8B is an 8 billion parameter Qwen3-based model developed for lexical query expansion, presented at EMNLP 2026. It introduces STORM (Stepwise Token Optimization with Reward-guided beaM search), a self-supervised framework that trains a rewriter through generation guided by retrieval metrics. This approach transforms retrieval rewards into a token-level signal, concentrating exploration on vocabulary effective for retrieval.

Key Capabilities

  • Lexical Query Expansion: Generates semantic-related keywords to enhance retrieval performance for systems like BM25.
  • Self-Supervised Training: Utilizes a novel reward-guided beam search to optimize keyword generation for retrieval effectiveness.
  • Multilingual Support: Achieves zero-shot transfer to 18 languages (MIRACL benchmark), outperforming dedicated multilingual dense retrievers.
  • Infrastructure-Light: Offers a competitive alternative to dense neural retrieval without requiring specialized corpus encoding or frequent index rebuilding.

Good For

  • Improving BM25 and Lexical Retrievers: Addresses vocabulary mismatch issues in traditional lexical search.
  • Efficient Information Retrieval: Provides fast retrieval speeds comparable to plain BM25 while enhancing result quality.
  • Multilingual Search Applications: Effective for expanding queries across various languages without specific training.
  • Research in Retrieval-Augmented Generation: Demonstrates an innovative method for integrating retrieval metrics into generation processes.

Finetunes
1 models