Arthur-75/storm-qwen3-1.7B
Arthur-75/storm-qwen3-1.7B is a 1.7 billion parameter Qwen3-based language model developed by Arthur-75, specifically fine-tuned for lexical query expansion. It utilizes the STORM (Stepwise Token Optimization with Reward-guided beaM search) framework to generate retrieval-effective keywords for search queries. This model excels at improving lexical retrieval systems like BM25 by addressing vocabulary mismatch, offering an infrastructure-light alternative to dense neural retrieval.
Loading preview...
STORM Qwen3 Keyword Generator
Arthur-75/storm-qwen3-1.7B is a 1.7 billion parameter model based on the Qwen3 architecture, designed for lexical query expansion. It implements the STORM (Stepwise Token Optimization with Reward-guided beaM search) framework, which is a self-supervised method for training rewriters through generation guided by retrieval metrics.
Key Capabilities
- Retrieval-Effective Keyword Generation: Generates semantic-related keywords for a given query, optimized for improving lexical retrieval systems like BM25.
- Token-Level Reward Signal: Converts retrieval rewards into a token-level signal, concentrating exploration on vocabulary effective for retrieval.
- Infrastructure-Light: Offers a competitive alternative to dense neural retrieval without requiring specialized indexes that need rebuilding with model changes.
- Multilingual Transfer: Demonstrated zero-shot transfer to 18 languages (MIRACL), outperforming dedicated multilingual dense retrievers.
How it Works
STORM trains the rewriter by scoring a beam of candidate expansions against a BM25 index at each step, pruning low-reward continuations. This process ensures that the generated terms are not just well-formed but also effective for retrieval.
Usage
The model can be loaded and run using the transformers library, with an example provided for generating keywords from a user query. It supports standard beam search and group beam search for generation, with group beam search often yielding slightly better performance.