aozalevsky/search_slm_demo
The aozalevsky/search_slm_demo is a 4.5 billion parameter model based on Qwen/Qwen3.5-4B, specifically designed to convert natural language requests for macromolecular structures into RCSB PDB Search API v2 queries. It operates efficiently on a single 16 GB GPU and excels at translating complex, free-text queries into structured API calls, including handling follow-up edits and providing structured errors for non-search requests. This model is optimized for precise biological and chemical structure data retrieval, offering a specialized alternative to general-purpose LLMs for scientific search applications.
Loading preview...
Overview
The aozalevsky/search_slm_demo is a specialized 4.5 billion parameter model, fine-tuned from Qwen/Qwen3.5-4B, designed to translate natural language requests into structured queries for the RCSB PDB Search API v2. It runs efficiently on a single 16 GB GPU, making it accessible for targeted scientific applications. The model is continuously refined, with version 4.10 representing a full retrain incorporating user feedback and expanding searchable fields.
Key Capabilities
- Natural Language to API Conversion: Transforms free-text requests (e.g., "find all heteroduplexes resolved by NMR") into precise JSON-formatted RCSB PDB Search API v2 queries.
- Query Editing and Refinement: Can modify previous queries based on short follow-up instructions (e.g., "only cryo-EM, released after 2020").
- Structured Error Handling: Provides structured error messages for requests that are not valid structure searches or lack necessary information.
- Extensive Search Coverage: Supports a wide range of biological and chemical search parameters, including organisms (virus families, common names), named objects (helicases, intrinsically disordered proteins), ligands (covalent inhibitors, ligand classes, approved drugs), quality metrics (clashscore, Ramachandran outliers), experimental conditions (crystallization pH, funding agencies), and post-translational modifications.
- Logical Operator Interpretation: Understands natural language logical operators (e.g., "X-ray and NMR structures", "neither A nor B") and interaction partners.
- High Accuracy: Achieves 100% acceptable answers on version 4 items in its gold set evaluation, with 99.7% valid queries.
Good For
- Biomedical Researchers: Ideal for scientists and researchers needing to quickly and accurately query the RCSB PDB for specific macromolecular structures without manual API construction.
- Scientific Data Platforms: Can be integrated into platforms requiring natural language interfaces for biological structure databases.
- Specialized Search Applications: Suitable for applications where precise, domain-specific natural language understanding is critical for data retrieval, offering a focused alternative to general-purpose LLMs.