Sparset/Synth-v1-qwen3.5-4b
Sparset/Synth-v1-qwen3.5-4b is a 4.5 billion parameter decoder-only vision-language transformer, fine-tuned by Sparset from Qwen/Qwen3.5-4B. It specializes in document synthesis, combining evidence, resolving conflicts, performing grounded calculations, and handling insufficient evidence from documents provided in the prompt. This model is optimized for grounded answering over text for research and evaluation in RAG pipelines, demonstrating improved performance on an internal 500-case document-synthesis benchmark.
Loading preview...
Sparset Synth v1 (Qwen3.5-4B) Overview
Sparset Synth v1 is a 4.5 billion parameter model developed by Sparset, fine-tuned from Qwen/Qwen3.5-4B. It is specifically designed for document-synthesis tasks, where it processes multiple documents supplied in the prompt to generate grounded answers. This model excels at synthesizing information, resolving conflicting sources, performing calculations based on provided data, and identifying when evidence is insufficient.
Key Capabilities
- Document Synthesis: Combines information from multiple supplied documents to form coherent answers.
- Grounded Reasoning: Resolves conflicting or superseded sources and performs calculations based on document content.
- Evidence Handling: Accurately identifies when supplied evidence is insufficient to answer a query, or asks for clarification.
- Conversational Context: Follows conversational references within multi-turn interactions.
- Improved Performance: Achieved a composite score of 95.22 on an internal 500-case document-synthesis benchmark, up from 91.16 for its base model, with significant gains in insufficient-evidence handling (+19.10) and cross-document calculation (+13.40).
Good For
- Research and Evaluation: Ideal for studying grounded answering over text, particularly with retrieved chunks, tool results, or document excerpts in RAG pipelines.
- Complex Information Extraction: Use cases requiring the synthesis of information from diverse sources, such as policy documents, financial records, or legal texts.
- Grounded QA Systems: Developing systems that need to provide answers strictly based on provided evidence, with citation capabilities.
Note: This is a research release for non-commercial use only, due to restrictions from a portion of its training data. It is not intended for commercial or production deployment.