alhparsa/qwen35-extract-tsv-v2
alhparsa/qwen35-extract-tsv-v2 is a 9 billion parameter Qwen3.5-based vision-language model, fine-tuned by alhparsa for highly accurate medical admission-fact extraction from document page images. It outputs compact TSV data with 98-100% recall and 100% precision, optimized for speed with a hybrid GatedDeltaNet architecture and DFlash speculative decoding. This model specializes in structured data extraction from scanned medical documents, providing a robust solution for clinical data processing.
Loading preview...
Model Overview
alhparsa/qwen35-extract-tsv-v2 is a specialized 9 billion parameter vision-language model, built upon the Qwen3.5 architecture. It is specifically fine-tuned for medical admission-fact extraction from document page images, delivering its output in a compact TSV format rather than JSON. The model demonstrates exceptional performance, achieving approximately 98–100% recall and 100% precision (post-deduplication) when evaluated by image-grounded judges.
Key Capabilities & Features
- High Accuracy Extraction: Achieves near-perfect recall and precision for medical fact extraction.
- TSV Output: Generates structured data in a compact TSV format, with each fact represented by a line containing page number, fact, tags, temporal date, temporal text, and confidence.
- Vision-Language Integration: Processes document page images directly, leveraging a hybrid GatedDeltaNet architecture (24 GDN + 8 full-attention) combined with a vision tower and MTP head.
- Optimized for Speed: Achieves 4.4–4.9 seconds per page using the DFlash drafter on 2xH100 GPUs, with lossless speculative decoding for speedup.
- Contextual Extraction: Handles multi-page documents by incorporating "Previous context" to ensure facts spanning page breaks are not lost.
- Configurable Tagging: Extracts facts based on a user-provided list of available tags, ensuring relevant information is captured.
- Patient Identity Handling: Patient identity details (name/MRN/DOB/CSN) are extracted to a dedicated CTX line, separate from fact rows.
Good For
- Automated Medical Data Extraction: Ideal for healthcare applications requiring precise and structured extraction of admission facts from scanned medical documents.
- Clinical Workflow Automation: Can significantly speed up the processing of patient records and reduce manual data entry.
- Research and Analytics: Provides high-quality, structured data for medical research and analytical purposes from unstructured image-based documents.