vadimbelsky/qwen3.5-medical-ft-stage3-dpo

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 23, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The vadimbelsky/qwen3.5-medical-ft-stage3-dpo is a 9 billion parameter Qwen3.5-based language model developed by vadimbelsky, specifically fine-tuned for emergency department (ED) triage. This model excels at processing ED SOAP intake notes to output structured triage decisions, including ESI levels, justifications, and intervention targets. It was developed through a three-stage pipeline, culminating in DPO alignment to reduce over-triage while preserving high-risk recall, achieving 75.0% accuracy on the MIETIC evaluation set.

Loading preview...

Overview

This model, vadimbelsky/qwen3.5-medical-ft-stage3-dpo, is a 9 billion parameter Qwen3.5-based language model specifically fine-tuned for emergency department (ED) triage. It processes ED SOAP (Subjective, Objective, Assessment, Plan) intake notes to generate structured triage decisions. The model's development involved a three-stage fine-tuning pipeline, with the final stage utilizing DPO (Direct Preference Optimization) to align its outputs, focusing on reducing over-triage while maintaining high recall for critical cases.

Key Capabilities

  • Structured ED Triage: Outputs ESI (Emergency Severity Index) levels (1-5) with justifications, key clinical findings, time-to-provider targets, and immediate intervention requirements.
  • Optimized for Accuracy & Safety: Achieves 75.0% accuracy on the MIETIC evaluation set, with a high-risk recall (ESI 1+2) of 92% and a reduced over-triage rate of 13.9%. This was achieved through a carefully balanced DPO dataset and a combined loss function.
  • Quantized for On-Device Inference: Available in Q4_K_M GGUF format, suitable for efficient local deployment using tools like llama.cpp.

Good For

  • Research in Medical AI: Ideal for researchers exploring AI applications in clinical decision support, particularly in emergency medicine triage.
  • Developing Triage Support Systems: Can serve as a core component for building experimental systems that assist clinicians in preliminary triage assessments.

Limitations

It is crucial to note that this model is for research purposes only and must not be used for clinical decision-making without licensed clinician oversight. It has an 11.1% under-triage rate and a small percentage of missed high-risk patients, and has not been validated on real ED populations.

Finetunes
1 models