ningpy/redflag-detection-V4.0

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The ningpy/redflag-detection-V4.0 is a 7.6 billion parameter medical red-flag symptom extraction model, fine-tuned from Qwen2.5-7B-Instruct. It is specifically aligned with the 2026-08 Medical Advisory "Red Flad_New" specification, incorporating 59 strict IF+AND rules for symptom detection. This model excels at extracting 96 atomic symptoms, new severity modifiers, and context flags, making it highly specialized for medical diagnostic support. It achieves an 89.2% exact match accuracy on strict validation, with a clinical-fair recall of 0.851.

Loading preview...

Overview

ningpy/redflag-detection-V4.0 is a specialized medical red-flag symptom extraction model, built upon the Qwen2.5-7B-Instruct architecture with 7.6 billion parameters. It has been meticulously fine-tuned to adhere to the 2026-08 Medical Advisory "Red Flad_New" specification, which includes 59 rules with strict IF+AND semantics.

Key Capabilities

  • Comprehensive Symptom Extraction: Identifies 96 distinct atomic symptoms, expanding upon a base of 78 with 18 new additions like coughing blood, collapse, and severe weakness.
  • Advanced Modifiers and Context: Incorporates new severity modifiers (e.g., has_severe_complication, is_life_threatening) and context flags (e.g., smoke_inhalation, venomous_bite).
  • Population-Specific Gates: Features flags for specific patient populations such as is_pregnant, is_elderly, is_child, and has_diabetes.
  • High Performance: Achieves an 89.2% Exact Match on strict validation, with a Precision of 0.918 and Recall of 0.749. In clinical-fair evaluation, it boasts a Recall of 0.851.

Training Details

The model was trained on 3766 samples, utilizing LoRA with r=32 and alpha=64 across 7 target projections. Training involved 3 epochs at a learning rate of 2e-5, with a sequence length of 1900, using 8xV100 DDP.

Good For

  • Automated Medical Red-Flag Detection: Ideal for systems requiring precise extraction of critical medical symptoms from text.
  • Clinical Decision Support: Can be integrated into tools that assist healthcare professionals in identifying potential urgent conditions based on patient descriptions.
  • Research and Development: Useful for researchers working on medical NLP and diagnostic AI, particularly those focused on structured symptom extraction and rule-based evaluation.