Medical red-flag symptom extraction model fine-tuned from Qwen2.5-7B-Instruct. Extracts structured medical information from user messages that then feeds a rule engine to determine emergency red-flag rules.

Performance (100-case OOD test set)

Exact Match: 90.0%
Micro F1: 0.950
Precision: 0.959
Recall: 0.940

Val Extraction (269 samples)

Exact Match: 88.5%
Micro F1: 0.870

Training data

8355 samples (V9 + patch14 x6 + patch15 x6, targeted RF-055/RF-049/RF-056)
Coverage: 70 RF rules (RF-001..RF-070)
Multi-lingual: English + Bahasa Melayu

Schema

Output is JSON with:

json

{
  "age": {"value": int, "unit": "years"|"months"|"weeks"|"days"},
  "patient": "self"|"third_party",
  "conditions": ["immunocompromised"|"diabetes"|"asthma"|"pregnant"|...],
  "symptoms": ["fever"|"chest_pain"|"breathlessness"|...],
  "denied_symptoms": [...],
  "modifiers": {"onset": "acute"|"chronic", "fever_celsius": float, ...},
  "context_flags": ["post_flight"|"post_surgery"|"post_trauma"|...]
}

Usage with vLLM

bash

python3 -m vllm.entrypoints.openai.api_server \
    --model ningpy/redflag-detection-V3.0 \
    --served-model-name redflag \
    --dtype float16