ningpy/redflag-detection-V4.1

Redflag Detection V4.1 (V20d, 65 rules)

Fine-tuned from Qwen2.5-7B-Instruct for medical red-flag extraction. Aligned with 2026-08 "Red Flad_New" spec, extended with 6 additional critical rules missing from the original spec.

What's new in V4.1 vs V4.0

• —Fixed `is_male` hallucination (0/269 val cases; V4.0 had schema gap where sex was in prompt but 0 training samples used it)
• —6 new critical rules covering previously uncovered surgical/OB emergencies:
• —V20-60 testicular_torsion (sex=male defensive gate)
• —V20-61 pyelonephritis (painful_urination + fever + severe)
• —V20-62 severepid (fouldischarge + fever + severe)
• —V20-63 eye_emergency (angle-closure glaucoma / sudden vision loss; any age)
• —V20-64 caudaequina (newincontinence + severe_weakness; any adult age)
• —V20-65 postpartum_hemorrhage (sex=female defensive gate)
• —Defensive sex gates on Row 34-39 (pregnancy) — blocks conflict only, not required
• —Pediatric age fallback — Row 40-46 accept is_child/is_baby boolean when age_band absent
• —50 new training samples covering sex extraction from "Patient Info: Age X, Male/Female" blocks (EN/中文/BM)

Performance (val 269, strict)

Performance (clinical-fair, SUSPECTED = TP for recall)

• —Precision: 0.753
• —Recall: 0.853
• —F1: 0.800

Usage

python3 -m vllm.entrypoints.openai.api_server \
    --model ningpy/redflag-detection-V4.1 \
    --served-model-name redflag \
    --dtype float16

Requires V20e rule engine (65 compiled rules with sex_gate + age fallback).

Training

• —3816 samples (V20c 3766 + patch20d 50 sex-field samples)
• —LoRA r=32 alpha=64, 7 target projections
• —3 epochs @ 2e-5, seq len 1900, 8×V100 DDP