xlmr-large-vimed-ner

XLM-R-large fine-tuned for BIO NER on Vietnamese clinical notes (bệnh án). It is layer L1 of the JSL-style FHIR pipeline (fhir_jsl).

Labels: SIGN (signs, vital signs, observations), SYMP (symptoms), DX (diagnoses), MED (medications), DOSE (dosage).

Training

• Data: airace, de-identified Vietnamese medical records. Checkpoint full is trained on all 92 dev files (ids in train_meta.json); the 9 blind test files are held out.
• 8 epochs, lr 2e-5, batch 8, 737 windows / 8,284 spans, ~103 s on one H200.

Results (span IoU ≥ 0.5, macro F1)

	test (9 blind)	dev (4-fold CV)
raw output	0.880	0.859
+ span post-processing	0.886	0.868
GLM few-shot (reference)	0.732	0.768

Miss rate obs / dx / med on test: 2.4% / 4.1% / 2.1%.

Post-processing matters. The XLM-R tokenizer splits accented Vietnamese syllables (Rối → ▁R + ối), so expand predicted spans to whole words and merge combined drugs written A + B.

Usage

from transformers import pipeline
ner = pipeline("token-classification", model="qminhlb/xlmr-large-vimed-ner", aggregation_strategy="simple")
ner("Bệnh nhân sốt 38,5 độ, ho khan. Chẩn đoán: Viêm phổi. Paracetamol 500mg x 2 viên/ngày")

Input is capped at 512 tokens; split long notes into overlapping windows.

Limitations

Trained on one highly templated dataset (101 records). Expect lower accuracy on other hospitals' formats and on OCR-noisy text. Not a medical device.