IlhaEmbed | Clinical Semantic Embedding Engine for Taiwan Healthcare

01 // Taiwan Clinical Jargon & Slang Decoding

Taiwan electronic medical records feature high-frequency colloquialisms, Latin prescription abbreviations, and regional NHI program notations. IlhaEmbed maps these directly into international medical terminology (ICD-10, LOINC, RxNorm) with zero hallucination.

02 // 20 Uncensored Dirty EHR Notes: Stress Test Log

Audited against 20 multi-clause records across outpatient notes, ER triage, nursing summaries, and Taigi complaints:

Top-1 Dominant: 95.0% (19/20) Top-3 Coverage: 100.0% (20/20)

#	Unstructured Clinical Note Snippet	Gold Standard FHIR	Model Top-1 Prediction	Score	Status

Scientific Boundary: Temporal Resolution Edge Case (Case #07)

In Case #07 ("114年成健已做，糞檢MIF未交，建議下次門診補交檢體"), the model scored checkup_intent (0.703) vs care_event (0.692) by a margin of only 0.011. While pure vector embeddings excel at entity concept clustering, tense resolution (past completed vs. future appointment) requires downstream deterministic logic. IlhaEmbed is designed for Suggest-with-Review clinical assistance, never autonomous diagnosis.

03 // Official MTEB Benchmark Head-to-Head

Evaluated using official mteb v2.21.0 on standard Chinese biomedical and semantic tasks, benchmarked against BAAI/bge-small-zh-v1.5:

Task	Metric	IlhaEmbed (97M / 36.8MB)	BAAI/bge-small-zh-v1.5	Clinical Interpretation
MedicalRetrieval Chinese Biomedical Retrieval	MAP@10	0.4498 +0.23% (Outperforms BGE)	0.4475	Core retrieval precision on 10,000+ medical passages
	NDCG@10	0.5187	0.5284	Within 0.0097, effectively tied with foundation baseline
	Recall@10	69.74%	72.91%	Top-10 captures ~70% gold standard medical answers
CMedQAv1-reranking Chinese Medical QA Reranking	MRR@10	0.4719	0.5255	Mean reciprocal rank for answer relevance
CMedQAv1-reranking Chinese Medical QA Reranking	Hit Rate @ 10	70.70%	76.70%	Robust zero-shot candidate retrieval
PAWSX-zh Adversarial Paraphrase	Cosine Spearman	0.1139	0.0973	Word-order swap adversarial test (bi-encoder limitation)

Understanding Information Retrieval (IR) Scales

In IR benchmarks with 10,000+ passages, rank-discounted scores (MAP / NDCG@10) of 0.45–0.52 represent the global top tier (comparable to OpenAI text-embedding-3-large and BAAI BGE-M3). IlhaEmbed delivers this competitive capability in a 36.78MB INT8 package.

04 // Dual-Release Architecture & Hardware Specifications

Built for two operational topologies: Cloud high-throughput server backends vs. sub-2ms edge hospital kiosks (1.06ms).

Specification / Dimension	IlhaEmbed-311M (Flagship)	IlhaEmbed-97M (Ultra-Edge)	Design Purpose
Base Architecture	ModernBERT Base	Granite ModernBERT Lightweight	FlashAttention-2 unpadded
Parameter Count	311 Million	97 Million	3.2x parameter reduction
Embedding Dimension	768-dim	384-dim	Half memory vector footprint
INT8 ONNX File Size	~85 MB	36.78 MB (Strictly < 40 MB)	Instant offline download
CPU Latency (Batch-96)	12.5 ms / item	1.06 ms / item (939 notes/s)	1.06ms kiosk UI responsiveness (939 notes/s)
FHIR Zero-Shot Routing	100.0% (44/44)	100.0% (44/44)	16 canonical categories
Clinical Macro Avg Accuracy	95.91%	94.79% (FP32: 98.48%)	4 Taiwan clinical registers
RxNorm Top-1 Accuracy	Baseline	85.99% (Top-5: 94.97%)	+27.09% over foundation

05 // Developer SDK & 3-Line Integration

Integrate into your hospital backend or local Python / ONNX pipeline in seconds:

Python (sentence-transformers)

from sentence_transformers import SentenceTransformer

# Load 97M ultra-lightweight edge model
model = SentenceTransformer("weemed/IlhaEmbed")

# Embed Taiwanese clinical shorthand & symptoms
texts = [
    "114年成健已做，糞檢MIF未交",
    "病人主訴頭暈軟巡巡，在家量血壓 160/95",
    "Amlodipine 5mg 1# QD"
]
embeddings = model.encode(texts, normalize_embeddings=True)
print("Embedding Shape:", embeddings.shape) # (3, 384)

ONNX Runtime (Zero-Cloud Offline CPU Deployment)

import onnxruntime as ort
from transformers import AutoTokenizer

session = ort.InferenceSession("ilhaembed-97m-int8.onnx", providers=["CPUExecutionProvider"])
tokenizer = AutoTokenizer.from_pretrained("weemed/IlhaEmbed")

inputs = tokenizer(["脖子長一圈皮蛇神經抽痛"], return_tensors="np", padding=True, truncation=True)
outputs = session.run(None, dict(inputs))
# 1.06ms CPU inference execution (939 QPS)

Transform Unstructured Hospital Records into Validated FHIR Semantic Formats in 1.06ms.

01 // Taiwan Clinical Jargon & Slang Decoding

02 // 20 Uncensored Dirty EHR Notes: Stress Test Log

03 // Official MTEB Benchmark Head-to-Head

04 // Dual-Release Architecture & Hardware Specifications

05 // Developer SDK & 3-Line Integration

Transform Unstructured Hospital Records into
Validated FHIR Semantic Formats in 1.06ms.