gloomcheng commited on
Commit
92e2fa4
·
verified ·
1 Parent(s): 2ae08aa

fix(style): remove block img from headings to ensure clean typography and proper heading alignment

Browse files
Files changed (1) hide show
  1. README.md +13 -13
README.md CHANGED
@@ -45,7 +45,7 @@ widget:
45
 
46
  ---
47
 
48
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/chart.svg" width="18" height="18" align="center" /> Technical Specifications & Benchmark Comparison
49
 
50
  | Metric / Specification | IlhaEmbed-311M (Flagship) | IlhaEmbed-97M (Ultra-Lightweight) | Release Gate / Baseline |
51
  |---|---:|---:|---:|
@@ -63,7 +63,7 @@ widget:
63
 
64
  ---
65
 
66
- ### <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/flask.svg" width="18" height="18" align="center" /> Official MTEB Chinese Benchmark Comparison
67
  Evaluated using official **MTEB (Massive Text Embedding Benchmark v2.21.0)** on Chinese biomedical and semantic tasks, benchmarked head-to-head against the leading open-source model **BAAI/bge-small-zh-v1.5**:
68
 
69
  | Task | Metric | **IlhaEmbed (97M / 37MB)** | **BAAI/bge-small-zh** | Analysis |
@@ -78,13 +78,13 @@ Evaluated using official **MTEB (Massive Text Embedding Benchmark v2.21.0)** on
78
 
79
  > **Score Interpretation**: In information retrieval benchmarks, MAP/NDCG@10 in the 0.45–0.52 range represents top-tier open-source performance (e.g. BGE-small, OpenAI text-embedding-3-large). IlhaEmbed delivers competitive retrieval at a fraction of the computational footprint.
80
 
81
- ### <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/clipboard.svg" width="18" height="18" align="center" /> Realistic Clinical Shorthand Evaluation (20 Arbitrary Multi-Clause Notes)
82
  To verify real-world robustness beyond curated single terms, IlhaEmbed was tested against 20 dirty, unstructured, multi-clause clinical notes (spanning emergency triage, Taigi spoken complaints, nursing discharge summaries, and ICU notes):
83
  - **Top-1 Dominant Intent Accuracy**: **95.0% (19/20)**
84
  - **Top-3 Intent Coverage**: **100.0% (20/20)**
85
  - **Boundary Edge Case**: Past checkup completion (*"114年成健已做"*) scored 0.703 for `checkup_intent` vs 0.692 for `care_event`. The small 0.011 margin demonstrates why pure semantic embeddings must be paired with downstream temporal validation rules and clinical human-in-the-loop oversight.
86
 
87
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/shield.svg" width="18" height="18" align="center" /> Safety & Regulatory Boundary (SaMD Exemption)
88
 
89
  - **Intended Use**: Assistive terminology alignment, semantic routing, and candidate recommendation (Suggest-with-Review).
90
  - **Non-SaMD Posture**: Under Taiwan TFDA / international SaMD regulatory guidance, IlhaEmbed does not diagnose, treat, or autonomously formulate clinical care decisions. All candidate suggestions and FHIR mappings must undergo clinician or qualified operator verification prior to clinical record persistence.
@@ -92,7 +92,7 @@ To verify real-world robustness beyond curated single terms, IlhaEmbed was teste
92
 
93
  ---
94
 
95
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/terminal.svg" width="18" height="18" align="center" /> Usage
96
 
97
  ### 1. Flagship (311M) — Python / Sentence-Transformers
98
  ```python
@@ -135,7 +135,7 @@ print(normed.shape) # (2, 384)
135
 
136
  ---
137
 
138
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/scale.svg" width="18" height="18" align="center" /> Data Governance & Open-Source Principles
139
 
140
  1. **Zero Protected Health Information (PHI)**: No patient records, electronic medical records (EMR), or private institutional data are included in training datasets or model checkpoints.
141
  2. **Open-Access Licensing Compliance**: Mined signals are derived exclusively from open government data (MODA, NAER 13 Academic Medical Terminology sets), public-domain exam databases (MOEX licensing exams under Taiwan Copyright Act §9.1.5), and licensed terminology descriptions.
@@ -156,7 +156,7 @@ print(normed.shape) # (2, 384)
156
 
157
  ---
158
 
159
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/sparkles.svg" width="18" height="18" align="center" /> 臨床痛點與核心升級亮點
160
 
161
  台灣各級醫療院所的電子病歷、護理交班、社區健檢表上充滿高度在地化的行話與臨床縮寫:
162
  - `L-CT`:代表低劑量胸部電腦斷層(肺癌早期篩檢)。
@@ -183,7 +183,7 @@ print(normed.shape) # (2, 384)
183
 
184
  ---
185
 
186
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/chart.svg" width="18" height="18" align="center" /> 雙版本評測成效對照 (Benchmark Comparison)
187
 
188
  | 評測維度/指標 | **IlhaEmbed-311M (Flagship)** | **IlhaEmbed-97M (Edge / Kiosk)** | 歷史開源基準 (jina/ckip/bge) | 發布門禁要求 |
189
  |---|---:|---:|---:|---:|
@@ -201,7 +201,7 @@ print(normed.shape) # (2, 384)
201
 
202
  ---
203
 
204
- ### <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/flask.svg" width="18" height="18" align="center" /> MTEB 官方基準實測對照 (Massive Text Embedding Benchmark)
205
  使用官方最新版 MTEB v2.21.0 針對中文醫療與語意任務進行同機同卡嚴謹對照,對比開源主流標竿模型 **BAAI/bge-small-zh-v1.5**:
206
 
207
  | MTEB 官方評測任務 | 核心指標 | **IlhaEmbed (97M / 37MB)** | **BAAI/bge-small-zh** | 實測結論與意義 |
@@ -216,13 +216,13 @@ print(normed.shape) # (2, 384)
216
 
217
  > **評測數值常識說明**:在資訊檢索(IR)領域中,MAP / NDCG@10 達 0.45 ~ 0.52 即為全球開源與商業模型(如 BGE-small, OpenAI)之頂尖梯隊。IlhaEmbed 僅以 37MB 體積即在醫療檢索任務超越 BGE-small。
218
 
219
- ### <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/clipboard.svg" width="18" height="18" align="center" /> 隨意病歷速記檢驗 (真實未美化 20 筆複合速記)
220
  使用正式發布版模型對 20 筆包含門診、急診、台語主訴、護理給藥等真實複合長句進行檢驗:
221
  - **主意圖命中率 (Top-1 Dominant Hit)**:**95.0% (19/20)**
222
  - **前三候選覆蓋率 (Top-3 Intent Coverage)**:**100.0% (20/20)**
223
  - **邊界失效案例分析**:`「114年成健已做,糞檢MIF未交,建議下次門診補交檢體」` 預測 `checkup_intent` (0.703) 與 `care_event` (0.692) 僅差 0.011。「已做」與「安排」的時態語意在純向量空間邊界模糊,證明臨床落地時必須保留下游規則過濾與醫事人員覆核(Fail-Closed 殘差機制)。
224
 
225
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/terminal.svg" width="18" height="18" align="center" /> 快速開始 (Quick Start)
226
 
227
  ### 1. Python / Sentence-Transformers(支援雙版本)
228
 
@@ -271,7 +271,7 @@ print("ONNX 輸出維度:", normed.shape) # (2, 384)
271
 
272
  ---
273
 
274
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/shield.svg" width="18" height="18" align="center" /> 醫療法規、SaMD 豁免與安全邊界
275
 
276
  - **預期用途(Intended Use)**:臨床輔助建議、攝取分流與候選重排(Suggest-with-Review Candidate Ranker)。
277
  - **非 SaMD 宣告(Non-SaMD Posture)**:依據台灣衛生福利部食品藥物管理署(TFDA)與國際醫療器材軟體(SaMD)法規指引,IlhaEmbed 不具備自主診斷、疾病處方或獨立醫療決策功能。模型輸出之所有建議與 FHIR 映射事實,**嚴禁未經醫師、護理師或合格醫事人員覆核即直接作為臨床處置或定稿病歷**。
@@ -279,7 +279,7 @@ print("ONNX 輸出維度:", normed.shape) # (2, 384)
279
 
280
  ---
281
 
282
- ## <img src="https://huggingface.co/weemed/IlhaEmbed/resolve/main/assets/icons/scale.svg" width="18" height="18" align="center" /> 資料治理與開源邊界原則
283
 
284
  1. **零受保護健康資訊(Zero PHI)**:模型訓練與評測全流程不包含任何真實病患姓名、身分證號、病歷號或可識別隱私個資。
285
  2. **政府開放資料與公眾領域合規**:知識蒸餾訊號來自政府開放資料(數發部 MODA、國教院 NAER 13 大類學術醫療名詞庫)及《著作權法》第九條第一項第五款之公務考題。
 
45
 
46
  ---
47
 
48
+ ## Technical Specifications & Benchmark Comparison
49
 
50
  | Metric / Specification | IlhaEmbed-311M (Flagship) | IlhaEmbed-97M (Ultra-Lightweight) | Release Gate / Baseline |
51
  |---|---:|---:|---:|
 
63
 
64
  ---
65
 
66
+ ### Official MTEB Chinese Benchmark Comparison
67
  Evaluated using official **MTEB (Massive Text Embedding Benchmark v2.21.0)** on Chinese biomedical and semantic tasks, benchmarked head-to-head against the leading open-source model **BAAI/bge-small-zh-v1.5**:
68
 
69
  | Task | Metric | **IlhaEmbed (97M / 37MB)** | **BAAI/bge-small-zh** | Analysis |
 
78
 
79
  > **Score Interpretation**: In information retrieval benchmarks, MAP/NDCG@10 in the 0.45–0.52 range represents top-tier open-source performance (e.g. BGE-small, OpenAI text-embedding-3-large). IlhaEmbed delivers competitive retrieval at a fraction of the computational footprint.
80
 
81
+ ### Realistic Clinical Shorthand Evaluation (20 Arbitrary Multi-Clause Notes)
82
  To verify real-world robustness beyond curated single terms, IlhaEmbed was tested against 20 dirty, unstructured, multi-clause clinical notes (spanning emergency triage, Taigi spoken complaints, nursing discharge summaries, and ICU notes):
83
  - **Top-1 Dominant Intent Accuracy**: **95.0% (19/20)**
84
  - **Top-3 Intent Coverage**: **100.0% (20/20)**
85
  - **Boundary Edge Case**: Past checkup completion (*"114年成健已做"*) scored 0.703 for `checkup_intent` vs 0.692 for `care_event`. The small 0.011 margin demonstrates why pure semantic embeddings must be paired with downstream temporal validation rules and clinical human-in-the-loop oversight.
86
 
87
+ ## Safety & Regulatory Boundary (SaMD Exemption)
88
 
89
  - **Intended Use**: Assistive terminology alignment, semantic routing, and candidate recommendation (Suggest-with-Review).
90
  - **Non-SaMD Posture**: Under Taiwan TFDA / international SaMD regulatory guidance, IlhaEmbed does not diagnose, treat, or autonomously formulate clinical care decisions. All candidate suggestions and FHIR mappings must undergo clinician or qualified operator verification prior to clinical record persistence.
 
92
 
93
  ---
94
 
95
+ ## Usage
96
 
97
  ### 1. Flagship (311M) — Python / Sentence-Transformers
98
  ```python
 
135
 
136
  ---
137
 
138
+ ## Data Governance & Open-Source Principles
139
 
140
  1. **Zero Protected Health Information (PHI)**: No patient records, electronic medical records (EMR), or private institutional data are included in training datasets or model checkpoints.
141
  2. **Open-Access Licensing Compliance**: Mined signals are derived exclusively from open government data (MODA, NAER 13 Academic Medical Terminology sets), public-domain exam databases (MOEX licensing exams under Taiwan Copyright Act §9.1.5), and licensed terminology descriptions.
 
156
 
157
  ---
158
 
159
+ ## 臨床痛點與核心升級亮點
160
 
161
  台灣各級醫療院所的電子病歷、護理交班、社區健檢表上充滿高度在地化的行話與臨床縮寫:
162
  - `L-CT`:代表低劑量胸部電腦斷層(肺癌早期篩檢)。
 
183
 
184
  ---
185
 
186
+ ## 雙版本評測成效對照 (Benchmark Comparison)
187
 
188
  | 評測維度/指標 | **IlhaEmbed-311M (Flagship)** | **IlhaEmbed-97M (Edge / Kiosk)** | 歷史開源基準 (jina/ckip/bge) | 發布門禁要求 |
189
  |---|---:|---:|---:|---:|
 
201
 
202
  ---
203
 
204
+ ### MTEB 官方基準實測對照 (Massive Text Embedding Benchmark)
205
  使用官方最新版 MTEB v2.21.0 針對中文醫療與語意任務進行同機同卡嚴謹對照,對比開源主流標竿模型 **BAAI/bge-small-zh-v1.5**:
206
 
207
  | MTEB 官方評測任務 | 核心指標 | **IlhaEmbed (97M / 37MB)** | **BAAI/bge-small-zh** | 實測結論與意義 |
 
216
 
217
  > **評測數值常識說明**:在資訊檢索(IR)領域中,MAP / NDCG@10 達 0.45 ~ 0.52 即為全球開源與商業模型(如 BGE-small, OpenAI)之頂尖梯隊。IlhaEmbed 僅以 37MB 體積即在醫療檢索任務超越 BGE-small。
218
 
219
+ ### 隨意病歷速記檢驗 (真實未美化 20 筆複合速記)
220
  使用正式發布版模型對 20 筆包含門診、急診、台語主訴、護理給藥等真實複合長句進行檢驗:
221
  - **主意圖命中率 (Top-1 Dominant Hit)**:**95.0% (19/20)**
222
  - **前三候選覆蓋率 (Top-3 Intent Coverage)**:**100.0% (20/20)**
223
  - **邊界失效案例分析**:`「114年成健已做,糞檢MIF未交,建議下次門診補交檢體」` 預測 `checkup_intent` (0.703) 與 `care_event` (0.692) 僅差 0.011。「已做」與「安排」的時態語意在純向量空間邊界模糊,證明臨床落地時必須保留下游規則過濾與醫事人員覆核(Fail-Closed 殘差機制)。
224
 
225
+ ## 快速開始 (Quick Start)
226
 
227
  ### 1. Python / Sentence-Transformers(支援雙版本)
228
 
 
271
 
272
  ---
273
 
274
+ ## 醫療法規、SaMD 豁免與安全邊界
275
 
276
  - **預期用途(Intended Use)**:臨床輔助建議、攝取分流與候選重排(Suggest-with-Review Candidate Ranker)。
277
  - **非 SaMD 宣告(Non-SaMD Posture)**:依據台灣衛生福利部食品藥物管理署(TFDA)與國際醫療器材軟體(SaMD)法規指引,IlhaEmbed 不具備自主診斷、疾病處方或獨立醫療決策功能。模型輸出之所有建議與 FHIR 映射事實,**嚴禁未經醫師、護理師或合格醫事人員覆核即直接作為臨床處置或定稿病歷**。
 
279
 
280
  ---
281
 
282
+ ## 資料治理與開源邊界原則
283
 
284
  1. **零受保護健康資訊(Zero PHI)**:模型訓練與評測全流程不包含任何真實病患姓名、身分證號、病歷號或可識別隱私個資。
285
  2. **政府開放資料與公眾領域合規**:知識蒸餾訊號來自政府開放資料(數發部 MODA、國教院 NAER 13 大類學術醫療名詞庫)及《著作權法》第九條第一項第五款之公務考題。