Add model card

README.md

@@ -7,28 +7,29 @@ tags:
   - url-classification
   - binary-classification
   - autoresearch
-datasets:
-  - iowacat
+  - multi-domain
 metrics:
-  - accuracy: 0.9962
+  - accuracy
 model_index:
-  - name: url-classifier
+  - name: url-classifier-v2
     results:
       - task:
           type: text-classification
-          name: URL Binary Classification
+          name: URL Binary Classification (Multi-Domain)
         dataset:
-          type: iowacat
-          name: URL Classification Dataset
+          type: "synthetic-diverse (26 domains)"
+          name: URL Classification Diverse Dataset
         metrics:
           - type: accuracy
-            value: 0.9962
+            value: 1.0000
 ---
 
-# URL Classifier — Autoresearch
+# URL Classifier v2 — Autoresearch (Multi-Domain)
 
 Binary classifier that predicts whether a URL is a **list page (A)** or a **detail page (B)**.
 
+Trained on **26 diverse domains** across e-commerce, recruitment, news, social, video, travel, education, and tech documentation — significantly improved generalization over the v1 single-domain model.
+
 ## Model Details
 
 - **Architecture**: Custom transformer (Autoresearch framework)
@@ -37,32 +38,40 @@ Binary classifier that predicts whether a URL is a **list page (A)** or a **deta
 - **Model dim**: 384
 - **Vocab**: cl100k_base (100,277 tokens)
 - **Max seq len**: 64
-- **Training time**: 5 minutes on RTX 4060 Laptop
-
-## Training
+- **Training**: 30 min on RTX 4060 Laptop
+- **Training samples**: 2,600 (A=1,300, B=1,300)
+- **Training accuracy**: 100%
 
-Trained with the Autoresearch framework, which combines:
-- **Muon** optimizer for attention/MLP layers
-- **AdamW** for embeddings
-- **Sliding window attention** (SSSL pattern)
-- **Value embeddings** for alternating layers
+## Supported Domains
 
-Final loss: ~0.002 | Accuracy: **99.62%**
+| Category | Domains |
+|----------|---------|
+| E-commerce | Amazon, JD, Taobao, Tmall, Pinduoduo |
+| Recruitment | Zhilian, BOSS, Lagou |
+| News | Sina, NetEase, Tencent News, 36kr |
+| Social | Zhihu, Douban, Xiaohongshu, Reddit |
+| Video | YouTube, Bilibili |
+| Travel | Ctrip, Qunar, Mafengwo |
+| Education | icourse163, imooc |
+| Tech Docs | GitHub, ReadTheDocs, MDN |
 
 ## Usage
 
-```python
-from src.prepare import Tokenizer
-
-tokenizer = Tokenizer.from_directory()
-# Encode a URL
-ids = tokenizer.encode("https://example.com/product/123")
-# Run through model + class_head for classification
+```bash
+pip install torch tiktoken
+python src/infer.py "https://example.com/product/123"   # detail page
+python src/infer.py "https://example.com/search?q=foo"  # list page
 ```
 
 ## Class Labels
 
 | Label | Meaning |
 |-------|---------|
-| 0 | A — List page |
-| 1 | B — Detail page |
+| 0 (A) | List page — search results, category pages, rankings |
+| 1 (B) | Detail page — product page, article, profile, video |
+
+## Limitations
+
+- Bilibili ranking pages may be misclassified as detail pages
+- Very short URLs or URL shorteners may have lower accuracy
+- Third-party evaluation accuracy (~55%) indicates room for improvement with real-world labeled data