V2: Update model card with new metrics

README.md

@@ -14,11 +14,24 @@ metrics:
 - precision
 - recall
 pipeline_tag: image-classification
+model-index:
+- name: document-moire-detector
+  results:
+  - task:
+      type: image-classification
+      name: Moiré Pattern Detection
+    metrics:
+    - type: accuracy
+      value: 0.9950
+      name: Test Accuracy
+    - type: f1
+      value: 0.9950
+      name: Test F1
 ---
 
-# Document Moiré Detection Model
+# Document Moiré Detection Model (V2)
 
-A fine-tuned **DeiT-tiny** (Vision Transformer) model for detecting moiré patterns in document images.
+A fine-tuned **DeiT-small** (Vision Transformer, 22M params) model for detecting moiré patterns in document images.
 
 ## Model Description
 
@@ -34,26 +47,29 @@ visual artifacts that commonly occur when:
 
 ## Training
 
-- **Base model:** `facebook/deit-tiny-patch16-224` (5.5M parameters)
-- **Training data:** 6,000 samples (3,000 clean + 3,000 synthetic moiré)
+- **Base model:** `facebook/deit-small-patch16-224` (22M parameters)
+- **Training data:** 8,000 samples (4,000 clean + 4,000 synthetic moiré)
 - **Source images:** [rvl-cdip document classification dataset](https://huggingface.co/datasets/hf-tuner/rvl-cdip-document-classification)
-- **Moiré generation:** 4 synthetic methods:
+- **Moiré generation:** 6 synthetic methods:
   1. Resize aliasing (screen-camera simulation)
   2. Frequency-domain pattern overlay
   3. Multi-frequency band interference with color fringing
   4. Screen pixel grid + capture simulation
+  5. **Subtle moiré** — low-strength single-frequency patterns (hard examples)
+  6. **Localized moiré** — partial-image patterns with gaussian mask
 - **Epochs:** 5
-- **Learning rate:** 5e-5 (cosine schedule)
+- **Learning rate:** 3e-5 (cosine schedule)
 - **Effective batch size:** 64
+- **Label smoothing:** 0.05
 
 ## Performance
 
 | Metric    | Validation | Test (held-out) |
 |-----------|-----------|-----------------|
-| Accuracy  | 99.8%     | 99.5% |
-| F1 Score  | 0.998     | 0.995 |
-| Precision | 100%      | 99.3% |
-| Recall    | 99.6%     | 99.7% |
+| Accuracy  | 98.5% | 99.5% |
+| F1 Score  | 0.985 | 0.995 |
+| Precision | 98.2% | 99.3% |
+| Recall    | 98.8% | 99.7% |
 
 ## Usage
 
@@ -86,6 +102,13 @@ with torch.no_grad():
 print(model.config.id2label[predicted_class])  # 'clean' or 'moire'
 ```
 
+## Version History
+
+| Version | Model | Train Size | Methods | Val F1 | Test F1 |
+|---------|-------|-----------|---------|--------|---------|
+| V1 | DeiT-tiny (5.5M) | 6,000 | 4 | 0.998 | 0.995 |
+| **V2** | **DeiT-small (22M)** | **8,000** | **6** | **0.985** | **0.995** |
+
 ## Limitations
 
 - Trained on synthetic moiré patterns — may not capture all real-world moiré variations