feat: add new arrows model

README.md

@@ -57,6 +57,23 @@ model-index:
             value: 0.8347
             name: F1 Score (Macro)
             verified: false
+  - name: chess-cv-arrows
+    results:
+      - task:
+          type: image-classification
+          name: Image Classification
+        dataset:
+          name: Chess CV Arrows Test Dataset
+          type: chess-cv-arrows-test
+        metrics:
+          - type: accuracy
+            value: 0.9997
+            name: Accuracy
+            verified: false
+          - type: f1
+            value: 0.9997
+            name: F1 Score (Macro)
+            verified: false
 pipeline_tag: image-classification
 ---
 
@@ -129,49 +146,28 @@ The pieces model classifies chess square images into 13 classes: 6 white pieces
 
 ### ↗ Arrows Model (`arrows.safetensors`)
 
-*Training ...*
-
-<!---->
-
-<!-- **Overview:** -->
-
-<!---->
-
-<!-- The arrows model classifies chess square images into 49 classes representing different arrow overlay patterns: 20 arrow heads (N, NE, E, SE, S, SW, W, NW, and intermediate directions), 12 arrow tails, 8 middle segments, 4 corner pieces, and empty squares (xx). This model enables detection and reconstruction of arrow annotations commonly used in chess analysis interfaces. -->
-
-<!---->
-
-<!-- **Training:** -->
-
-<!---->
-
-<!-- - **Architecture**: SimpleCNN (156k parameters, same as pieces model) -->
-
-<!-- - **Input**: 32×32px RGB square images -->
-
-<!-- - **Data**: ~93,000 synthetic images from 55 board styles × arrow overlays -->
-
-<!-- - **Augmentation**: Conservative augmentation with highlight overlays (25%), random crops, and minimal color jitter/noise. No horizontal flips or rotation to preserve arrow directionality -->
-
-<!-- - **Optimizer**: AdamW (lr=0.0003, weight_decay=0.0003) -->
-
-<!-- - **Training**: 200 epochs, batch size 64 -->
-
-<!---->
+**Overview:**
 
-<!-- **Performance:** -->
+The arrows model classifies chess square images into 49 classes representing different arrow overlay patterns: 20 arrow heads, 12 arrow tails, 8 middle segments (for straight and diagonal arrows), 4 corner pieces (for knight-move arrows), and empty squares (xx). This model enables detection and reconstruction of arrow annotations commonly used in chess analysis interfaces. The NSEW naming convention (North/South/East/West) indicates arrow orientation and direction.
 
-<!---->
+**Training:**
 
-<!-- | Dataset               | Accuracy | F1-Score (Macro) | -->
+- **Architecture**: SimpleCNN (156k parameters, same as pieces model)
+- **Input**: 32×32px RGB square images
+- **Data**: ~4.5M synthetic images from 55 board styles × arrow overlays (~3.14M train, ~672K val, ~672K test)
+- **Augmentation**: Conservative augmentation with highlight overlays (25%), random crops, and minimal color jitter/noise. No horizontal flips to preserve arrow directionality
+- **Optimizer**: AdamW (lr=0.0005, weight_decay=0.00005)
+- **Training**: 20 epochs, batch size 128
 
-<!-- | --------------------- | -------- | ---------------- | -->
+**Performance:**
 
-<!-- | Test Data (synthetic) | ~99%     | ~99%             | -->
+| Dataset               | Accuracy | F1-Score (Macro) |
+| --------------------- | -------- | ---------------- |
+| Test Data (synthetic) | 99.97%   | 99.97%           |
 
-<!---->
+The arrows model is optimized for detecting directional annotations while maintaining spatial consistency across the board.
 
-<!-- The arrows model is optimized for detecting directional annotations while maintaining spatial consistency across the board. -->
+**Limitation:** Classification accuracy degrades when multiple arrow components overlap in a single square.
 
 ## Training Your Own Model
 

arrows.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b575e3d87c698b04c1cbe097179c9de3f7f37568e78570367eea549196d1218
+size 645164

config.json

@@ -1,7 +1,7 @@
 {
-  "model_id": "pieces",
+  "model_id": "arrows",
   "architecture": "SimpleCNN",
-  "num_classes": 13,
+  "num_classes": 49,
   "input_size": [
     32,
     32,
@@ -11,18 +11,54 @@
   "framework": "mlx",
   "task": "image-classification",
   "classes": [
-    "bB",
-    "bK",
-    "bN",
-    "bP",
-    "bQ",
-    "bR",
-    "wB",
-    "wK",
-    "wN",
-    "wP",
-    "wQ",
-    "wR",
+    "corner-E-S",
+    "corner-N-E",
+    "corner-S-W",
+    "corner-W-N",
+    "head-E",
+    "head-ENE",
+    "head-ESE",
+    "head-N",
+    "head-NE",
+    "head-NNE",
+    "head-NNW",
+    "head-NW",
+    "head-S",
+    "head-SE",
+    "head-SSE",
+    "head-SSW",
+    "head-SW",
+    "head-W",
+    "head-WNW",
+    "head-WSW",
+    "middle-E-NNE",
+    "middle-E-SSE",
+    "middle-E-W",
+    "middle-N-ENE",
+    "middle-N-S",
+    "middle-N-WNW",
+    "middle-S-ESE",
+    "middle-S-WSW",
+    "middle-SE-NW",
+    "middle-SW-NE",
+    "middle-W-NNW",
+    "middle-W-SSW",
+    "tail-E",
+    "tail-ENE",
+    "tail-ESE",
+    "tail-N",
+    "tail-NE",
+    "tail-NNE",
+    "tail-NNW",
+    "tail-NW",
+    "tail-S",
+    "tail-SE",
+    "tail-SSE",
+    "tail-SSW",
+    "tail-SW",
+    "tail-W",
+    "tail-WNW",
+    "tail-WSW",
     "xx"
   ]
 }