Update README.md

README.md

@@ -42,6 +42,46 @@ making near–real-time pathological analysis feasible at scale.
 
 ![image](https://cdn-uploads.huggingface.co/production/uploads/633a801b7646c9f51a05cc92/pynK0OWbjH5WUlu8L7OTj.png)
 
+## Training Details
+
+The model was trained on ImageNet-1K using a high-resolution training
+pipeline adapted from common ImageNet training practices.
+
+- Dataset: ImageNet-1K
+- Input resolution: 512×512
+- Model: CSATv2
+- Optimizer: AdamW
+- Learning rate: 2e-3
+- Learning rate schedule: Cosine
+- Epochs: 300
+- Warmup epochs: 5
+- Weight decay: 2e-2
+- Batch size: 128 (per GPU)
+- Mixed precision training: Enabled (AMP)
+
+### Data Augmentation
+- Random resized crop (scale: 0.08–1.0, ratio: 3/4–4/3)
+- Horizontal flip (p = 0.5)
+- RandAugment (rand-m7-mstd0.5-inc1)
+- Mixup (α = 0.8)
+- CutMix (α = 1.0)
+- Bicubic interpolation
+
+### Regularization
+- Label smoothing: Disabled (handled implicitly via Mixup / CutMix)
+- Dropout / DropPath: Disabled
+- Random erase: Disabled
+
+### Optimization Details
+- Exponential Moving Average (EMA): Enabled  
+  - EMA decay: 0.99996
+- Gradient clipping: Disabled
+- Channels-last memory format: Optional
+
+The training pipeline was adapted from publicly available ImageNet
+training repositories (Solving ImageNet),
+with task-specific modifications for high-resolution and
+high-throughput training.
 
 
 ## Usage