training/train_config.yaml

# YOLOv8 Fine-Tuning Configuration for E-Waste Cannibalization Index

# Dataset Path Specification
data: splits/data.yaml

# Model & Training Strategy
model: yolov8n.pt     # Nano backbone optimized for edge deployment (<15MB TFLite)
imgsz: 960            # High resolution to resolve micro desoldered pads
epochs: 50
batch: 8              # Memory-safe batch size
workers: 2            # CPU dataloader threads

# Loss & Augmentation Parameters (Class Imbalance Fix)
box: 7.5              # Box loss gain
cls: 2.5              # Elevated class loss gain for rare component classes
dfl: 1.5              # Distribution Focal Loss gain

# Synthetic Data Augmentations
copy_paste: 0.8       # Copy-paste rare desoldered pads onto clean boards
mosaic: 1.0           # 4-image mosaic blending
mixup: 0.15           # Mixup image blending
hsv_h: 0.015          # Color hue jitter for flux/burn marks
hsv_s: 0.7            # Color saturation jitter
hsv_v: 0.4            # Brightness jitter
degrees: 15.0         # Random rotation
translate: 0.1        # Translation
scale: 0.5            # Multi-scale zooming
flipud: 0.5           # Vertical flip for 2-sided PCB inspection
fliplr: 0.5           # Horizontal flip