| Parameter | Description | Recommended |
|---|
| Dataset Snapshot | The snapshot version to use | |
| Base Model | YOLO pretrained model scale | |
| Epochs | Number of complete passes through the dataset | 100-300 |
| Batch Size | Images fed to GPU per step | 16 |
| Image Size | Unified resize dimension during training | 640 |
| Model | Params | Speed | Accuracy | Best For |
|---|
yolov8n.pt | 3.2M | Fastest | ★★ | Extreme real-time needs, few classes |
yolov8s.pt | 11.2M | Fast | ★★★ | Recommended — balanced choice |
yolov8m.pt | 25.9M | Medium | ★★★★ | Higher accuracy required |
yolov8l.pt | 43.7M | Slow | ★★★★★ | High-precision scenarios |
n/s/m/l/x stand for nano/small/medium/large/xlarge — larger models are more accurate but slower at inference.
| Metric | Meaning | Desired Trend |
|---|
| Box Loss | Bounding box localization error | ↓ Decreasing |
| Class Loss | Classification error | ↓ Decreasing |
| mAP50 | Mean Average Precision (IoU=0.5) | ↑ Increasing |
| mAP50-95 | Mean Average Precision (IoU=0.5~0.95) | ↑ Increasing |
| Precision | Accuracy rate | ↑ Increasing |
| Recall | Coverage rate | ↑ Increasing |
| Status | Meaning |
|---|
queued | Waiting for GPU resources to free up |
preparing | Setting up training data and environment |
running | Training in progress |
completed | Training finished normally |
failed | Training terminated abnormally (check logs for diagnosis) |
stopped | You manually stopped training |
| Hardware | Spec |
|---|
| GPU | 4× NVIDIA RTX 6000 Ada Generation |
| VRAM per GPU | 48 GB |
| Total VRAM | 192 GB |
