How to get more precision in detection in training custom model in Google colab?

See more:

computer-vision

, +

I have a dataset created in roboflow.

There are 2 cases.

case training in roboflow.

manual training in googleColab.

the same dataset after training in both I get far better results in roboflow.

these are roboflow metrics.

Roboflow Metrics
mAP: 96.4%

Precision: 94.8%

Recall: 93.7%

these are google colab metrics.

Performance Metrics:

mAP (Mean Average Precision) at 50% IoU: 90.1%

mAP (Mean Average Precision) at 50-95% IoU: 67.6%

Precision: 89.8%

Recall: 87.2%

i dont know how roboflow reached 95 precision , in google colab it is not going over 90% precision even if I train for more epochs instead it slightly reduces due to overfitting.

# Use the mounted drive path as the save directory
!yolo task=detect mode=train model=yolov8s.pt data=/content/ocrscale-6/data.yaml epochs=75 imgsz=640 batch=16 save_period=10 project=/content/drive/MyDrive/YOLOv8_Checkpoints

this is the command I am using, what extra is roboflow doing that is providing better results.

should I switch to yolov8m instead yolov8s. or any other alternatives like adding augmentation.

What I have tried:

Trained model 4 to 5 times but results are similar and not reaching the same with roboflow ones.