While training i see in preview that mask differ from what was on manual tool.
I use next gudeline from howto:
1) extract faces from image/video
2) sort extracted faces and remove wrong face detection
3) trune dots and mask using manual tool
4) extract training set from alignment file
After that i see that some mask not actually fit faces. Here example
And here corresponding mask in manual tool.
Why mask in training and in manual tool differs?
I can't tell for sure due to the quality of the screenshot, but it looks like you might be using the "extended" mask for training. If you didn't specify another mask that is the one that you'd be using. To change it, open the GUI then Settings -> Configure Settings -> Train -> Loss -> Mask Type and set it to the one you used in the manual tool (probably BiseNet-Fp Face)