The API uses scale=false by default. But the online ocr has scaling turned ON by default (see the checkbox in the web form).
The OCR engine is pre-trained, so you (as a user) can not train it. But of course we continually improve it. => Do you have a few more images for us where the detection has problems?