I can't train the model to recognize a single character

[ViktorZam]

Hello, no model can recognize the number 4 in this image: https://drive.google.com/file/d/1xlKTDl0aBUootWFWxIYFJbE1xYjVOfS_/view?usp=sharing
I trained the cyrillic_g2.pth model with the following data: https://drive.google.com/drive/folders/1P1sM0HqML-UlDsmCtO6DQC4dkreqYoO6 ?usp=sharing
I also tried to train the model with such data: https://drive.google.com/drive/folders/1_7O_FEpIcbDdd9gIR4pSZuQ34tVZabCt?usp=sharing

log_dataset.txt
log_train.txt
opt.txt

but the number 4 is also not recognizable...
to use a custom model:

reader = easyocr.Reader(["ru"], False, 
                        model_storage_directory="EasyOCR/model",
                        user_network_directory="EasyOCR/user_network",
                        recog_network='custom_model')

img = cv.resize(img, None, fx=IncSizeImg, fy=IncSizeImg, interpolation=cv.INTER_LANCZOS4)#cv.INTER_LANCZOS4
    if filters:
        gray_img = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
        sharpen_kernel = np.array([[0,-1,0], [-1,5,-1], [0,-1,0]])
        gray_img = cv.filter2D(gray_img, -1, sharpen_kernel)
    
    #binary_img = cv.adaptiveThreshold(gray_img, 255, cv.ADAPTIVE_THRESH_MEAN_C, cv.THRESH_BINARY, 12, 4)
        binary_img = cv.threshold(gray_img,thresh,255,cv.THRESH_BINARY_INV)[1]
        img = binary_img
 

    cv.imwrite("Debug/" + str(time.time()) + ".png", img)
    
    result = reader.readtext(img, detail=0, allowlist=allowchars, paragraph=check_paragraph)

[ViktorZam]

I would also like to know how to determine which images to put in train_filtered and which in val.