Image too small to scale!! (2x48 vs min width of 3)

[mrobe]

I've been getting very poor results with jpn_vert, but when training to improve it I'm getting a million errors (below).

I've checked all the existing issues on this here and tried all the suggestions, but I'm still stuck.

Trying this:

gmake training MODEL_NAME=jpn_vert_1 START_MODEL=jpn_vert FINETUNE_TYPE=Impact

I get endless errors like this:

Image data/jpn_vert_1-ground-truth/seg-017-0001.lstmf not trainable
Image too small to scale!! (2x48 vs min width of 3)
Line cannot be recognized!!
[...]

I don't understand this error, as none of my image files for training have these dimensions. They are all vertical strips of Japanese, generally around 120 x 3200 px, all PNG files at 600 dpi. Here is a ZIP (small) of my ground-truth folder.

EDIT: reading elsewhere on the site it sounds like tesstrain will scale every training image to 48px height — is that correct? If so, how should I train vertical Japanese? Rotate everything 90°? Otherwise, cutting the vertical images by hand would be a big pain. I know there are hocr tools for this but I have checked the hocr output and the bounding boxes cut through characters all the time (which I assume is why the OCR is giving poor results), so AFAICT that is really not a viable solution either.

How can I proceed?

My system: macOS Monterey v12.7.6.

tesseract --version
tesseract 5.4.1
   leptonica-1.84.1
   libgif 5.2.1 : libjpeg 8d (libjpeg-turbo 3.0.0) : libpng 1.6.43 : libtiff 4.6.0 : zlib 1.2.11 : libwebp 1.4.0 : libopenjp2 2.5.2