MFTRAINING(1)

💅 MFTRAINING(1) — the mftraining girlies' manual

okay so bestie, welcome to mftraining from the MFTRAINING(1) section of the manual — she's the feature-training bestie for Tesseract and she's lowkey carrying the whole shape-prototype pipeline. let's get into it 💖

---

✨ NAME ✨

mftraining — feature training for Tesseract. that's it, that's the vibe. 🎀

---

✨ SYNOPSIS ✨

mftraining -U unicharset -O lang.unicharset FILE...

like, you hand her a unicharset, you tell her where the output unicharset goes, and then you yeet a whole list of FILE... at her. slay.

---

✨ DESCRIPTION ✨

so here's the tea 🫖: mftraining takes a list of .tr files and from those she generates three iconic little files:

• inttemp — the shape prototypes 👑
• shapetable — the shape table, obviously
• pffmtable — the number of expected features for each character

there's also a fourth file called Microfeat that this program writes too — but ngl, nobody uses it. it's just there. sitting. being written. not being used. 💀 (not me getting attached to an unused file though 😭)

---

✨ OPTIONS ✨

• -U FILE — (Input) the unicharset, the one generated by unicharset_extractor(1). this is the source-of-truth character set, bestie, treat her with respect 💅
• -F font_properties_file — (Input) the font properties file. every line looks like this, and every field except the font name is straight up 0 or 1 — no in-between, no "maybe", just binary energy:

*font_name* *italic* *bold* *fixed_pitch* *serif* *fraktur*

• -X xheights_file — (Input) the x heights file. every line is this, where xheight is calculated as the pixel x height of a character drawn at 32pt on 300 dpi:

*font_name* *xheight*

and the manpage literally spells out the vibe check for you: [ that is, if base x height + ascenders + descenders = 133, how much is x height? ] 🧚 so yes, they really do want you to do the maths, no cap.

• -D dir — the directory to write output files to. pick a folder, any folder, just make it a good one 🦋
• -O FILE — (Output) the output unicharset that will be given to combine_tessdata(1). this is the glow-up version of your unicharset and she's going places ✨

---

✨ SEE ALSO ✨

go read these too, iykyk:

• tesseract(1)
• cntraining(1)
• unicharset_extractor(1)
• combine_tessdata(1)
• shapeclustering(1)
• unicharset(5)

and the docs link, because of course:

https://tesseract-ocr.github.io/tessdoc/Training-Tesseract.html

---

✨ COPYING ✨

Copyright (C) Hewlett-Packard Company, 1988 — Licensed under the Apache License, Version 2.0. periodt. 📜

---

✨ AUTHOR ✨

The Tesseract OCR engine was written by Ray Smith and his research groups at Hewlett Packard (1985-1995) and Google (2006-2018). literal main character energy from both eras, we love a cross-decade icon 🌈👑

---

okay that's the whole page, you're fully caught up 💜 this was MFTRAINING(1), dated 2026-09-28 — that's the source/version and date right there in the footer, and now it's in a cute little sign-off instead of a boring bottom line. mwah, go train some features bestie 🫶💖