SpandZero

SpandTranslate started on off-the-shelf models. None of them recognised SFX, the sound effects drawn into the art; they wrapped several texts in one giant box and missed webtoon-specific elements. Some were closed to commercial use.

On right-place-and-right-class, the ready-made stack scored 38%. Nothing better could be downloaded; it had to be built.

We wrote a labelling constitution that splits every piece of text into four classes: bubble, text in bubble, free text and SFX. We built a human-approved gold set, then trained models generation by generation: each one from scratch, on data labelled by the previous one. No generation could touch the next one's data before passing our own exam.

The rule was simple: no paid training until the dataset is proven clean.

• Twin boxes. The model gave the same text two classes and the translation was printed twice. It had leaked in from the previous generation's labels and our class-blind scorer never saw it. We rewrote the scorer: live twins dropped from 1,083 to 7.
• The model was fine; the last step ate it. One filter rule after detection was deleting correct boxes. Fixing it took the same model from 66% to 78%; even the old model gained 11 points.
• Training without burning money. The dataset went through ten cleaning rounds; training did not start until 40 out of 40 random samples came back clean.

The labelling standard matters more than the architecture. Longer training is not the fix; data is. And you have to measure the ruler too: a class-blind scorer can show 85% for a model that is really at 75%.