Don't Correct MeBenchmark
FAITHFUL HANDWRITING TRANSCRIPTION

Don't Correct Me

How well do models preserve what was actually written—even when it’s misspelled?

31 systems845 handwritten samples947 scored misspellingsUpdated

Leaderboard

All languages
RankModel
Loading results…
Ranked by preservation · higher is better

Preservation measures faithfulness to the writer. Read overcorrection alongside CER and WER: fewer corrections alone do not mean better recognition. Metric definitions

READING THE RESULTS

What do the scores mean?

Preservation ↑
The annotated misspelling remains in the transcription.
Overcorrection ↓
The model replaces the misspelling with an accepted corrected spelling.
CER / WER ↓
Character and word error rates against the literal reference, averaged across samples.
THE RESEARCH

Faithful transcription of contemporary handwriting

A student’s spelling mistake is part of the handwritten record. This benchmark examines when recognition systems preserve that record and when they silently correct it.

Jonathan Pagel · Khaoula Otmani · Enkelejda Kasneci
Technical University of Munich

MODEL DETAILS