DacteloBy the Dactelo team at Taktek.
A first adaptation of an existing Arabic recognizer
On 30 October 2026, Dactelo adapted the published OpenITI Arabic base checkpoint using 512 Egyptian handwritten word images. The split contained 410 training, 49 validation and 53 test words. Identical normalized words stayed together so that the test held out word forms from training.
After one adaptation epoch, character error on the 53 test words fell from 69.71% to 64.23%. That is 15 fewer character edits across 274 reference characters. Error remained high, and the tiny sample cannot establish reliable superiority. Writer identities were unavailable, so this was not an unseen-writer test.
The run preserved its checkpoint, source licences, split and scoring records. It used no Lebanese evaluation pages or noncommercial Muharaf training data. This is an internal adaptation prototype, with no public weight release or court-document performance claim. It establishes a working route for further experiments with stronger evaluation.