DacteloBy the Dactelo team at Taktek.
A GPU baseline on held-out Arabic manuscript lines
On 30 October 2026, Dactelo ran its reproducible manuscript-line experiment on a Colab T4 GPU. The prepared Baybars material contained 11,462 training lines, 1,392 validation lines and 1,402 test lines. Source pages stayed within one split, with source mappings and image checks preserved.
The starting OpenITI model completed all 1,402 test lines. Against 49,609 reference characters, strict character error was 66.39%; word error was 96.89%. These are baseline errors on historical manuscript lines. They do not measure Lebanese court handwriting, and the provisional line geometry can affect the result.
A capped adaptation run began afterward. At this journal checkpoint, no adapted-model score had been verified. Recording the starting score separately prevents a training launch from becoming an improvement claim. Dataset provenance, recovered export checks and runtime records accompany the experiment; writer-disjoint evaluation remains unestablished.