DacteloBy the Dactelo team at Taktek.
A first research training run, with its errors recorded
On 28 October 2026, Dactelo recorded the result of a five-epoch Kraken handwriting training experiment on a subset of Muharaf. It used 6,995 training lines, 368 validation lines and 1,334 test lines. CPU training took about four hours.
The measured test character error rate was 45.69%, and the word error rate was 94.40%. Those values show that this short experiment remained far from useful transcription quality. The run tested Dactelo’s training procedure on that dataset; it did not evaluate a released pretrained recognizer or Lebanese court handwriting.
The result and checkpoints were retained rather than described as a successful reader. Muharaf’s noncommercial terms also keep this experiment in the research pool. The useful milestone is a completed, measured training path, together with evidence that a completed run alone says little about readiness for the target task.