Abstract <p>This paper considers the influence of hyperparameters on the efficiency of models for optical handwriting recognition of pre-reform texts by the example of 19th-century handwritten reports of Yenisei province governors. The comparative analysis of model configurations with different architectural components, including normalization modules, feature extraction blocks, and predictors, is carried out. Particular attention is paid to the role of input image resolution and hidden layer size in achieving the optimal tradeoff between prediction accuracy and computational cost. The results allow us to identify key parameters for the development of optical character recognition systems adapted to historical texts with nonstandard orthography and complex structure. Directions for further research include the evaluation of synthetic methods to expand training data and the analysis of alternative architectures, e.g., transformers.</p>