SELECTION AND TRAINING OF A LANGUAGE MODEL FOR GENERATING A PLOT-BASED BOARD ROLE-PLAYING GAME

Abstract

The article examines modern foreign and domestic scientific works on automating the game master’s functions in generating plots and supporting interactive narratives. The analysis of methods of teaching language models, the choice of suitable architectures is carried out. The assessment of the quality of the selected model on various tasks of natural language processing was carried out. Language models capable of generating text, analyzing and understanding its meaning are considered. The architecture of the model is defined, which is capable of solving the problem of generating text for plotting a board role-playing game with the possibility of learning on its own Dataset. The process of retraining a pre-trained neural network to adapt to new data or tasks is presented. The sberbank-ai/rugpt3small_based_on_gpt2 model was used to conduct the study. The stages of model development consisted of data preparation, model training and evaluation of its quality. The criteria for choosing the architecture of the model are defined: it is able to train and work in conditions of limited resources. The main stages of further training are described, including: preparation of the Dataset, architecture selection, installation of hyperparameters, model training and evaluation of the results. A negative experience was gained due to the lack of narrative logic and semantic coherence in the text. Based on the negative experience of implementing a pre-trained neural network, it was decided to use YandexGPT. The approach used made it possible to create and rework texts, offer new ideas and take into account the context of communication with the user. The structure of promt requests has been compiled. An API has been implemented that performs the function of a link between the user and the neural network, providing transmission of requests and responses. A neural network has been selected that generates an image based on a text description corresponding to a given promt. A request template has been prepared for GigaChat to generate images based on a pre-created text description. There are restrictions in the script (the maximum number of characters is 3000).