From training data to outputs: How language models inherit patterns from text data

The stylistic behavior of large language models can change considerably when fine-tuned on narrowly focused datasets. This thesis examines whether adapting a model to a stylistically consistent data corpus of a single author leads to measurable changes in generated texts, and how such changes can be systematically evaluated. The main objective is to experimentally assess the extent of stylistic shift induced by fine-tuning and to verify it using multidimensional quantitative analysis. The thesis is divided into six chapters. The introductory part outlines the theoretical foundations of discourse analysis, stylometry, and the functioning of large language models. This is followed by the construction of a reference discourse profile of the author, based on qualitative analysis of selected segments. The core of the thesis describes the experimental setup, including the preparation of the training corpus, fine-tuning of three model architectures (GPT-2, Mistral, Qwen), the design of an evaluation prompt set, and the generation of model outputs. The final analytical section evaluates linguistic shifts across lexical, syntactic, pragmatic, emotional, and embedding-based dimensions, comparing model behavior before and after fine-tuning. The results show that fine-tuning induces consistent stylistic changes across all three models. The most notable shifts appear in the stabilization of vocabulary, changes in the distribution of parts of speech, adjustments in overall tonal expression, and modifications to sentence organization. Embedding-based metrics confirm a global shift of generated texts toward the reference author’s style, with Qwen and Mistral displaying the most stable adaptation. Importantly, the analysis does not indicate any increase in toxicity or degradation of technical text quality.