LyrAIX - tweaking the style

Recently, Large Language Models (LLMs) have been tested in different do- mains. This research examined the capabilities of a fine-tuned LLM to gen- erate lyrics in different styles. A pre-trained medium (355M of learning parameters) size GPT-2 model was fine-tuned with a compiled dataset of song parts and corresponding stylistic labels. The dataset was constructed from lyrics collected from the Genius website, which were later filtered and labeled with the help of unsupervised and rule-based classifiers. The model was tested to generate lyrics defined by such stylistic parameters as affect, topic, rhyme scheme, and content explicitness. Additionally, the model was assessed on the preservation of the author’s style originality and distinctive- ness. The results have shown that a fine-tuned LLM is more capable of lyrics generation with defined text explicitness and affect, rather than topics and rhyming scheme. Furthermore, a positive indication of the original author’s style preservation was discovered with the reported average similarity score of 6.167 on a 1 to 10 Likert scale