This paper explores methods for using large language models (LLM) to detect provocative comments on social media during mass protests. We analyze existing approaches to text data processing and develop a methodology for using LLama3 8B, Mistral 7B, and Gemma 7B models. The effectiveness of the models is evaluated using comments related to the protests in Mongolia in December 2022. The results demonstrate that the Mistral 7B model has the highest accuracy and efficiency in classifying provocative comments. The findings confirm that the use of large language models significantly improves the accuracy and speed of content analysis on social media, which is important for public opinion management and conflict prevention.

Keywords: large language model, social network, provocative comment, mass protest, text analysis, natural language processing, LLama3 8B, Mistral 7B, Gemma 7B, classification, accuracy, recall, F1-score, political instability, public opinion management