In recent years, social media has rapidly become an essential platform for daily life and emotional expression. However, with the rise of social pressures, particularly during the COVID-19 pandemic, public mental health has been significantly impacted. Traditional methods for detecting negative sentiment on social media have several limitations, such as difficulty in accurately capturing subtle emotional nuances, lack of adaptability to diverse contexts, and inefficiency in handling large-scale data. These shortcomings hinder timely intervention and effective psychological support. To address these challenges, this study proposes a detection method based on large models to improve the accuracy and management of negative sentiment detection on social media. We collected and analyzed a large amount of user text data from the Weibo platform, identifying seven major negative sentiments and their corresponding features. Based on this, a high-quality dataset of 800 entries was constructed, combining expert manual screening with automatic annotation techniques. We fine-tuned large language models like Baichuan, Qwen, and GPT on the llama-factory platform and systematically evaluated their performance in detecting negative sentiment. Leveraging the advanced capabilities of large language models in semantic understanding and generation, our method overcomes the limitations of traditional techniques, better capturing emotional differences, adapting to diverse contexts, and efficiently handling large-scale data. The key contributions of this research include summarizing and annotating seven major negative sentiments and their features, constructing a high-quality abnormal psychological dataset, and systematically evaluating the effectiveness of large language models in this field.