This paper presents our solution for subtask2, which focuses on the automated detection of conspiracy in text. Unlike traditional toxic text detection, conspiracy identification is particularly challenging as it often lacks explicit semantic indicators. To address this, we leveraged a Large Language Model (LLM) as our backbone and employed Low-Rank Adaptation (LoRA) for fine-tuning to enhance detection performance. To generate probabilistic confidence scores while maintaining the generative capabilities of the LLM, we implemented a hybrid loss function that integrates both generative and token classification losses. Additionally, semi-supervised learning with unlabeled data was incorporated to further refine classification accuracy. Our approach achieved a test accuracy of 0.87, ranking 2nd in both stages of the competition leaderboard.