Cyberbullying poses a severe threat to individual mental health and social stability. Existing detection methods suffer from poor domain adaptability, a contradiction between accuracy and lightweight performance, and insufficient interpretability. To address these issues, this study proposed the Chain-of-Thought Optimized LoRA-Quantized DeepSeek-8B (CLQP) Model, a scenario-adaptive joint optimization framework integrating Chain-of-Thought (CoT), Low-Rank Adaptation (LoRA), and quantization for efficient cross-scene detection. Specifically, CoT enhances interpretability and implicit bullying identification; LoRA fine-tunes core attention layers to reduce trainable parameters; NF4 quantization balances lightweight deployment and semantic retention. By integrating and unifying annotations for an external database and adopting targeted data augmentation, CLQP was deployed on an RTX 4060 (8GB VRAM). Experimental results showed that CLQP's F1-score exceeded 85% in four scenarios; the false negative rate (FN) decreased by 26%; the false positive rate (FPR) was controlled within 5.2%; the inference speed reached 2 s per sample; and the GPU memory usage was only 5.3GB.
Si et al. (Mon,) studied this question.