Evaluation of Prompting Strategies for Cyberbullying Detection Using Various Large Language Models

Sentiment analysis detects toxic language for safer online spaces and helps businesses refine strategies through customer feedback analysis [1, 2]. Advancements in Large Language Models (LLMs) and prompt engineering have introduced novel approaches to sentiment analysis, cyberbullying detection, and toxicity classification. However, several challenges persist, particularly in handling text ambiguity, sarcasm, multilingual contexts, and nuanced emotional comprehension, which limit the ability to achieve accurate and human-aligned results. This study uses the CYBY23 dataset, which contains 112 human-annotated threads. To balance the dataset, synthetic threads were generated using ChatGPT, resulting in a final dataset of 148 threads evenly distributed across two labels: 0 (bullying with no aggression) and 1 (bullying with aggression). Three publicly available LLMs—Deepseek-r1-distillllama-70b (Deepseek), Qwen-2.5-32b (Qwen) and llama3-70b-8192 (Llama)—were systematically evaluated using zero-shot, one-shot, and few-shot prompting strategies, with all models accessed via Groq Cloud APIs. The model outputs were assessed using recall, precision, F1 scores, and accuracy to measure performance in different prompting techniques (PT). In this report, Qwen achieved the highest overall accuracy at 82.43% in few-shot 2, while Llama matched that accuracy in one-shot 2, demonstrating solid performance in few-shot tasks as well. Deepseek showed high variability, thriving with contextual enhancements in zero-shot 2 but struggling in one-shot and fluctuating in few-shot settings. one-shot prompting proved most effective for Llama, while few-shot methods worked best for both Qwen and Llama.