Multimodal Hate Speech Detection using Fine-Tuned Llama 2 Model

Hate speech detection is crucial for secure online environments, especially with recent policy changes on social media. This study enhances the Llama 2 model by fine-tuning it with a comprehensive Twitter dataset to improve hate speech detection accuracy. The system's frontend allows users to upload or record audio, upload videos, or provide YouTube URLs, with processing managed by the Google Speech Recognition API and the Whisper library. The backend integrates advanced speech recognition and video processing technologies. The refined Llama 2 model achieved an F1 score of 0.94, with accuracy, precision, and recall metrics surpassing other state-of-the-art models. These results demonstrate the model's high effectiveness and reliability in hate speech detection. This research highlights the potential of transformer-based models in enhancing online safety and inclusivity. Future work will focus on expanding the dataset, improving model generalization, supporting additional languages, and enhancing system scalability to better handle evolving hate speech forms.