I am working with the VUA20 dataset on Hugging Face and have a baseline metaphor-detection model in place. I now want to refine that system, turn it into a stronger paraphrase-aware model, and benchmark it against two large-language-model backbones—BERT and RoBERTa. The work I need completed is three-fold: 1. Data & Model • Re-train or fine-tune BERT and RoBERTa on the VUA20. • Integrate a paraphrasing component so the model not only flags a metaphor but can restate the sentence in literal form. 2. Explanation Layer • For every detected metaphor, generate a concise textual explanation describing why the phrase is figurative and how the paraphrase conveys the literal meaning. 3. Evaluation & Comparison • Report Accuracy, Recall and F1 Score for each model. • Summarise gains or trade-offs versus my existing baseline. Deliverables • Clean, reproducible code (Python, PyTorch or TensorFlow). • A short technical report with the metric tables and discussion. • Sample output file showing original sentence, model decision, explanation, and paraphrase. • README with setup instructions and commands to replicate the results. If anything in the dataset handling or metric calculation needs special care, note it in the documentation so I can audit it quickly. Looking forward to seeing how far we can push metaphor detection with these architectures.