Multi-Style Text Generation for Enhanced Text Transformation

• December, 2025

We built a single controllable style-transfer model that rewrites any input into five styles—Professional, Casual, Polite, Social, and Emojify using a simple instruction format like “convert to {style}: {text}.” We also created a synthetic multi-style parallel dataset (13,176 train / 1,464 val / 2,500 test) in JSONL and fine-tuned FLAN-T5-base with a reproducible Hugging Face Transformers pipeline on OrangeGrid. The model maintains strong content preservation (ROUGE-1 ≈ 0.736, BLEU ≈ 0.538), and we deployed a React + Flask demo UI for real-time rewriting.