WMT 2025 CreoleMT Systems Description : Martinican Creole and French

AbstractThis paper describes our submissions to the constrained subtask of the WMT25 Creole Machine Translation shared task. We participated with a bidirectional Martinican Creole – French system. Our work explores training-time strategies tailored for low-resource MT, including LoRA fine-tuning, curriculum sampling, gradual unfreezing, and multitask learning. We report competitive results against the baseline on both translation directions.