Qwen models often experience language contamination during Arabic generation, particularly with less frequently trained languages. To address this, I conducted an experiment using
Smoothie-Qwen to implement language suppression specifically during Arabic text generation.
