[Srikarmk]

While I was training my project, I edited a copy of the training config and accidentally deleted its model: line. When mlx-lm starts training, any setting missing from your config is filled in from its built-in defaults, and the default model is a small one, Qwen/Qwen3-0.6b. So instead of failing or asking, it downloaded that model (about 1.2 GB) and trained it, when you meant Qwen2.5-Coder-1.5B. It did print the model’s name, but only inside a summary box shown after the download had finished.

My change makes mlx-lm say, before anything downloads, “no model was given, so I’m using the default”, which would have caught the mistake straight away.