Train on files that you think are good - this way, the model will learn to emulate good style and practices. A fine-tuned code completion model generates more relevant suggestions. As a result, you will see a measurable increase in the percentage of code written by the model: from about 30% to 45%.