Expert models and fine-tuning

In the docs for fine-tuning it is suggested that you can add in completions with an empty prompt to build knowledge of a domain: OpenAI API

I am wondering if this would help increase the quality if you also have fine tuning with prompt/completion examples (eg you are training for categorising something, but you also provide some extra background knowledge as completions only?), or would it be better to only have prompts and completions as the completion-only knowledge would confuse things?