In practice if you're making up tags on the fly it's not much better than untagged data. A LLM that can figure out what the tags mean can probably just infer it from the data anyway.